diff --git a/.helix/languages.toml b/.helix/languages.toml new file mode 100644 index 00000000..529cdcf3 --- /dev/null +++ b/.helix/languages.toml @@ -0,0 +1,19 @@ +[[language]] +auto-format = true +file-types = [ "zig" ] +language-servers = [ "zls" ] +name = "zig" +roots = [ "build.zig", "build.zig.zon", "zls.build.json" ] +scope = "source.zig" + +[language.indent] +tab-width = 4 +unit = " " + +[language-server.zls] +command = "sh" +args = [ + "-c", + "exec \"$(git rev-parse --show-toplevel)/tools/zls.sh\" \"$@\"", + "zls" +] diff --git a/bazel/BUILD.bazel b/bazel/BUILD.bazel index 434ecfa9..eb2b2ab2 100644 --- a/bazel/BUILD.bazel +++ b/bazel/BUILD.bazel @@ -3,7 +3,9 @@ load("@rules_zig//zig:defs.bzl", "zig_library") cc_library( name = "retain_runfiles_init", - # Mach-O adds an underscore to the exported Zig symbol. + srcs = ["runfiles_init.c"], + # Retain the constructor even when its object is in a static archive. + # Mach-O adds an underscore to the exported C symbol. linkopts = select({ "@platforms//os:macos": ["-Wl,-u,__bazel_runfiles_init"], "//conditions:default": ["-Wl,-u,_bazel_runfiles_init"], diff --git a/bazel/bazel.zig b/bazel/bazel.zig index 2ae3324a..67513227 100644 --- a/bazel/bazel.zig +++ b/bazel/bazel.zig @@ -1,35 +1,26 @@ const std = @import("std"); -const builtin = @import("builtin"); const bazel_runfiles = @import("runfiles"); const stdx = @import("stdx"); var runfiles_global: bazel_runfiles.Runfiles = undefined; -const init_array_section = switch (builtin.object_format) { - .macho => "__DATA,__mod_init_func", - .elf => ".init_array", - else => "", -}; - -export const _bazel_runfiles_init align(@alignOf(usize)) linksection(init_array_section) = &struct { - fn call(argc: c_int, argv: [*c][*:0]u8, envp: [*:null]?[*:0]u8) callconv(.c) void { - _ = argc; - _ = envp; - var threaded: std.Io.Threaded = .init_single_threaded; - runfiles_global = bazel_runfiles.Runfiles.create(.{ - .allocator = std.heap.c_allocator, - .io = threaded.io(), - .argv0 = std.mem.span(argv[0]), - .directory = if (std.c.getenv("RUNFILES_DIR")) |directory| std.mem.span(directory) else null, - .manifest = if (std.c.getenv("RUNFILES_MANIFEST_FILE")) |manifest| std.mem.span(manifest) else null, - }) catch { - std.debug.panic("Unable to find runfiles", .{}); - } orelse { - std.debug.panic("Unable to initialize runfiles", .{}); - }; - } -}.call; +pub export fn _bazel_runfiles_init_impl(argc: c_int, argv: [*c][*:0]u8, envp: [*:null]?[*:0]u8) callconv(.c) void { + _ = argc; + _ = envp; + var threaded: std.Io.Threaded = .init_single_threaded; + runfiles_global = bazel_runfiles.Runfiles.create(.{ + .allocator = std.heap.c_allocator, + .io = threaded.io(), + .argv0 = std.mem.span(argv[0]), + .directory = if (std.c.getenv("RUNFILES_DIR")) |directory| std.mem.span(directory) else null, + .manifest = if (std.c.getenv("RUNFILES_MANIFEST_FILE")) |manifest| std.mem.span(manifest) else null, + }) catch { + std.debug.panic("Unable to find runfiles", .{}); + } orelse { + std.debug.panic("Unable to initialize runfiles", .{}); + }; +} pub fn runfiles(source_repository: []const u8) !bazel_runfiles.Runfiles.WithSourceRepo { return runfiles_global.withSourceRepo(source_repository); diff --git a/bazel/runfiles_init.c b/bazel/runfiles_init.c new file mode 100644 index 00000000..770a4b91 --- /dev/null +++ b/bazel/runfiles_init.c @@ -0,0 +1,15 @@ +extern void _bazel_runfiles_init_impl(int argc, char **argv, char **envp); + +// Preserve the existing startup ABI. Registration happens in C; +// runfiles discovery and initialization remain in Zig. +// glibc and Darwin's dyld pass argc/argv/envp to constructors; musl does not. +// The original Zig initializer already made this assumption to read argv[0]. +// +// Let C emit the platform's constructor registration. Zig's incremental ELF +// linker does not yet handle .init_array's section type, flags, and size correctly. +// TODO: Revert this workaround once Zig issue #37453 is fixed: +// https://codeberg.org/ziglang/zig/issues/37453 +__attribute__((constructor)) +void _bazel_runfiles_init(int argc, char **argv, char **envp) { + _bazel_runfiles_init_impl(argc, argv, envp); +} diff --git a/devenv.lock b/devenv.lock index bb3ef660..6c3be332 100644 --- a/devenv.lock +++ b/devenv.lock @@ -3,11 +3,11 @@ "devenv": { "locked": { "dir": "src/modules", - "lastModified": 1779636362, - "narHash": "sha256-Q/hy/ybvbe91fjA1MbulUwPjJH9yKaFZnSaTkwLE8S0=", + "lastModified": 1793199252, + "narHash": "sha256-LRgdAdMHL3uIX+kix/oyMzG+gMxzSYOzLrrRaoZthog=", "owner": "cachix", "repo": "devenv", - "rev": "9d6e68dac81fab1f7f2ec6784220151b4fba4b59", + "rev": "0d6c556e5f862c14a75a543e2ee4fdb7acc005ad", "type": "github" }, "original": { @@ -22,11 +22,11 @@ "nixpkgs-src": "nixpkgs-src" }, "locked": { - "lastModified": 1778507786, - "narHash": "sha256-HzSQCKMsMr8r55LwM1JuzIOB+8bzk0FEv6sItKvsfoY=", + "lastModified": 1793043970, + "narHash": "sha256-m14odYTdkBUEkJOiEYs2zjGjn8fcjvdCJsm45+k084U=", "owner": "cachix", "repo": "devenv-nixpkgs", - "rev": "8f24a228a782e24576b155d1e39f0d914b380691", + "rev": "d97234e58d4e90e3bf77c082fc3b546e51200938", "type": "github" }, "original": { @@ -39,11 +39,11 @@ "nixpkgs-src": { "flake": false, "locked": { - "lastModified": 1778274207, - "narHash": "sha256-I4puXmX1iovcCHZlRmztO3vW0mAbbRvq4F8wgIMQ1MM=", + "lastModified": 1792812153, + "narHash": "sha256-T2+XBfNbf9pecLUWS4+RtuVNZ3W8K5taG5QdV71mWlk=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "b3da656039dc7a6240f27b2ef8cc6a3ef3bccae7", + "rev": "674a2028661e550a97400c1cc992b9cca9a668a1", "type": "github" }, "original": { diff --git a/dsl/cute/README.md b/dsl/cute/README.md index c7828abe..946c542d 100644 --- a/dsl/cute/README.md +++ b/dsl/cute/README.md @@ -88,19 +88,20 @@ the actual IR regions. | `cute.arch.griddepcontrol_launch_dependents()` | `b.launchDependents()` | | `x.to(Float32)` | `x.to(.f32)` | | `if cond:` / `for i in range(...)` | `b.openIf(cond)` / `b.openFor(0, n, 1, .{})` | -| `cpasync.make_tiled_tma_atom(...)` | `b.makeTiledTmaLoadAtom(...)` | +| `cpasync.make_tiled_tma_atom(...)` | `b.makeTiledTmaAtom(...)` | | `cpasync.tma_partition(...)` | `b.tmaPartition(...)` | -| `cute.copy(atom, src, dst)` | `b.copy(atom, src, dst, pred)` | +| `cute.copy(atom, src, dst)` | `b.copy(atom, src, dst, .{})` | | `cute.make_tiled_mma(atom)` | `b.makeTiledMma(atom, result_type)` | | `tiled_mma.make_fragment_A/B/C(...)` | `b.mmaMakeFragment(...)` | | `cute.gemm(...)` | `b.gemm(...)` | | `cute.arch.alloc_tmem(...)` | `b.allocTmem(...)` | | `cute.arch.retrieve_tmem_ptr(...)` | `b.retrieveTmemPtr(...)` | -* **Layouts are static.** `makeLayout` takes Zig slices and emits one - `cute.static`; the tensor keeps its shape and stride in Zig, so `dim`, - `size` and coordinates need no MLIR queries. Coordinates are tuples of - `Value`s and ints; ints stay static in the `!cute.coord` type. +* **Layouts accept static and dynamic extents.** `makeLayout` takes Zig slices + for a tracked static `Layout`, or nested tuples of integers and SSA `Value`s + for a `View`. Dynamic shape and stride leaves retain their integer widths. + The tracked `Tensor` keeps its shape and stride in Zig; use `makeTensorView` + with dynamic layouts. Integers stay static in coordinate types. * **Stride defaults differ by origin.** `makeLayout` defaults to CuTe's compact left-most stride like `cute.make_layout`; a `tensor` argument defaults to XLA's row-major layout, which is what the buffer holds. @@ -122,7 +123,11 @@ the actual IR regions. copies, tensor-memory fragments, and SM100 block-scaled MMA. CuTe derives their layouts in the compiler, so callers spell out the resulting CuTe type with `View` instead of relying on Python's dynamic type objects. -* **Not covered yet:** dynamic tensor shapes and automatic CuTe layout algebra. +* **TMA tilers follow Python.** `makeTiledTmaAtom` accepts a shape tuple, tile, + or layout, composes it with the GMEM identity layout, and removes the SMEM + stage mode when present. The lower-level `makeTiledTmaLoadAtom` and + `makeTiledTmaStoreAtom` take an already composed coordinate map. `copy` + makes store atoms executable automatically; loads require `tma_bar_ptr`. * **Typed result types.** Operations whose result type CuTe derives (`localTile`, `slice`, `tmaPartition`, ...) take an `*mlir.Type` assembled with `layoutSpec`, `layoutType`, `memrefType`, `coordTensorType`, diff --git a/dsl/cute/builder.zig b/dsl/cute/builder.zig index d19819fa..8f791f80 100644 --- a/dsl/cute/builder.zig +++ b/dsl/cute/builder.zig @@ -16,6 +16,49 @@ const vector = @import("mlir/dialects/vector"); const mlir = @import("mlir"); const stdx = @import("stdx"); +/// `CUtensorMapDataType`. +pub const TensorMapDataType = enum(u32) { + uint8 = 0, + uint16 = 1, + uint32 = 2, + int32 = 3, + uint64 = 4, + int64 = 5, + float16 = 6, + float32 = 7, + float64 = 8, + bfloat16 = 9, + float32_ftz = 10, + tfloat32 = 11, + tfloat32_ftz = 12, + @"16u4_align8b" = 13, + @"16u4_align16b" = 14, + @"16u6_align16b" = 15, +}; + +/// `CUtensorMapSwizzle`. +pub const TensorMapSwizzle = enum(u32) { none = 0, @"32b" = 1, @"64b" = 2, @"128b" = 3, @"128b_atom_32b" = 4 }; +/// `CUtensorMapL2promotion`. +pub const TensorMapL2Promotion = enum(u32) { none = 0, @"64b" = 1, @"128b" = 2, @"256b" = 3 }; +/// `CUtensorMapInterleave`. +pub const TensorMapInterleave = enum(u32) { none = 0, @"16b" = 1, @"32b" = 2 }; +/// `CUtensorMapFloatOOBfill`. +pub const TensorMapOobFill = enum(u32) { none = 0, nan_request_zero_fma = 1 }; + +pub const TensorMapSpec = struct { + data_type: TensorMapDataType, + rank: u32, + global_dims: [5]u64 = @splat(1), + /// Byte strides of dimensions 1..rank-1. + global_strides: [4]u64 = @splat(0), + box_dims: [5]u32 = @splat(1), + element_strides: [5]u32 = @splat(1), + interleave: TensorMapInterleave = .none, + swizzle: TensorMapSwizzle, + l2_promotion: TensorMapL2Promotion = .none, + oob_fill: TensorMapOobFill = .none, +}; + const dtypes = @import("dtype.zig"); pub const DType = dtypes.DType; const isFloatDtype = dtypes.isFloatDtype; @@ -51,6 +94,8 @@ pub const AlgebraExpr = union(enum) { static: i64, /// A dynamic leaf known to be a multiple of this, 1 when nothing is known. dynamic: i64, + /// A dynamic integer with an explicit bit width and divisibility. + dynamic_int: struct { width: u8, divisible_by: i64 = 1 }, all, basis: ScaledBasis, tuple: []const AlgebraExpr, @@ -466,8 +511,122 @@ pub const TmaFormat = enum { u16, }; -/// Typed description of a tiled TMA atom: the element type, the bits moved -/// per copy, the basis of one box, and the coordinate layout of the tensor. +/// CuTe's 64-bit TMA `.L2::cache_hint` policies for SM90. +pub const CacheHintSm90 = enum(u64) { + evict_normal = 0x1000000000000000, + evict_first = 0x12F0000000000000, + evict_last = 0x14F0000000000000, +}; + +/// CuTe's 64-bit TMA `.L2::cache_hint` policies for SM100. +pub const CacheHintSm100 = enum(u64) { + evict_normal = 0x1000000000000000, + evict_first = 0x12F0000000000000, + evict_last = 0x14F0000000000000, +}; + +pub const CtaGroup = enum(u8) { + one = 1, + two = 2, + + fn toNvvm(self: CtaGroup) nvvm.CtaGroup { + return switch (self) { + .one => .cta_1, + .two => .cta_2, + }; + } +}; +pub const MBarrierScope = nvvm.MBarrierScope; +pub const OperandMajorMode = enum(u1) { k = 0, mn = 1 }; + +/// Number of CTAs cooperating on one instruction, as in CuTe's CtaGroup.value. +pub fn ctaCount(group: CtaGroup) u8 { + return @backingInt(group); +} + +pub const Tcgen05SmemSwizzle = enum(u3) { + none = 0, + swizzle_128b_atom_32b = 1, + swizzle_128b = 2, + swizzle_64b = 4, + swizzle_32b = 6, +}; + +/// CuTe's packed SM100 shared-memory descriptor. All offsets are in bytes. +pub const Tcgen05SmemDesc = struct { + pub const Options = struct { + leading_byte_offset: u32 = 0, + stride_byte_offset: u32 = 0, + version: u2 = 1, + base_offset: u3 = 0, + layout: Tcgen05SmemSwizzle = .none, + }; + + pub fn build(b: *Builder, start_address: Value, opts: Options) Value { + std.debug.assert(opts.leading_byte_offset % 16 == 0 and opts.leading_byte_offset < 0x40000); + std.debug.assert(opts.stride_byte_offset % 16 == 0 and opts.stride_byte_offset < 0x40000); + const fields = (@as(u64, opts.leading_byte_offset >> 4) << 16) | + (@as(u64, opts.stride_byte_offset >> 4) << 32) | + (@as(u64, opts.version) << 46) | + (@as(u64, opts.base_offset) << 49) | + (@as(u64, @backingInt(opts.layout)) << 61); + const address = if (start_address.isInt()) start_address else b.ptrToInt(start_address, .i32); + return address.to(.i64).shrLogical(4).bitAnd(0x3fff).bitOr(b.constant(.i64, @as(i64, @bitCast(fields)))); + } +}; + +/// Dense SM100 instruction descriptor; the MMA kind disambiguates dtype encodings. +pub const Tcgen05InstrDesc = struct { + pub const Options = struct { + a_dtype: DType, + b_dtype: DType, + c_dtype: DType = .f32, + m_dim: u16, + n_dim: u16, + a_major: OperandMajorMode = .k, + b_major: OperandMajorMode = .k, + a_negate: bool = false, + b_negate: bool = false, + }; + + pub fn build(comptime opts: Options) u32 { + comptime std.debug.assert(opts.m_dim > 0 and opts.m_dim <= 256 and opts.m_dim % 16 == 0); + comptime std.debug.assert(opts.n_dim > 0 and opts.n_dim <= 256 and opts.n_dim % 8 == 0); + const half_inputs = (opts.a_dtype == .f16 or opts.a_dtype == .bf16) and (opts.b_dtype == .f16 or opts.b_dtype == .bf16); + const fp8_inputs = (opts.a_dtype == .f8e4m3fn or opts.a_dtype == .f8e5m2) and (opts.b_dtype == .f8e4m3fn or opts.b_dtype == .f8e5m2); + const valid = (half_inputs and opts.c_dtype == .f32) or + (opts.a_dtype == .f16 and opts.b_dtype == .f16 and opts.c_dtype == .f16) or + (opts.a_dtype == .f32 and opts.b_dtype == .f32 and opts.c_dtype == .f32) or + (fp8_inputs and (opts.c_dtype == .f16 or opts.c_dtype == .f32)) or + (opts.a_dtype == .i8 and opts.b_dtype == .i8 and opts.c_dtype == .i32); + if (!valid) @compileError("unsupported dense tcgen05 dtype combination"); + const c: u32 = switch (opts.c_dtype) { + .f16 => 0, + .f32 => 1, + .i32 => 2, + else => @compileError("unsupported tcgen05 accumulator dtype"), + }; + return (c << 4) | (inputFormat(opts.a_dtype) << 7) | (inputFormat(opts.b_dtype) << 10) | + (@as(u32, @intFromBool(opts.a_negate)) << 13) | + (@as(u32, @intFromBool(opts.b_negate)) << 14) | + (@as(u32, @backingInt(opts.a_major)) << 15) | + (@as(u32, @backingInt(opts.b_major)) << 16) | + (@as(u32, opts.n_dim >> 3) << 17) | (@as(u32, opts.m_dim >> 4) << 24); + } + + fn inputFormat(comptime dtype: DType) u32 { + return switch (dtype) { + .f16, .f8e4m3fn => 0, + .bf16, .f8e5m2, .i8 => 1, + .f32 => 2, + .f4e2m1fn => 5, + else => @compileError("unsupported tcgen05 input dtype"), + }; + } +}; + +/// Typed description of a non-multicast tiled TMA atom: the element type, +/// bits moved per copy, basis of one box, and coordinate layout of the tensor. pub const TmaConfig = struct { dtype: DType, copy_bits: u32, @@ -475,18 +634,29 @@ pub const TmaConfig = struct { coordinate_layout: LayoutSpec, coordinate_rank: u8, format: TmaFormat = .default, + /// Load completion group: with `.two`, each CTA's load completes on + /// the mbarrier of the pair's even CTA. Ignored by TMA stores. + cta_group: CtaGroup = .one, + mode: enum { tiled, gather4 } = .tiled, }; +pub const CopyBulkTensorTileG2SOp = struct { cta_group: CtaGroup = .one }; +pub const CopyBulkTensor2DGather4G2SOp = struct { cta_group: CtaGroup = .one }; +pub const CopyBulkTensorTileS2GOp = struct {}; + pub const BlockScaledMma = struct { atom: Atom, tiled: Atom, }; -/// SM100 block-scaled MMA instruction: one CTA computes a 128 x `n` tile, -/// FP4 weights as A. The instruction `kind` fixes B, the scales and K. +/// SS, K-major SM100 block-scaled MMA instruction with FP4 weights as A. +/// The instruction `kind` fixes B, the scales and K. pub const BlockScaledMmaConfig = struct { kind: Kind = .mxfp4_mxfp8, + /// M is 128 for one CTA, or 128 or 256 for a CTA pair. + m: u16 = 128, n: u16, + cta_group: CtaGroup = .one, pub const Kind = enum { /// MXFP4 weights x MXFP8 activations: E8M0 scales every 32 @@ -524,8 +694,23 @@ pub const BlockScaledMmaConfig = struct { .nvfp4_nvfp4 => 16, }; } + + fn validate(self: BlockScaledMmaConfig) void { + const m_ok = self.m == 128 or (self.cta_group == .two and self.m == 256); + const n_multiple: u16 = if (self.cta_group == .two) 16 else 8; + if (!m_ok or self.n < n_multiple or self.n > 256 or self.n % n_multiple != 0) { + std.debug.panic("unsupported SM100 block-scaled MMA shape {d}x{d}x{d} on {d} CTAs", .{ self.m, self.n, self.k(), ctaCount(self.cta_group) }); + } + } }; +fn tmaLoadKind(config: TmaConfig) cute_nvgpu.TiledTmaLoad { + return switch (config.cta_group) { + .one => .sm_90, + .two => .sm_100_2sm, + }; +} + fn tmaDataFormat(format: TmaFormat, dtype: DType) cute_nvgpu.TmaDataFormat { return switch (format) { .u4 => .U4, @@ -573,6 +758,7 @@ pub const ArgSpec = struct { pub const MlirTypeSpec = struct { type_: *const mlir.Type, grid_constant: bool = false, + cuda_grid_constant: bool = false, }; pub const TensorSpec = struct { @@ -681,6 +867,7 @@ pub const Builder = struct { .mlir_type => .{ .mlir_type = .{ .type_ = inner.type_, .grid_constant = if (@hasField(@TypeOf(inner), "grid_constant")) inner.grid_constant else false, + .cuda_grid_constant = if (@hasField(@TypeOf(inner), "cuda_grid_constant")) inner.cuda_grid_constant else false, } }, } }; } @@ -707,7 +894,7 @@ pub const Builder = struct { return @Struct(.auto, null, &names, &types, &@splat(.{})); } - fn declareArgsLow(self: *Builder, args: []const ArgSpec) FinishError!void { + pub fn declareArgsLow(self: *Builder, args: []const ArgSpec) FinishError!void { std.debug.assert(self.entry_block == null); const ctx = self.ctx; const scratch = self.arena.allocator(); @@ -729,7 +916,7 @@ pub const Builder = struct { var any_arg_attr = false; for (arg_attrs, args) |*attr, a| { attr.* = switch (a.kind) { - .mlir_type => |typed| if (typed.grid_constant) .dict(ctx, &.{cute_nvgpu.gridConstant(ctx)}) else empty_dict, + .mlir_type => |typed| if (typed.cuda_grid_constant) .dict(ctx, &.{.named(ctx, "cuda.grid_constant", .unit(ctx))}) else if (typed.grid_constant) .dict(ctx, &.{cute_nvgpu.gridConstant(ctx)}) else empty_dict, else => empty_dict, }; if (attr.* != empty_dict) any_arg_attr = true; @@ -941,6 +1128,75 @@ pub const Builder = struct { })); } + /// `!cuda.tensor_map`: the kernel argument type of an explicitly encoded map. + pub fn tensorMapType(self: *Builder) *const mlir.Type { + return (cuda.ops.TensorMapType.get(self.ctx, .{}) catch unreachable).type_(); + } + + /// A host stack array of `values` (u64 -> i64 slots, u32 -> i32 slots): a `cuuint64_t*` / + /// `cuuint32_t*` argument of `cuTensorMapEncodeTiled`. + fn tensorMapHostArray(self: *Builder, comptime T: type, values: []const T) Value { + const ctx = self.ctx; + const elem: *const mlir.Type = if (T == u64) .int(ctx, .i64) else .int(ctx, .i32); + const ptr_type = llvm.pointerType(ctx, llvm.NVPTXAddressSpace.generic); + const array = self.emit(mlir.Operation.make(ctx, "llvm.alloca", .{ + .operands = .{ .flat = &.{self.constant(.i32, @as(i32, @intCast(@max(values.len, 1)))).inner} }, + .results = .{ .flat = &.{ptr_type} }, + .attributes = &.{ .named(ctx, "elem_type", .typeAttr(elem)), .named(ctx, "alignment", .int(ctx, .i64, 8)) }, + .location = self.loc(), + })); + for (values, 0..) |v, i| { + const slot = self.emit(mlir.Operation.make(ctx, "llvm.getelementptr", .{ + .operands = .{ .flat = &.{array.inner} }, + .results = .{ .flat = &.{ptr_type} }, + .attributes = &.{ .named(ctx, "elem_type", .typeAttr(elem)), .named(ctx, "rawConstantIndices", .denseArray(ctx, .i32, &[_]i32{@intCast(i)})) }, + .location = self.loc(), + })); + const value = if (T == u64) self.constant(.i64, @as(i64, @intCast(v))) else self.constant(.i32, @as(i32, @intCast(v))); + self.emitVoid(llvm.store(ctx, value.inner, slot.inner, .{ .alignment = @sizeOf(T) }, self.loc())); + } + return array; + } + + /// Encode a tiled map in a host function and return both the map and CUDA status. + /// The caller supplies the final device address and handles the returned status. + pub fn encodeTensorMap(self: *Builder, spec: TensorMapSpec, address: Value) struct { map: Value, result: Value } { + const ctx = self.ctx; + // The `cuda` dialect's address operand is a `ptr` dialect pointer, which zml does not register. + ctx.setAllowUnregisteredDialects(true); + const addr = self.emit(cuda.cast(ctx, (mlir.Type.parse(self.ctx, "!ptr.ptr<#ptr.generic_space>") catch @panic("invalid tensor-map IR")), address.inner, self.loc())); + const results = self.emitResults(2, cuda.ops.tensor_map_encode_tiled( + ctx, + self.constant(.i32, @as(i32, @intCast(@backingInt(spec.data_type)))).inner, + self.constant(.i32, @as(i32, @intCast(spec.rank))).inner, + addr.inner, + self.tensorMapHostArray(u64, spec.global_dims[0..spec.rank]).inner, + self.tensorMapHostArray(u64, spec.global_strides[0 .. spec.rank - 1]).inner, + self.tensorMapHostArray(u32, spec.box_dims[0..spec.rank]).inner, + self.tensorMapHostArray(u32, spec.element_strides[0..spec.rank]).inner, + self.constant(.i32, @as(i32, @intCast(@backingInt(spec.interleave)))).inner, + self.constant(.i32, @as(i32, @intCast(@backingInt(spec.swizzle)))).inner, + self.constant(.i32, @as(i32, @intCast(@backingInt(spec.l2_promotion)))).inner, + self.constant(.i32, @as(i32, @intCast(@backingInt(spec.oob_fill)))).inner, + cuda.resultType(ctx), + self.tensorMapType(), + self.loc(), + )); + return .{ .map = results[1], .result = results[0] }; + } + + /// Device side: the generic address (i64) of the `!cuda.tensor_map` kernel argument `arg` + /// (`&tensor_map` for a `__grid_constant__` parameter). + pub fn tensorMapAddress(self: *Builder, descriptor: Value) Value { + const ctx = self.ctx; + const llvm_ptr = self.emit(cuda.cast(ctx, llvm.pointerType(ctx, llvm.NVPTXAddressSpace.generic), descriptor.inner, self.loc())); + return self.emit(mlir.Operation.make(ctx, "llvm.ptrtoint", .{ + .operands = .{ .flat = &.{llvm_ptr.inner} }, + .results = .{ .flat = &.{DType.i64.toMlir(ctx)} }, + .location = self.loc(), + })); + } + pub fn cudaResultStatus(self: *Builder, result: Value) Value { return self.emit(cuda.cast(self.ctx, .int(self.ctx, .i32), result.inner, self.loc())); } @@ -1019,7 +1275,8 @@ pub const Builder = struct { /// `cute.ptrtoint`: the integer address of a pointer (`i32` for shared /// memory, `i64` for global memory), for hand-written PTX operands. pub fn ptrToInt(self: *Builder, pointer: anytype, dt: DType) Value { - return self.emit(cute_ir.ptrtoint(self.ctx, self.asValue(pointer).inner, dt.toMlir(self.ctx), self.loc())); + const ptr = self.asValue(pointer); + return self.emit(if (ptr.isPtr()) cute_ir.ptrtoint(self.ctx, ptr.inner, dt.toMlir(self.ctx), self.loc()) else llvm.ptrtoint(self.ctx, ptr.inner, dt.toMlir(self.ctx), self.loc())); } /// Reinterpret a pointer without changing its address. This is the typed @@ -1074,6 +1331,7 @@ pub const Builder = struct { const result = switch (e) { .static => |v| if (@hasDecl(A, "getStatic")) A.getStatic(self.ctx, v) else std.debug.panic("{s} has no integer leaf", .{@typeName(A)}), .dynamic => |divisibility| if (@hasDecl(A, "getDynamic")) A.getDynamic(self.ctx, 32, divisibility) else std.debug.panic("{s} has no dynamic leaf", .{@typeName(A)}), + .dynamic_int => |v| if (@hasDecl(A, "getDynamic")) A.getDynamic(self.ctx, v.width, v.divisible_by) else std.debug.panic("{s} has no dynamic leaf", .{@typeName(A)}), .all => if (@hasDecl(A, "getUnderscore")) A.getUnderscore(self.ctx) else std.debug.panic("{s} has no `_` leaf", .{@typeName(A)}), .basis => |b| if (@hasDecl(A, "getBasis")) blk: { const modes = self.arena.allocator().alloc(i32, b.mode_count) catch @panic("OOM"); @@ -1238,8 +1496,59 @@ pub const Builder = struct { // ==================== cute.core ==================== - /// `cute.make_layout(shape, stride=...)`, static only. - pub fn makeLayout(self: *Builder, shape: []const i64, opts: LayoutOpts) Layout { + fn layoutResult(comptime Shape: type) type { + return switch (@typeInfo(Shape)) { + .pointer => Layout, + else => View, + }; + } + + /// `cute.make_layout(shape, stride=...)`. Slices keep the tracked static + /// Layout API; nested tuples accept mixed integer and SSA shape/stride leaves. + pub fn makeLayout(self: *Builder, shape: anytype, opts: anytype) layoutResult(@TypeOf(shape)) { + if (comptime layoutResult(@TypeOf(shape)) == Layout) return self.makeStaticLayout(shape, .{ + .stride = if (@hasField(@TypeOf(opts), "stride")) opts.stride else null, + }); + var shape_values: stdx.BoundedArray(*const mlir.Value, MAX_RANK) = .empty; + var stride_values: stdx.BoundedArray(*const mlir.Value, MAX_RANK) = .empty; + const shape_expr = self.coordExpr(&shape_values, shape); + const stride_expr = if (@hasField(@TypeOf(opts), "stride") and @TypeOf(opts.stride) != @TypeOf(null)) self.coordExpr(&stride_values, opts.stride) else blk: { + var stride: AlgebraExpr = .{ .static = 1 }; + var stride_value: ?Value = null; + var index: usize = 0; + break :blk self.compactLayoutStride(shape_expr, shape_values.constSlice(), &index, &stride, &stride_value, &stride_values); + }; + const shape_value = self.emit(cute_ir.make_shape(self.ctx, shape_values.constSlice(), self.shapeType(shape_expr), self.loc())); + const stride_type = (cute_ir.StrideType.get(self.ctx, .{ .attr = self.algebraAttr(cute_ir.StrideAttr, stride_expr).attribute() }) catch @panic("invalid stride")).type_(); + const stride_value = self.emit(cute_ir.make_stride(self.ctx, stride_values.constSlice(), stride_type, self.loc())); + const result = self.emit(cute_ir.make_layout(self.ctx, self.layoutType(.{ .shape = shape_expr, .stride = stride_expr }), .{ .shape = shape_value.inner, .stride = stride_value.inner }, self.loc())); + return .{ .inner = result.inner, .kernel = self }; + } + + fn compactLayoutStride(self: *Builder, shape: AlgebraExpr, values: []const *const mlir.Value, index: *usize, stride: *AlgebraExpr, stride_value: *?Value, dynamic: *stdx.BoundedArray(*const mlir.Value, MAX_RANK)) AlgebraExpr { + if (shape == .tuple) { + const elements = self.arena.allocator().alloc(AlgebraExpr, shape.tuple.len) catch @panic("OOM"); + for (shape.tuple, elements) |extent, *element| element.* = self.compactLayoutStride(extent, values, index, stride, stride_value, dynamic); + return .{ .tuple = elements }; + } + if (shape == .static and shape.static == 1) return .{ .static = 0 }; + const result = stride.*; + if (stride_value.*) |v| dynamic.append(v.inner) catch @panic("too many stride leaves"); + if (shape == .static and stride.* == .static) { + stride.static *= shape.static; + } else { + const extent = if (shape == .dynamic or shape == .dynamic_int) blk: { + const v: Value = .{ .inner = values[index.*], .kernel = self }; + index.* += 1; + break :blk v; + } else self.lift(shape.static); + stride_value.* = (stride_value.* orelse self.lift(stride.static)).mul(extent); + stride.* = dynamicExpr(stride_value.*.?); + } + return result; + } + + fn makeStaticLayout(self: *Builder, shape: []const i64, opts: LayoutOpts) Layout { if (shape.len == 0 or shape.len > MAX_RANK) std.debug.panic("makeLayout: rank {d} not in 1..{d}", .{ shape.len, MAX_RANK }); const stride = opts.stride orelse self.leftMost(shape); if (stride.len != shape.len) std.debug.panic("makeLayout: stride rank {d} != shape rank {d}", .{ stride.len, shape.len }); @@ -1262,8 +1571,8 @@ pub const Builder = struct { fn coordExpr(self: *Builder, dyn: *stdx.BoundedArray(*const mlir.Value, MAX_RANK), coord: anytype) AlgebraExpr { const T = @TypeOf(coord); if (T == Value) { - dyn.append(self.lift(coord).to(.i32).inner) catch @panic("makeCoord: too many dynamic leaves"); - return .{ .dynamic = 1 }; + dyn.append(coord.inner) catch @panic("makeCoord: too many dynamic leaves"); + return dynamicExpr(coord); } if (T == AlgebraToken) return switch (coord) { .all => .all, @@ -1282,6 +1591,12 @@ pub const Builder = struct { } } + fn dynamicExpr(value: Value) AlgebraExpr { + std.debug.assert(value.isInt()); + const width = dtypeBitwidth(value.dtype()); + return if (width == 32) .{ .dynamic = 1 } else .{ .dynamic_int = .{ .width = @intCast(width) } }; + } + /// A `!cute.coord` from nested tuples of `Value`s, integers, and `.all` /// placeholders. Integer leaves stay static in the dependent CuTe type. pub fn makeCoord(self: *Builder, coord: anytype) *const mlir.Value { @@ -1516,20 +1831,27 @@ pub const Builder = struct { /// `projection`, like Python's `proj`, selects the modes of `tile` that /// apply, such as `b.expr(.{ 1, AlgebraToken.all, 1 })`. - pub fn localTile(self: *Builder, input: anytype, tile: anytype, coord: anytype, result_type: *const mlir.Type, projection: ?AlgebraExpr) View { + pub fn localTile(self: *Builder, input: anytype, tile: anytype, coord: anytype, result_type: ?*const mlir.Type, projection: ?AlgebraExpr) View { + if (result_type == null) return self.inferView("cute.local_tile", &.{ self.asValue(input).inner, self.asValue(tile).inner, if (@TypeOf(coord) == View) coord.inner else self.makeCoord(coord) }, if (projection) |p| &.{.named(self.ctx, "proj", self.algebraAttr(cute_ir.CoordAttr, p).attribute())} else &.{}); const coord_value = if (@TypeOf(coord) == View) coord.inner else self.makeCoord(coord); const result = self.emit(cute_ir.local_tile( self.ctx, self.asValue(input).inner, self.asValue(tile).inner, coord_value, - result_type, + result_type.?, .{ .proj = if (projection) |p| self.algebraAttr(cute_ir.CoordAttr, p).attribute() else null }, self.loc(), )); return .{ .inner = result.inner, .kernel = self }; } + // A null result type on the algebra operations below requests dialect inference. + fn inferView(self: *Builder, name: []const u8, operands: []const *const mlir.Value, attrs: []const mlir.NamedAttribute) View { + const value = self.emit(mlir.Operation.make(self.ctx, name, .{ .operands = .{ .flat = operands }, .attributes = attrs, .result_type_inference = true, .location = self.loc() })); + return .{ .inner = value.inner, .kernel = self }; + } + pub fn localPartition(self: *Builder, input: anytype, tiler: anytype, index: anytype, result_type: *const mlir.Type) View { const result = self.emit(cute_ir.local_partition( self.ctx, @@ -1542,12 +1864,13 @@ pub const Builder = struct { return .{ .inner = result.inner, .kernel = self }; } - pub fn slice(self: *Builder, input: anytype, coord: anytype, result_type: *const mlir.Type) View { + pub fn slice(self: *Builder, input: anytype, coord: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.slice", &.{ self.asValue(input).inner, self.makeCoord(coord) }, &.{}); const result = self.emit(cute_ir.slice( self.ctx, self.asValue(input).inner, self.makeCoord(coord), - result_type, + result_type.?, self.loc(), )); return .{ .inner = result.inner, .kernel = self }; @@ -1566,13 +1889,17 @@ pub const Builder = struct { (cute_ir.MemRefType.get(self.ctx, .{ .ptr = memref.ptr_(), .layout = layout_value.type_() }) catch @panic("invalid regrouped memref")).type_() else if (input_type.isA(cute_ir.CoordTensorType)) |coords| (cute_ir.CoordTensorType.get(self.ctx, .{ .arithTuple = coords.arithTuple(), .layout = layout_value.type_() }) catch @panic("invalid regrouped coordinate tensor")).type_() + else if (input_type.isA(cute_nvgpu.SmemDescViewType)) |desc| + (cute_nvgpu.SmemDescViewType.get(self.ctx, .{ .desc = desc.desc(), .layout = layout_value.type_().isA(cute_ir.LayoutType).?.attr() }) catch @panic("invalid regrouped descriptor")).type_() else - std.debug.panic("withLayout: expected memref or coordinate tensor, got {f}", .{input_type}); + std.debug.panic("withLayout: expected memref, coordinate tensor or shared-memory descriptor view, got {f}", .{input_type}); const iterator_type: *const mlir.Type = if (input_type.isA(cute_ir.MemRefType)) |memref| memref.ptr_() else if (input_type.isA(cute_ir.CoordTensorType)) |coords| (cute_ir.ArithTupleIteratorType.get(self.ctx, .{ .arithTuple = coords.arithTuple() }) catch @panic("invalid arithmetic tuple iterator")).type_() + else if (input_type.isA(cute_nvgpu.SmemDescViewType)) |desc| + desc.desc() else unreachable; const iterator = self.emit(cute_ir.get_iter(self.ctx, value.inner, iterator_type, self.loc())); @@ -1580,11 +1907,12 @@ pub const Builder = struct { return .{ .inner = result.inner, .kernel = self }; } - pub fn groupModes(self: *Builder, input: anytype, begin: u32, end: u32, result_type: *const mlir.Type) View { + pub fn groupModes(self: *Builder, input: anytype, begin: u32, end: u32, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.group_modes", &.{self.asValue(input).inner}, &.{ .named(self.ctx, "begin", .int(self.ctx, .i32, begin)), .named(self.ctx, "end", .int(self.ctx, .i32, end)) }); const result = self.emit(cute_ir.group_modes( self.ctx, self.asValue(input).inner, - result_type, + result_type.?, .int(self.ctx, .i32, begin), .int(self.ctx, .i32, end), self.loc(), @@ -1603,12 +1931,13 @@ pub const Builder = struct { return .{ .inner = result.inner, .kernel = self }; } - pub fn flatDivide(self: *Builder, input: anytype, tiler: anytype, result_type: *const mlir.Type) View { + pub fn flatDivide(self: *Builder, input: anytype, tiler: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.flat_divide", &.{ self.asValue(input).inner, self.asValue(tiler).inner }, &.{}); const result = self.emit(cute_ir.flat_divide( self.ctx, self.asValue(input).inner, self.asValue(tiler).inner, - result_type, + result_type.?, self.loc(), )); return .{ .inner = result.inner, .kernel = self }; @@ -1625,39 +1954,44 @@ pub const Builder = struct { return .{ .inner = result.inner, .kernel = self }; } - pub fn filterZeros(self: *Builder, input: anytype, target_profile: ?View, result_type: *const mlir.Type) View { + pub fn filterZeros(self: *Builder, input: anytype, target_profile: ?View, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.filter_zeros", if (target_profile) |target| &.{ self.asValue(input).inner, target.inner } else &.{self.asValue(input).inner}, &.{}); const result = self.emit(cute_ir.filter_zeros( self.ctx, self.asValue(input).inner, - result_type, + result_type.?, .{ .target_profile = if (target_profile) |target| target.inner else null }, self.loc(), )); return .{ .inner = result.inner, .kernel = self }; } - pub fn getLayout(self: *Builder, input: anytype, result_type: *const mlir.Type) View { - const result = self.emit(cute_ir.get_layout(self.ctx, self.asValue(input).inner, result_type, self.loc())); + pub fn getLayout(self: *Builder, input: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.get_layout", &.{self.asValue(input).inner}, &.{}); + const result = self.emit(cute_ir.get_layout(self.ctx, self.asValue(input).inner, result_type.?, self.loc())); return .{ .inner = result.inner, .kernel = self }; } /// `cute.get_iter`: the iterator of `input`, of type `result_type`. - pub fn getIter(self: *Builder, input: anytype, result_type: *const mlir.Type) View { - const result = self.emit(cute_ir.get_iter(self.ctx, self.asValue(input).inner, result_type, self.loc())); + pub fn getIter(self: *Builder, input: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.get_iter", &.{self.asValue(input).inner}, &.{}); + const result = self.emit(cute_ir.get_iter(self.ctx, self.asValue(input).inner, result_type.?, self.loc())); return .{ .inner = result.inner, .kernel = self }; } - pub fn getShape(self: *Builder, input: anytype, result_type: *const mlir.Type) View { - const result = self.emit(cute_ir.get_shape(self.ctx, self.asValue(input).inner, result_type, self.loc())); + pub fn getShape(self: *Builder, input: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.get_shape", &.{self.asValue(input).inner}, &.{}); + const result = self.emit(cute_ir.get_shape(self.ctx, self.asValue(input).inner, result_type.?, self.loc())); return .{ .inner = result.inner, .kernel = self }; } - pub fn addOffset(self: *Builder, input: anytype, offset: anytype, result_type: *const mlir.Type) View { + pub fn addOffset(self: *Builder, input: anytype, offset: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.add_offset", &.{ self.asValue(input).inner, self.asValue(offset).inner }, &.{}); const result = self.emit(cute_ir.add_offset( self.ctx, self.asValue(input).inner, self.asValue(offset).inner, - result_type, + result_type.?, self.loc(), )); return .{ .inner = result.inner, .kernel = self }; @@ -1694,44 +2028,94 @@ pub const Builder = struct { return .{ .inner = result.inner, .kernel = self }; } - pub fn tiledCopyRetile(self: *Builder, tiled: Atom, input: anytype, result_type: *const mlir.Type) View { + pub fn tiledCopyRetile(self: *Builder, tiled: Atom, input: anytype, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.tiled.copy.retile", &.{ tiled.inner, self.asValue(input).inner }, &.{}); const result = self.emit(cute_ir.tiled_copy_retile( self.ctx, tiled.inner, self.asValue(input).inner, - result_type, + result_type.?, self.loc(), )); return .{ .inner = result.inner, .kernel = self }; } - pub fn tiledMmaPartition(self: *Builder, tiled: Atom, input: anytype, coord: anytype, operand_id: u2, result_type: *const mlir.Type) View { + pub fn tiledMmaPartition(self: *Builder, tiled: Atom, input: anytype, coord: anytype, operand_id: u2, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.tiled.mma.partition", &.{ tiled.inner, self.asValue(input).inner, self.makeCoord(coord) }, &.{.named(self.ctx, "operand_id", .int(self.ctx, .i32, operand_id))}); const result = self.emit(cute_ir.tiled_mma_partition( self.ctx, tiled.inner, self.asValue(input).inner, self.makeCoord(coord), - result_type, + result_type.?, .int(self.ctx, .i32, operand_id), self.loc(), )); return .{ .inner = result.inner, .kernel = self }; } - pub fn mmaMakeFragment(self: *Builder, tiled: Atom, input: anytype, operand_id: u2, result_type: *const mlir.Type) View { + pub fn mmaMakeFragment(self: *Builder, tiled: Atom, input: anytype, operand_id: u2, result_type: ?*const mlir.Type) View { + if (result_type == null) return self.inferView("cute.mma.make_fragment", &.{ tiled.inner, self.asValue(input).inner }, &.{.named(self.ctx, "operand_id", .int(self.ctx, .i32, operand_id))}); const result = self.emit(cute_ir.mma_make_fragment( self.ctx, tiled.inner, self.asValue(input).inner, - result_type, + result_type.?, .int(self.ctx, .i32, operand_id), self.loc(), )); return .{ .inner = result.inner, .kernel = self }; } - pub fn copy(self: *Builder, atom: Atom, src: anytype, dst: anytype, pred: ?Value) void { - self.emitVoid(cute_ir.copy(self.ctx, atom.inner, self.innerSlice(src), self.innerSlice(dst), .{ .pred = if (pred) |p| p.inner else null }, self.loc())); + pub const CopyOptions = struct { pred: ?Value = null, tma_bar_ptr: ?Value = null, cache_policy: ?Value = null }; + + pub fn copy(self: *Builder, atom: Atom, src: anytype, dst: anytype, opts: CopyOptions) void { + var exec = atom; + if (atom.type_().isA(cute_nvgpu.CopyAtomNonExecTiledTmaStoreType)) |ty| { + const exec_type = (cute_nvgpu.CopyAtomTmaStoreType.get(self.ctx, .{ + .valType = ty.valType(), + .copyBits = ty.copyBits(), + .mode = .tiled, + .gStride = self.emptyStrideType(), + .tmaGbasis = ty.tmaGbasis(), + }) catch @panic("invalid executable TMA store atom")).type_(); + exec = self.makeExecTma(atom, exec_type); + } + if (opts.tma_bar_ptr) |mbar| { + if (atom.type_().isA(cute_nvgpu.CopyAtomNonExec2DGather4TmaLoadType)) |ty| { + const exec_type = (cute_nvgpu.CopyAtomTmaLoadType.get(self.ctx, .{ + .valType = ty.valType(), + .copyBits = ty.copyBits(), + .mode = .gather4, + .numCta = if (ty.kind() == .sm_100_2sm) 2 else 1, + .gStride = self.emptyStrideType(), + .tmaGbasis = ty.tmaGbasis(), + }) catch @panic("invalid executable gather atom")).type_(); + exec = self.makeExecTma(atom, exec_type); + } else if (atom.type_().isA(cute_nvgpu.CopyAtomNonExecTiledTmaLoadType)) |ty| { + const exec_type = (cute_nvgpu.CopyAtomTmaLoadType.get(self.ctx, .{ + .valType = ty.valType(), + .copyBits = ty.copyBits(), + .mode = .tiled, + .numCta = if (ty.kind() == .sm_100_2sm) 2 else 1, + .gStride = self.emptyStrideType(), + .tmaGbasis = ty.tmaGbasis(), + }) catch @panic("invalid executable TMA atom")).type_(); + exec = self.makeExecTma(atom, exec_type); + } + exec = self.setTmaBarrier(exec, mbar); + } + if (opts.cache_policy) |policy| { + const field = if (exec.type_().isA(cute_nvgpu.CopyAtomTmaLoadType) != null) + (cute_nvgpu.AtomCopyFieldTmaLoadAttr.get(self.ctx, .{ .value = .cache_policy }) catch unreachable).attribute() + else if (exec.type_().isA(cute_nvgpu.CopyAtomTmaStoreType) != null) + (cute_nvgpu.AtomCopyFieldTmaStoreAttr.get(self.ctx, .{ .value = .cache_policy }) catch unreachable).attribute() + else + @panic("cache_policy requires an executable TMA copy atom"); + const updated = self.emit(cute_nvgpu.atom_set_value(self.ctx, exec.inner, policy.to(.i64).inner, exec.type_(), field, self.loc())); + exec = .{ .inner = updated.inner, .kernel = self }; + } + self.emitVoid(cute_ir.copy(self.ctx, exec.inner, self.innerSlice(src), self.innerSlice(dst), .{ .pred = if (opts.pred) |p| p.inner else null }, self.loc())); } pub fn gemm(self: *Builder, tiled: Atom, d: anytype, a: anytype, rhs: anytype, c: anytype) void { @@ -1762,9 +2146,67 @@ pub const Builder = struct { // the kernel receives them as grid-constant arguments and makes them // executable before issuing copies. + pub const TmaAtomOptions = struct { gmem_coord_tensor: ?View = null }; + + fn inferValue(self: *Builder, name: []const u8, operands: []const *const mlir.Value, attributes: []const mlir.NamedAttribute) Value { + return self.emit(mlir.Operation.make(self.ctx, name, .{ + .operands = .{ .flat = operands }, + .attributes = attributes, + .result_type_inference = true, + .location = self.loc(), + })); + } + + fn valueShape(self: *Builder, value: Value) Value { + if (value.type_().isA(cute_ir.ShapeType) != null) return value; + return self.inferValue("cute.get_shape", &.{value.inner}, &.{}); + } + + fn shapeRank(shape: Value) usize { + return shape.type_().isA(cute_ir.ShapeType).?.attr().isA(cute_ir.ShapeAttr).?.rank(); + } + + /// Common CuTe TMA atom constructor. Gather uses the index tensor's layout + /// at construction and passes data/index tensors together to partition/copy. + pub fn makeTiledTmaAtom(self: *Builder, op: anytype, gmem: anytype, smem_layout: anytype, cta_tiler: anytype, config: TmaConfig, opts: TmaAtomOptions) TmaDescriptor { + const Tiler = @TypeOf(cta_tiler); + const tiler = if (Tiler == Value or Tiler == View or Tiler == Layout) + self.asValue(cta_tiler) + else if (@typeInfo(Tiler) == .pointer) + self.makeShape(self.dimsExpr(cta_tiler)).value() + else + self.makeShape(cta_tiler).value(); + var smem = self.asValue(smem_layout); + const tiler_rank = shapeRank(self.valueShape(tiler)); + const smem_rank = shapeRank(self.valueShape(smem)); + std.debug.assert(smem_rank == tiler_rank or smem_rank == tiler_rank + 1); + if (smem_rank == tiler_rank + 1) { + const modes = self.arena.allocator().alloc(i32, tiler_rank) catch @panic("OOM"); + for (modes, 0..) |*mode, i| mode.* = @intCast(i); + smem = self.inferValue("cute.select", &.{smem.inner}, &.{.named(self.ctx, "mode", .denseArray(self.ctx, .i32, modes))}); + } + const shape = self.valueShape(self.asValue(gmem)); + const identity = self.inferValue("cute.make_identity_layout", &.{shape.inner}, &.{}); + const cta_map = self.inferValue("cute.composition", &.{ identity.inner, tiler.inner }, &.{}); + var cfg = config; + cfg.mode = .tiled; + if (@TypeOf(op) == CopyBulkTensorTileS2GOp) return self.makeTiledTmaStoreAtom(gmem, smem, cta_map, cfg); + cfg.cta_group = op.cta_group; + if (@TypeOf(op) == CopyBulkTensorTileG2SOp) return self.makeTiledTmaLoadAtom(gmem, smem, cta_map, cfg); + if (@TypeOf(op) != CopyBulkTensor2DGather4G2SOp) @compileError("unsupported TMA copy operation"); + cfg.mode = .gather4; + const indices = opts.gmem_coord_tensor orelse @panic("gather4 requires gmem_coord_tensor"); + const index_type = indices.type_().isA(cute_ir.MemRefType) orelse @panic("gather4 indices must be a tensor"); + const layout = self.emit(cute_ir.get_layout(self.ctx, indices.inner, index_type.layout(), self.loc())); + const kind: cute_nvgpu.GatherScatterTmaLoad = if (cfg.cta_group == .two) .sm_100_2sm else .sm_100; + const results = self.emitResults(2, cute_nvgpu.atom_make_non_exec_2d_gather4_tma_load(self.ctx, self.asValue(gmem).inner, layout.inner, smem.inner, cta_map.inner, self.tmaLoadAtomType(cfg), self.coordTensorType(cfg.coordinate_rank, cfg.coordinate_layout), (cute_nvgpu.GatherScatterTmaLoadAttr.get(self.ctx, .{ .value = kind }) catch unreachable).attribute(), .{ .num_multicast = .int(self.ctx, .i32, 1), .tma_format = tmaFormatAttribute(self, cfg) }, self.loc())); + return .{ .atom = .{ .inner = results[0].inner, .kernel = self }, .tensor = .{ .inner = results[1].inner, .kernel = self } }; + } + /// Build the host-visible TMA descriptor and its coordinate tensor, whose /// types are assembled from `config`. pub fn makeTiledTmaLoadAtom(self: *Builder, gmem: anytype, smem_layout: anytype, cta_map: anytype, config: TmaConfig) TmaDescriptor { + std.debug.assert(config.mode == .tiled); const results = self.emitResults(2, cute_nvgpu.atom_make_non_exec_tiled_tma_load( self.ctx, self.asValue(gmem).inner, @@ -1772,7 +2214,7 @@ pub const Builder = struct { self.asValue(cta_map).inner, self.tmaLoadAtomType(config), self.coordTensorType(config.coordinate_rank, config.coordinate_layout), - (cute_nvgpu.TiledTmaLoadAttr.get(self.ctx, .{ .value = .sm_90 }) catch @panic("invalid tiled TMA load")).attribute(), + (cute_nvgpu.TiledTmaLoadAttr.get(self.ctx, .{ .value = tmaLoadKind(config) }) catch @panic("invalid tiled TMA load")).attribute(), .{ .num_multicast = .int(self.ctx, .i32, 1), .tma_format = tmaFormatAttribute(self, config) }, self.loc(), )); @@ -1802,8 +2244,15 @@ pub const Builder = struct { /// Host-side tiled TMA load atom type described by `config`. pub fn tmaLoadAtomType(self: *Builder, config: TmaConfig) *const mlir.Type { + if (config.mode == .gather4) return (cute_nvgpu.CopyAtomNonExec2DGather4TmaLoadType.get(self.ctx, .{ + .kind = if (config.cta_group == .two) .sm_100_2sm else .sm_100, + .valType = config.dtype.toMlir(self.ctx), + .copyBits = @intCast(config.copy_bits), + .tmaGbasis = self.layoutType(config.global_basis), + .tmaFormat = if (config.format == .default) null else tmaDataFormat(config.format, config.dtype), + }) catch @panic("invalid gather TMA atom")).type_(); return (cute_nvgpu.CopyAtomNonExecTiledTmaLoadType.get(self.ctx, .{ - .kind = .sm_90, + .kind = tmaLoadKind(config), .valType = config.dtype.toMlir(self.ctx), .copyBits = @intCast(config.copy_bits), .tmaGbasis = self.layoutType(config.global_basis), @@ -1818,8 +2267,8 @@ pub const Builder = struct { return (cute_nvgpu.CopyAtomTmaLoadType.get(self.ctx, .{ .valType = config.dtype.toMlir(self.ctx), .copyBits = @intCast(config.copy_bits), - .mode = .tiled, - .numCta = 1, + .mode = if (config.mode == .gather4) .gather4 else .tiled, + .numCta = ctaCount(config.cta_group), .gStride = self.emptyStrideType(), .tmaGbasis = self.layoutType(config.global_basis), }) catch @panic("invalid TMA load atom")).type_(); @@ -1851,6 +2300,12 @@ pub const Builder = struct { return (cute_ir.StrideType.get(self.ctx, .{ .attr = self.algebraAttr(cute_ir.StrideAttr, .{ .tuple = &.{} }).attribute() }) catch @panic("invalid CuTe stride")).type_(); } + /// Generic address (i64) of the tensor map in an executable tiled TMA atom. + pub fn tmaDescriptorAddress(self: *Builder, descriptor: Atom) Value { + const ptr = self.emit(cute_nvgpu.get_tma_desc_addr(self.ctx, descriptor.inner, (cute_ir.PtrType.get(self.ctx, .{ .valueType = (cute_nvgpu.TmaDescriptorTiledType.get(self.ctx, .{}) catch unreachable).type_(), .addressSpace = .generic, .alignment = 64 }) catch unreachable).type_(), self.loc())); + return self.ptrToInt(ptr, .i64); + } + pub fn prefetchTmaDesc(self: *Builder, descriptor: Atom) void { self.emitVoid(cute_nvgpu.prefetch_tma_desc(self.ctx, descriptor.inner, self.loc())); } @@ -1861,11 +2316,11 @@ pub const Builder = struct { } /// Attach the mbarrier that an executable TMA load atom completes. - pub fn setTmaBarrier(self: *Builder, atom: Atom, barrier: Value) Atom { + pub fn setTmaBarrier(self: *Builder, atom: Atom, mbar: Value) Atom { const result = self.emit(cute_nvgpu.atom_set_value( self.ctx, atom.inner, - barrier.inner, + mbar.inner, atom.type_(), (cute_nvgpu.AtomCopyFieldTmaLoadAttr.get(self.ctx, .{ .value = .tma_bar }) catch @panic("invalid TMA load field")).attribute(), self.loc(), @@ -1915,10 +2370,10 @@ pub const Builder = struct { } fn blockScaledAtomType(self: *Builder, cfg: BlockScaledMmaConfig) *const mlir.Type { - if (cfg.n < 8 or cfg.n > 256 or cfg.n % 8 != 0) std.debug.panic("unsupported block-scaled MMA N {d}", .{cfg.n}); + cfg.validate(); return (cute_nvgpu.MmaAtomSM100UMMABlockScaledType.get(self.ctx, .{ - .shapeMnk = self.algebraAttr(cute_ir.ShapeAttr, self.expr(.{ 128, cfg.n, cfg.k() })).attribute(), - .numCta = 1, + .shapeMnk = self.algebraAttr(cute_ir.ShapeAttr, self.expr(.{ cfg.m, cfg.n, cfg.k() })).attribute(), + .numCta = ctaCount(cfg.cta_group), .aMajor = .k, .bMajor = .k, .aType = DType.f4e2m1fn.toMlir(self.ctx), @@ -1952,10 +2407,27 @@ pub const Builder = struct { return .{ .inner = result.inner, .kernel = self }; } - /// Commit the preceding tcgen05 MMA operations to an mbarrier. One elected - /// lane of the MMA warp issues this after each K-stage. - pub fn tcgen05Commit(self: *Builder, barrier: Value) void { - self.emitVoid(nvvm.tcgen05_commit(self.ctx, self.sharedBarrierPtr(barrier).inner, .cta_1, self.loc())); + pub const Tcgen05CommitOptions = struct { + /// Must match the group of the operations being tracked. + cta_group: CtaGroup = .one, + /// Arrive at the same barrier offset in each CTA whose rank bit is set. + /// Without a mask, arrive only on the issuing CTA's barrier. + multicast_mask: ?Value = null, + }; + + /// Track this thread's preceding tcgen05 MMA/copy operations of the + /// selected group, signaling one arrival per destination on completion. + /// The same elected thread that issues those operations must commit them. + /// For `.two`, either CTA in the pair can be the issuer. All tcgen05 + /// operations in a kernel must use the same group. + pub fn tcgen05Commit(self: *Builder, mbar: Value, opts: Tcgen05CommitOptions) void { + self.emitVoid(nvvm.tcgen05_commit( + self.ctx, + self.sharedBarrierPtr(mbar).inner, + opts.cta_group.toNvvm(), + if (opts.multicast_mask) |mask| mask.to(.i16).inner else null, + self.loc(), + )); } /// Wait until asynchronous tensor-memory loads have reached registers. @@ -1963,14 +2435,14 @@ pub const Builder = struct { self.emitVoid(nvvm.tcgen05_wait(self.ctx, .load, self.loc())); } - pub const TmemOptions = struct { two_cta: bool = false, exclusive: bool = false }; + pub const TmemOptions = struct { cta_group: CtaGroup = .one, exclusive: bool = false }; pub fn allocTmem(self: *Builder, columns: u16, holding_ptr: anytype, opts: TmemOptions) void { self.emitVoid(cute_nvgpu.arch_sm100_alloc_tmem( self.ctx, self.constant(.i32, columns).inner, self.asValue(holding_ptr).inner, - .{ .is_two_cta = if (opts.two_cta) .unit(self.ctx) else null, .exclusive = if (opts.exclusive) .unit(self.ctx) else null }, + .{ .is_two_cta = if (opts.cta_group == .two) .unit(self.ctx) else null, .exclusive = if (opts.exclusive) .unit(self.ctx) else null }, self.loc(), )); } @@ -1981,7 +2453,7 @@ pub const Builder = struct { } pub fn relinquishTmemAllocPermit(self: *Builder, opts: TmemOptions) void { - self.emitVoid(cute_nvgpu.arch_sm100_relinquish_tmem_alloc_permit(self.ctx, .{ .is_two_cta = if (opts.two_cta) .unit(self.ctx) else null }, self.loc())); + self.emitVoid(cute_nvgpu.arch_sm100_relinquish_tmem_alloc_permit(self.ctx, .{ .is_two_cta = if (opts.cta_group == .two) .unit(self.ctx) else null }, self.loc())); } pub fn deallocTmem(self: *Builder, pointer: anytype, columns: u16, opts: TmemOptions) void { @@ -1989,7 +2461,7 @@ pub const Builder = struct { self.ctx, self.asValue(pointer).inner, self.constant(.i32, columns).inner, - .{ .is_two_cta = if (opts.two_cta) .unit(self.ctx) else null, .exclusive = if (opts.exclusive) .unit(self.ctx) else null }, + .{ .is_two_cta = if (opts.cta_group == .two) .unit(self.ctx) else null, .exclusive = if (opts.exclusive) .unit(self.ctx) else null }, self.loc(), )); } @@ -2182,22 +2654,37 @@ pub const Builder = struct { _ = nvvm.barrier(self.ctx, self.loc()).appendTo(self.currentBlock()); } - /// Synchronize a fixed subset of the CTA on a named hardware barrier. - /// `participants` must be a warp multiple and every participating thread - /// must execute the same barrier id. - pub fn namedBarrier(self: *Builder, id: i32, participants: i32) void { - // The embedded CuTe compiler currently reads NVVM through a different - // MLIR revision. A generic `nvvm.barrier` with optional operands is - // accepted, but those operands are discarded and it lowers to - // `bar.sync 0`. Constant inline PTX preserves the named barrier ABI - // used by CUTLASS (`2,160` for TMEM allocation and `1,128` for the - // epilogue pipeline). - const assembly = std.fmt.allocPrint( - self.arena.allocator(), - "bar.sync {d}, {d};", - .{ id, participants }, - ) catch @panic("namedBarrier OOM"); - _ = llvm.inline_asm(self.ctx, assembly, &.{}, .{ .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + /// `bar.sync id, participants` (`barrier.cta.sync.aligned`). + /// `id` is an integer or an i32 Value in [0, 15]. `participants` must be + /// a positive warp multiple, and every participating thread must execute + /// the same barrier id. The aligned form requires CTA-uniform control flow. + pub fn barrierCtaSyncAligned(self: *Builder, id: anytype, participants: anytype) void { + if (@TypeOf(participants) == ?Value) { + if (participants) |count| return self.barrierCtaSyncAligned(id, count); + return self.barrierCtaSyncAligned(id, null); + } + if (@TypeOf(participants) != Value and @TypeOf(participants) != @TypeOf(null)) { + std.debug.assert(participants > 0 and participants <= 1024 and @mod(participants, 32) == 0); + } + if (@TypeOf(id) == Value) { + std.debug.assert(id.dtype() == .i32); + } else { + const immediate: i32 = id; + std.debug.assert(immediate >= 0 and immediate < 16); + } + // The LLVM intrinsic preserves convergence and both operands across + // the embedded compiler's NVVM dialect version boundary. + if (@TypeOf(participants) == @TypeOf(null)) { + _ = self.callIntrinsic("llvm.nvvm.barrier.cta.sync.aligned", &.{self.lift(id).to(.i32)}, null); + } else { + _ = self.callIntrinsic("llvm.nvvm.barrier.cta.sync.aligned.count", &.{ self.lift(id).to(.i32), self.lift(participants).to(.i32) }, null); + } + } + + pub const BarrierOptions = struct { barrier_id: ?Value = null, number_of_threads: ?Value = null }; + + pub fn barrier(self: *Builder, opts: BarrierOptions) void { + self.barrierCtaSyncAligned(opts.barrier_id orelse self.constant(.i32, 0), opts.number_of_threads); } /// Build the CuTe first-class divisor used by CUTLASS's persistent tile @@ -2368,48 +2855,101 @@ pub const Builder = struct { /// mbarrier operands are `!llvm.ptr<3>` for NVVM. A barrier is either a CuTe /// shared-memory pointer or a raw `i32` shared-memory address. - fn sharedBarrierPtr(self: *Builder, barrier: Value) Value { - if (barrier.isInt()) return self.addressPtr(barrier, .shared); - return self.emit(mlir.unrealized_conversion_cast(self.ctx, &.{barrier.inner}, &.{llvm.pointerType(self.ctx, NVPTXAddressSpace.shared)}, self.loc())); + fn sharedBarrierPtr(self: *Builder, mbar: Value) Value { + if (mbar.isInt()) return self.addressPtr(mbar, .shared); + return self.emit(mlir.unrealized_conversion_cast(self.ctx, &.{mbar.inner}, &.{llvm.pointerType(self.ctx, NVPTXAddressSpace.shared)}, self.loc())); } /// Initialize a shared-memory mbarrier expecting `arrivals` per phase. - pub fn mbarrierInit(self: *Builder, barrier: Value, arrivals: u32) void { - self.emitVoid(nvvm.mbarrier_init(self.ctx, self.sharedBarrierPtr(barrier).inner, self.constant(.i32, arrivals).inner, self.loc())); + pub fn mbarrierInit(self: *Builder, mbar: Value, arrivals: u32) void { + self.emitVoid(nvvm.mbarrier_init(self.ctx, self.sharedBarrierPtr(mbar).inner, self.constant(.i32, arrivals).inner, self.loc())); } /// Arrive and expect `bytes` of asynchronous transactions (SM90+), as a /// TMA producer does before issuing its copies. - pub fn mbarrierArriveExpectTx(self: *Builder, barrier: Value, bytes: u32) void { - self.emitVoid(nvvm.mbarrier_arrive_expect_tx(self.ctx, self.sharedBarrierPtr(barrier).inner, self.constant(.i32, bytes).inner, .cta, self.loc())); + pub fn mbarrierArriveExpectTx(self: *Builder, mbar: Value, bytes: anytype) void { + self.emitVoid(nvvm.mbarrier_arrive_expect_tx(self.ctx, self.sharedBarrierPtr(mbar).inner, self.lift(bytes).to(.i32).inner, .cta, self.loc())); } - pub fn mbarrierArrive(self: *Builder, barrier: Value, count: u32) void { - self.emitVoid(nvvm.mbarrier_arrive(self.ctx, self.sharedBarrierPtr(barrier).inner, self.constant(.i32, count).inner, .cta, self.loc())); + pub const MBarrierArriveOptions = struct { + peer_cta_rank_in_cluster: ?Value = null, + arrive_count: ?Value = null, + scope: MBarrierScope = .cta, + relaxed: bool = false, + }; + + /// Arrive locally or at the same offset in a peer CTA. Local arrivals + /// return the opaque state token; remote arrivals have no state result. + pub fn mbarrierArrive(self: *Builder, mbar: Value, opts: MBarrierArriveOptions) ?Value { + var ptr = self.sharedBarrierPtr(mbar); + if (opts.peer_cta_rank_in_cluster) |rank| { + ptr = self.emit(nvvm.mapa(self.ctx, ptr.inner, rank.inner, llvm.pointerType(self.ctx, NVPTXAddressSpace.shared_cluster), self.loc())); + } + const op = nvvm.ops.mbarrier_arrive(self.ctx, ptr.inner, .{ + .count = if (opts.arrive_count) |count| count.to(.i32).inner else null, + .scope = (nvvm.ops.MemScopeKindAttr.get(self.ctx, .{ .value = switch (opts.scope) { + .cta => .cta, + .cluster => .cluster, + } }) catch unreachable).attribute(), + .relaxed = .boolean(self.ctx, opts.relaxed), + .res_type = if (opts.peer_cta_rank_in_cluster == null) mlir.Type.int(self.ctx, .i64) else null, + }, self.loc()); + if (opts.peer_cta_rank_in_cluster == null) return self.emit(op); + self.emitVoid(op); + return null; } - /// Test whether phase `phase` has completed, without suspending (SM90+). - pub fn mbarrierWaitParity(self: *Builder, barrier: Value, phase: Value) Value { - return self.emit(nvvm.mbarrier_wait_parity(self.ctx, self.sharedBarrierPtr(barrier).inner, phase.inner, .@"try", .cta, self.loc())); + pub const MBarrierTryWaitOptions = struct { + scope: MBarrierScope = .cta, + time_limit: ?Value = null, + }; + + /// Try an acquire wait for phase `phase` (i32 parity, 0 or 1). May suspend + /// briefly and return false before completion. `scope` selects which + /// threads this wait can synchronize with; the barrier remains local. + pub fn mbarrierTryWait(self: *Builder, mbar: Value, phase: Value, opts: MBarrierTryWaitOptions) Value { + const ptr = self.sharedBarrierPtr(mbar); + if (opts.time_limit) |ticks| { + return self.emit(nvvm.mbarrier_try_wait_parity_timelimit(self.ctx, ptr.inner, phase.inner, ticks.to(.i32).inner, opts.scope, self.loc())); + } + return self.emit(nvvm.mbarrier_wait_parity(self.ctx, ptr.inner, phase.inner, .@"try", opts.scope, self.loc())); } - /// Wait for phase `phase`, letting the warp sleep for up to - /// `suspend_time` cycles per attempt (SM90+). - pub fn mbarrierTryWaitParity(self: *Builder, barrier: Value, phase: Value, suspend_time: u32) void { - self.emitVoid(nvvm.mbarrier_try_wait_parity(self.ctx, self.sharedBarrierPtr(barrier).inner, phase.inner, self.constant(.i32, suspend_time).inner, self.loc())); + pub const MBarrierWaitOptions = struct { + scope: MBarrierScope = .cta, + /// Hardware suspend hint in nanoseconds per attempt, not a total timeout. + time_limit: u32 = 10_000_000, + /// Probe before suspending, as in CuTe's pipeline consumer wait. + test_first: bool = true, + }; + + fn mbarrierWaitLoop(self: *Builder, mbar: Value, phase: Value, opts: MBarrierWaitOptions) void { + const ptr = self.sharedBarrierPtr(mbar); + const ticks = self.constant(.i32, opts.time_limit); + switch (opts.scope) { + // Preserve the operation recognized by CuTe's pipeline compiler. + .cta => self.emitVoid(nvvm.mbarrier_try_wait_parity(self.ctx, ptr.inner, phase.inner, ticks.inner, self.loc())), + .cluster => { + var pending = self.openWhile(.{}, .{}); + const ready = self.mbarrierTryWait(mbar, phase, .{ .scope = opts.scope, .time_limit = ticks }); + pending.yieldBefore(ready.eq(false), .{}); + pending.yieldAfter(.{}); + }, + } } - /// Wait for a pipeline phase with the two-step sequence of CuTe's - /// `PipelineTmaUmma`: test the barrier without suspending, then suspend - /// only when the phase is still outstanding. The initial test matters when + /// Acquire a pipeline phase at `scope`: test without suspending, then + /// block only when the phase is still outstanding. The initial test matters when /// a producer reaches a full ring and must yield to the consumer that /// recycles its next stage. - pub fn mbarrierWait(self: *Builder, barrier: Value, phase: Value) void { - const ready = self.mbarrierWaitParity(barrier, phase); - var pending = self.openIf(ready.eq(false)); - { - self.mbarrierTryWaitParity(barrier, phase, 10_000_000); + pub fn mbarrierWait(self: *Builder, mbar: Value, phase: Value, opts: MBarrierWaitOptions) void { + if (opts.test_first) { + const ready = self.mbarrierTestWait(mbar, phase, .{ .scope = opts.scope }); + var pending = self.openIf(ready.eq(false)); + self.mbarrierWaitLoop(mbar, phase, opts); pending.close(); + } else { + self.mbarrierWaitLoop(mbar, phase, opts); } } @@ -2438,7 +2978,7 @@ pub const Builder = struct { // Hand-scheduled kernels address shared memory with `i32` offsets in the // shared window, global memory with `i64` addresses and tensor memory with // `i32` TMEM addresses (lane << 16 | column). These helpers emit the NVVM / - // LLVM operations for them directly (no inline PTX). + // LLVM operations for them directly, using inline PTX where needed. pub const NVPTXAddressSpace = llvm.NVPTXAddressSpace; @@ -2500,11 +3040,6 @@ pub const Builder = struct { self.emitVoid(llvm.store(self.ctx, word.to(.i16).inner, self.addressPtr(address, space).inner, .{ .alignment = 2 }, self.loc())); } - /// `bar.sync id, threads`: named barrier over `threads` threads. - pub fn barrierSync(self: *Builder, id: i32, threads: i32) void { - _ = self.callIntrinsic("llvm.nvvm.barrier.cta.sync.aligned.count", &.{ self.constant(.i32, id), self.constant(.i32, threads) }, null); - } - /// `bar.red.or.pred`: named barrier over `threads` threads that also returns whether /// `predicate` (i1) is true in any of them. pub fn barrierReduceOr(self: *Builder, id: i32, threads: i32, predicate: Value) Value { @@ -2516,9 +3051,10 @@ pub const Builder = struct { self.emitVoid(nvvm.bar_warp_sync(self.ctx, self.constant(.i32, -1).inner, self.loc())); } - /// Non-blocking test of mbarrier phase `parity` (i1). - pub fn mbarrierTestParity(self: *Builder, barrier: Value, parity: Value) Value { - return self.emit(nvvm.mbarrier_wait_parity(self.ctx, self.sharedBarrierPtr(barrier).inner, parity.inner, .@"test", .cta, self.loc())); + /// Non-blocking acquire test of mbarrier phase `parity` (i32, 0 or 1). + pub fn mbarrierTestWait(self: *Builder, mbar: Value, parity: Value, opts: MBarrierTryWaitOptions) Value { + std.debug.assert(opts.time_limit == null); + return self.emit(nvvm.mbarrier_wait_parity(self.ctx, self.sharedBarrierPtr(mbar).inner, parity.inner, .@"test", opts.scope, self.loc())); } /// Call an LLVM intrinsic by name (e.g. an `llvm.nvvm.*` intrinsic the NVVM dialect of @@ -2549,8 +3085,8 @@ pub const Builder = struct { /// Arrive on `barrier` once this thread's prior cp.async copies have landed; with /// `noinc`, the arrival is counted in the barrier's expected count. - pub fn cpAsyncMbarrierArrive(self: *Builder, barrier: Value, comptime noinc: bool) void { - _ = self.callIntrinsic(if (noinc) "llvm.nvvm.cp.async.mbarrier.arrive.noinc.shared" else "llvm.nvvm.cp.async.mbarrier.arrive.shared", &.{self.sharedBarrierPtr(barrier)}, null); + pub fn cpAsyncMbarrierArrive(self: *Builder, mbar: Value, comptime noinc: bool) void { + _ = self.callIntrinsic(if (noinc) "llvm.nvvm.cp.async.mbarrier.arrive.noinc.shared" else "llvm.nvvm.cp.async.mbarrier.arrive.shared", &.{self.sharedBarrierPtr(mbar)}, null); } /// Wait until all of this thread's cp.async copies have landed. @@ -2705,13 +3241,271 @@ pub const Builder = struct { return self.gridDim().z.mul(value); } + /// SM100 `tile::gather4` TMA load from a tiled 2-D descriptor address `desc` (i64). + /// Gather four rows `r` (i32) at column `col` into 128-byte-aligned shared `dst` (i32), + /// completing their bytes on an 8-byte-aligned local mbarrier `bar` (i32). + /// The descriptor must have row box extent 1 and no interleaving. `col` and `r` + /// are element coordinates; row length and out-of-bounds fill come from the + /// descriptor (zero or OOB-NaN). Arm the barrier for all four rows' bytes before waiting. + pub fn tmaGather4(self: *Builder, desc: Value, dst: Value, bar: Value, col: Value, r: [4]Value, policy: CacheHintSm100) void { + std.debug.assert(desc.dtype() == .i64 and dst.dtype() == .i32 and bar.dtype() == .i32 and col.dtype() == .i32); + for (r) |row| std.debug.assert(row.dtype() == .i32); + _ = llvm.inline_asm(self.ctx, "cp.async.bulk.tensor.2d.shared::cta.global.tile::gather4.mbarrier::complete_tx::bytes.cta_group::1.L2::cache_hint [$0], [$1, {$2, $3, $4, $5, $6}], [$7], $8;", &.{ dst.inner, desc.inner, col.inner, r[0].inner, r[1].inner, r[2].inner, r[3].inner, bar.inner, self.constant(.i64, policy).inner }, .{ .constraints = "r,l,r,r,r,r,r,r,l,~{memory}", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + pub const BulkPrefetchOptions = struct { l2_cache_hint: ?Value = null }; + + /// Best-effort byte-range prefetch into L2; size must be a multiple of 16. + pub fn cpAsyncBulkPrefetch(self: *Builder, addr: Value, bytes: anytype, opts: BulkPrefetchOptions) void { + if (@TypeOf(bytes) != Value) std.debug.assert(bytes > 0 and bytes % 16 == 0); + const address = if (addr.isInt()) self.addressPtr(addr, .global) else self.llvmPointer(addr, .global); + self.emitVoid(nvvm.ops.cp_async_bulk_prefetch(self.ctx, address.inner, self.lift(bytes).to(.i32).inner, .{ + .l2CacheHint = if (opts.l2_cache_hint) |hint| hint.to(.i64).inner else null, + }, self.loc())); + } + + pub fn prefetchL1(self: *Builder, addr: Value) void { + const address = if (addr.isInt()) addr else self.ptrToInt(addr, .i64); + _ = llvm.inline_asm(self.ctx, "prefetch.global.L1 [$0];", &.{address.inner}, .{ .constraints = "l", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + pub fn prefetchTensorMap(self: *Builder, desc: Value) void { + const address = if (desc.isInt()) desc else self.ptrToInt(desc, .i64); + _ = llvm.inline_asm(self.ctx, "prefetch.tensormap [$0];", &.{address.inner}, .{ .constraints = "l", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + pub const OperandSource = union(enum) { smem: Value, tmem: Value }; + pub const Tcgen05MmaWsOptions = struct { + collector_b_buffer: ?nvvm.ops.Tcgen05MMACollectorBBuffer = null, + collector_op: ?nvvm.ops.Tcgen05MMACollectorOp = null, + col_b_zero_mask: ?Value = null, + }; + + /// Single-thread, asynchronous weight-stationary MMA, as in Python primitives. + /// The caller elects the issuer and commits/waits before consuming D. + pub fn tcgen05MmaWs(self: *Builder, kind: nvvm.Tcgen05MmaKind, d: Value, a: OperandSource, bdesc: Value, idesc: anytype, enable_input_d: Value, opts: Tcgen05MmaWsOptions) void { + std.debug.assert(kind == .f16 or kind == .tf32 or kind == .f8f6f4 or kind == .i8); + const adesc = switch (a) { + .smem => |desc| self.smemDescriptorToInt(desc), + .tmem => |ptr| if (ptr.isInt()) self.addressPtr(ptr, .tensor) else self.llvmPointer(ptr, .tensor), + }; + const dest = if (d.isInt()) self.addressPtr(d, .tensor) else self.llvmPointer(d, .tensor); + self.emitVoid(nvvm.ops.tcgen05_mma_ws(self.ctx, dest.inner, adesc.inner, self.smemDescriptorToInt(bdesc).inner, self.lift(idesc).to(.i32).inner, enable_input_d.to(.i1).inner, (nvvm.ops.Tcgen05MMAKindAttr.get(self.ctx, .{ .value = kind }) catch unreachable).attribute(), .{ + .collectorBBuffer = if (opts.collector_b_buffer) |v| (nvvm.ops.Tcgen05MMACollectorBBufferAttr.get(self.ctx, .{ .value = v }) catch unreachable).attribute() else null, + .collectorOp = if (opts.collector_op) |v| (nvvm.ops.Tcgen05MMACollectorOpAttr.get(self.ctx, .{ .value = v }) catch unreachable).attribute() else null, + .zeroColMask = if (opts.col_b_zero_mask) |v| v.to(.i64).inner else null, + }, self.loc())); + } + + pub fn makeUmmaSmemDesc(self: *Builder, src: Value, layout: LayoutSpec, major: OperandMajorMode, next_src: ?Value) Value { + return self.emit(cute_nvgpu.make_umma_smem_desc(self.ctx, src.inner, self.smemDescType(), self.layoutAttr(layout).attribute(), (cute_nvgpu.MajorModeAttr.get(self.ctx, .{ .value = switch (major) { + .k => .k, + .mn => .mn, + } }) catch unreachable).attribute(), .{ .nextSrc = if (next_src) |v| v.inner else null }, self.loc())); + } + + pub fn intToSmemDescriptor(self: *Builder, value: anytype) Value { + return self.emit(mlir.unrealized_conversion_cast(self.ctx, &.{self.lift(value).to(.i64).inner}, &.{self.smemDescType()}, self.loc())); + } + + pub fn smemDescriptorToInt(self: *Builder, desc: Value) Value { + if (desc.isInt()) return desc.to(.i64); + return self.emit(mlir.unrealized_conversion_cast(self.ctx, &.{desc.inner}, &.{mlir.Type.int(self.ctx, .i64)}, self.loc())); + } + + pub fn setMaxRegisterIncrease(self: *Builder, comptime count: u32) void { + self.setMaxRegister(count, .increase); + } + + pub fn setMaxRegisterDecrease(self: *Builder, comptime count: u32) void { + self.setMaxRegister(count, .decrease); + } + + fn setMaxRegister(self: *Builder, comptime count: u32, comptime action: enum { increase, decrease }) void { + comptime std.debug.assert(count >= 24 and count <= 256 and count % 8 == 0); + const text = std.fmt.comptimePrint("setmaxnreg.{s}.sync.aligned.u32 {d};", .{ if (action == .increase) "inc" else "dec", count }); + _ = llvm.inline_asm(self.ctx, text, &.{}, .{ .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + /// Specialized busy wait for hand-scheduled kernels; never hardware-suspends. + pub fn mbarrierSpin(self: *Builder, bar: Value, parity: Value, scope: MBarrierScope) void { + std.debug.assert(bar.dtype() == .i32 and parity.dtype() == .i32); + _ = llvm.inline_asm(self.ctx, switch (scope) { + .cta => "{ .reg .pred p; spin${:uid}: mbarrier.test_wait.parity.acquire.cta.shared::cta.b64 p, [$0], $1; @!p bra spin${:uid}; }", + .cluster => "{ .reg .pred p; spin${:uid}: mbarrier.test_wait.parity.acquire.cluster.shared::cta.b64 p, [$0], $1; @!p bra spin${:uid}; }", + }, &.{ bar.inner, parity.inner }, .{ .constraints = "r,r,~{memory}", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + /// Convert a CuTe or LLVM pointer to the LLVM pointer required by NVVM. + fn llvmPointer(self: *Builder, ptr: Value, comptime space: NVPTXAddressSpace) Value { + if (!ptr.isPtr()) return ptr; + return self.emit(mlir.unrealized_conversion_cast(self.ctx, &.{ptr.inner}, &.{llvm.pointerType(self.ctx, space)}, self.loc())); + } + + /// Map a local shared pointer/address to a typed pointer in peer CTA memory. + pub fn mapDsmemPtr(self: *Builder, smem: Value, peer_cta_rank: Value) Value { + const ptr = if (smem.isInt()) self.addressPtr(smem, .shared) else self.llvmPointer(smem, .shared); + const remote = self.emit(nvvm.mapa(self.ctx, ptr.inner, peer_cta_rank.to(.i32).inner, llvm.pointerType(self.ctx, NVPTXAddressSpace.shared_cluster), self.loc())); + if (smem.type_().isA(cute_ir.PtrType)) |ty| { + std.debug.assert(ty.addressSpace() == .smem); + const result_type = (cute_ir.PtrType.get(self.ctx, .{ + .valueType = ty.valueType(), + .addressSpace = .dsmem, + .alignment = ty.alignment(), + .swizzle = ty.swizzle(), + .bitlayout = ty.bitlayout(), + }) catch @panic("invalid DSMEM pointer")).type_(); + return self.emit(cute_ir.inttoptr(self.ctx, self.ptrToInt(remote, .i32).inner, result_type, self.loc())); + } + return remote; + } + + /// Store one, two, or four i32 words in a peer CTA, completing bytes on its + /// mbarrier. Both inputs name local shared memory; both are mapped to the peer. + pub fn storeAsyncDsmem(self: *Builder, smem: Value, value: anytype, mbar: Value, peer_cta_rank: Value) void { + const n = if (@TypeOf(value) == Value) 1 else common.arity(@TypeOf(value), "storeAsyncDsmem: value"); + comptime std.debug.assert(n == 1 or n == 2 or n == 4); + if (smem.type_().isA(cute_ir.PtrType)) |ty| { + std.debug.assert(ty.addressSpace() == .smem and ty.alignment() % (n * 4) == 0); + } + if (mbar.type_().isA(cute_ir.PtrType)) |ty| { + std.debug.assert(ty.addressSpace() == .smem and ty.alignment() % 8 == 0); + } + const packed_value = if (@TypeOf(value) == Value) value else blk: { + var words: [n]Value = undefined; + inline for (0..n) |i| words[i] = self.lift(value[i]).to(.i32); + break :blk self.joinWords(n, words); + }; + self.emitVoid(nvvm.ops.store_async(self.ctx, self.llvmPointer(self.mapDsmemPtr(smem, peer_cta_rank), .shared_cluster).inner, packed_value.inner, self.llvmPointer(self.mapDsmemPtr(mbar, peer_cta_rank), .shared_cluster).inner, self.loc())); + } + + /// Explicit aligned variants retain the convergence requirement of raw PTX. + pub fn barrierClusterArriveRelaxedAligned(self: *Builder) void { + _ = llvm.inline_asm(self.ctx, "barrier.cluster.arrive.relaxed.aligned;", &.{}, .{ .constraints = "~{memory}", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + pub fn barrierClusterWaitAligned(self: *Builder) void { + self.emitVoid(nvvm.cluster_wait(self.ctx, true, self.loc())); + } + + pub const ClusterArriveOptions = struct { aligned: bool = false }; + + pub fn clusterArriveRelaxed(self: *Builder, opts: ClusterArriveOptions) void { + if (opts.aligned) return self.barrierClusterArriveRelaxedAligned(); + _ = llvm.inline_asm(self.ctx, "barrier.cluster.arrive.relaxed;", &.{}, .{ .constraints = "~{memory}", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + } + + pub fn clusterWait(self: *Builder) void { + self.emitVoid(nvvm.cluster_wait(self.ctx, false, self.loc())); + } + + /// A null mask selects CuTe's FULL_MASK; integers and SSA masks also work. + pub fn voteAnySync(self: *Builder, pred: Value, mask: anytype) Value { + const active = if (@TypeOf(mask) == @TypeOf(null)) self.constant(.i32, -1) else self.lift(mask).to(.i32); + return self.callIntrinsic("llvm.nvvm.vote.any.sync", &.{ active, pred }, mlir.Type.int(self.ctx, .i1)).?; + } + + /// A null mask selects CuTe's FULL_MASK. + pub fn voteBallotSync(self: *Builder, pred: Value, mask: anytype) Value { + const active = if (@TypeOf(mask) == @TypeOf(null)) self.constant(.i32, -1) else self.lift(mask).to(.i32); + return self.callIntrinsic("llvm.nvvm.vote.ballot.sync", &.{ active, pred }, mlir.Type.int(self.ctx, .i32)).?; + } + + pub const FmaxOptions = struct { abs: bool = false, nan: bool = false, ftz: bool = false }; + + pub fn fmax(self: *Builder, x: Value, y: Value, opts: FmaxOptions) Value { + if (!opts.abs and !opts.nan and !opts.ftz) return self.callIntrinsic("llvm.nvvm.fmax.f", &.{ x, y }, mlir.Type.float(self.ctx, .f32)).?; + return self.emit(nvvm.ops.fmax(self.ctx, x.inner, y.inner, mlir.Type.float(self.ctx, .f32), .{ + .abs = if (opts.abs) .unit(self.ctx) else null, + .nan = if (opts.nan) .unit(self.ctx) else null, + .ftz = if (opts.ftz) .unit(self.ctx) else null, + }, self.loc())); + } + + fn vec2(self: *Builder, x: [2]Value) Value { + const ty = mlir.Type.vector(&.{2}, mlir.Type.float(self.ctx, .f32)); + return self.emit(vector.from_elements(self.ctx, &.{ x[0].inner, x[1].inner }, ty, self.loc())); + } + + fn unvec2(self: *Builder, v: Value) [2]Value { + return .{ self.vectorExtract(v, 0, .f32), self.vectorExtract(v, 1, .f32) }; + } + + pub const PackedF32x2Options = struct { rnd: nvvm.ops.FPRoundingMode = .rn, ftz: bool = false }; + + fn packedF32x2Attrs(self: *Builder, opts: PackedF32x2Options) struct { rnd: ?*const mlir.Attribute, ftz: ?*const mlir.Attribute } { + return .{ + .rnd = (nvvm.ops.FPRoundingModeAttr.get(self.ctx, .{ .value = opts.rnd }) catch unreachable).attribute(), + .ftz = if (opts.ftz) .unit(self.ctx) else null, + }; + } + + pub fn fmaPackedF32x2(self: *Builder, a: [2]Value, b: [2]Value, c: [2]Value, opts: PackedF32x2Options) [2]Value { + return self.unvec2(self.emit(nvvm.ops.fma_packed_f32x2(self.ctx, self.vec2(a).inner, self.vec2(b).inner, self.vec2(c).inner, mlir.Type.vector(&.{2}, mlir.Type.float(self.ctx, .f32)), .{ .rnd = self.packedF32x2Attrs(opts).rnd, .ftz = self.packedF32x2Attrs(opts).ftz }, self.loc()))); + } + + pub fn addPackedF32x2(self: *Builder, a: [2]Value, b: [2]Value, opts: PackedF32x2Options) [2]Value { + return self.unvec2(self.emit(nvvm.ops.add_packed_f32x2(self.ctx, self.vec2(a).inner, self.vec2(b).inner, mlir.Type.vector(&.{2}, mlir.Type.float(self.ctx, .f32)), .{ .rnd = self.packedF32x2Attrs(opts).rnd, .ftz = self.packedF32x2Attrs(opts).ftz }, self.loc()))); + } + + pub fn mulPackedF32x2(self: *Builder, a: [2]Value, b: [2]Value, opts: PackedF32x2Options) [2]Value { + return self.unvec2(self.emit(nvvm.ops.mul_packed_f32x2(self.ctx, self.vec2(a).inner, self.vec2(b).inner, mlir.Type.vector(&.{2}, mlir.Type.float(self.ctx, .f32)), .{ .rnd = self.packedF32x2Attrs(opts).rnd, .ftz = self.packedF32x2Attrs(opts).ftz }, self.loc()))); + } + + pub const StoreOptions = struct { + /// Address space for raw integer addresses; typed pointers retain theirs. + space: NVPTXAddressSpace = .global, + level1_eviction_priority: nvvm.ops.EvictKind = .evict_normal, + }; + + /// Scalar/vector store. Eight i32 words with no_allocate keep STG.E.256. + pub fn store(self: *Builder, addr: Value, value: anytype, comptime opts: StoreOptions) void { + if (@TypeOf(value) != Value) { + const n = common.arity(@TypeOf(value), "store: value"); + comptime std.debug.assert(n == 1 or n == 2 or n == 4 or n == 8); + var words: [n]Value = undefined; + inline for (0..n) |i| words[i] = self.lift(value[i]).to(.i32); + const is_global = if (addr.type_().isA(cute_ir.PtrType)) |ty| ty.addressSpace() == .gmem else addr.isInt() and opts.space == .global; + if (n == 8 and is_global and opts.level1_eviction_priority == .evict_no_allocate) { + const address = if (addr.isInt()) addr else self.ptrToInt(addr, .i64); + var args: [9]*const mlir.Value = undefined; + args[0] = address.inner; + for (words, 0..) |word, i| args[i + 1] = word.inner; + _ = llvm.inline_asm(self.ctx, "st.global.L1::no_allocate.v8.b32 [$0], {$1, $2, $3, $4, $5, $6, $7, $8};", &args, .{ .constraints = "l,r,r,r,r,r,r,r,r,~{memory}", .has_side_effects = true }, self.loc()).appendTo(self.currentBlock()); + return; + } + self.store(addr, self.joinWords(n, words), opts); + return; + } + const ptr = if (addr.isInt()) self.addressPtr(addr, opts.space) else if (addr.type_().isA(cute_ir.PtrType)) |ty| blk: { + const space: NVPTXAddressSpace = @fromBackingInt(@intCast(@backingInt(ty.addressSpace()))); + break :blk self.emit(mlir.unrealized_conversion_cast(self.ctx, &.{addr.inner}, &.{llvm.pointerType(self.ctx, space)}, self.loc())); + } else addr; + self.emitVoid(nvvm.ops.store_ext(self.ctx, value.inner, ptr.inner, .{ + .evict = if (opts.level1_eviction_priority == .evict_normal) null else (nvvm.ops.EvictKindAttr.get(self.ctx, .{ .value = opts.level1_eviction_priority }) catch unreachable).attribute(), + }, self.loc())); + } + + pub fn fma(self: *Builder, x: Value, y: Value, z: Value) Value { + return self.callIntrinsic("llvm.fma.f32", &.{ x, y, z }, mlir.Type.float(self.ctx, .f32)).?; + } + + pub fn clock(self: *Builder) Value { + return self.emit(llvm.inline_asm(self.ctx, "mov.u32 $0, %clock;", &.{}, .{ .result_type = mlir.Type.int(self.ctx, .i32), .constraints = "=r", .has_side_effects = true }, self.loc())); + } + + pub fn globalTimer(self: *Builder) Value { + return self.emit(llvm.inline_asm(self.ctx, "mov.u64 $0, %globaltimer;", &.{}, .{ .result_type = mlir.Type.int(self.ctx, .i64), .constraints = "=l", .has_side_effects = true }, self.loc())); + } + // ==================== scalars ==================== - /// `value` as a constant of `dt`; a Value passes through. An integer is - /// signed or an unsigned bit pattern of `dt`'s width; anything else - /// panics. + /// `value` as a constant of `dt`; a Value passes through and enums use + /// their backing integers. An integer is signed or an unsigned bit + /// pattern of `dt`'s width; anything else panics. pub fn constant(self: *Builder, dt: DType, value: anytype) Value { if (@TypeOf(value) == Value) return value; + if (@typeInfo(@TypeOf(value)) == .@"enum") return self.constant(dt, @backingInt(value)); const ctx = self.ctx; const T = @TypeOf(value); if (isFloatDtype(dt)) { @@ -2957,7 +3751,7 @@ test "guard, shared memory, sync, layouts, for loop" { const i = b.blockIdx().x.mul(128).add(tid); smem.set(.{tid}, gx.get(.{i})); b.syncThreads(); - b.namedBarrier(2, 128); + b.barrierCtaSyncAligned(2, 128); b.launchDependents(); b.waitForDependency(); const j = b.constant(.i32, 127).sub(tid); @@ -2976,7 +3770,9 @@ test "guard, shared memory, sync, layouts, for loop" { try expectContains(kernel, "cute_nvgpu.arch.alloc_smem"); try expectContains(kernel, "!cute.ptr>"); try expectContains(kernel, "\"nvvm.barrier\"() <{operandSegmentSizes = array}> : () -> ()"); - try expectContains(kernel, "bar.sync 2, 128;"); + try expectContains(kernel, "llvm.nvvm.barrier.cta.sync.aligned.count"); + try expectContains(kernel, "value = 2 : i32"); + try expectContains(kernel, "value = 128 : i32"); try expectContains(kernel, "griddepcontrol.launch_dependents;"); try expectContains(kernel, "griddepcontrol.wait;"); try expectContains(kernel, "scf.if"); @@ -3113,30 +3909,1045 @@ test "SM100 block-scaled MMA atoms use the CuTe DSL instruction shapes" { var b = try Builder.open(std.testing.allocator, ctx, "sm100_blockscaled_atoms"); defer b.deinit(); - const descriptor = try cute_nvgpu.CopyAtomNonExecTiledTmaLoadType.get(ctx, .{ - .kind = .sm_90, - .valType = DType.f8e4m3fn.toMlir(ctx), - .copyBits = 16384, - .tmaGbasis = b.layoutType(b.layoutSpec(.{ 128, 16, 1 }, .{ b.basis(1, 1, .{1}), b.basis(1, 1, .{0}), b.basis(1, 1, .{2}) })), - .tmaFormat = .U8, + const tma: TmaConfig = .{ + .dtype = .f8e4m3fn, + .copy_bits = 16384, + .global_basis = b.layoutSpec(.{ 128, 16, 1 }, .{ b.basis(1, 1, .{1}), b.basis(1, 1, .{0}), b.basis(1, 1, .{2}) }), + .coordinate_layout = b.layoutSpec(.{ 16, 128, 8 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}), b.basis(1, 1, .{2}) }), + .coordinate_rank = 3, + }; + var pair_tma = tma; + pair_tma.cta_group = .two; + const args = try b.declareArgs(.{ + .single = .{ .mlir_type = .{ .type_ = b.tmaLoadAtomType(tma), .grid_constant = true } }, + .pair = .{ .mlir_type = .{ .type_ = b.tmaLoadAtomType(pair_tma), .grid_constant = true } }, }); - _ = try b.declareArgs(.{ .descriptor = .{ .mlir_type = .{ .type_ = descriptor.type_(), .grid_constant = true } } }); + _ = b.makeExecTma(.{ .inner = args.single.inner, .kernel = &b }, b.tmaLoadExecAtomType(tma)); + _ = b.makeExecTma(.{ .inner = args.pair.inner, .kernel = &b }, b.tmaLoadExecAtomType(pair_tma)); const sf_ptr = b.intToPtr(0, .f8e8m0fnu, .tmem, 1); - _ = b.makeBlockScaledMma(.{ .n = 16 }, sf_ptr); - _ = b.makeBlockScaledMma(.{ .n = 128 }, sf_ptr); + for ([_]CtaGroup{ .one, .two }) |group| { + _ = b.makeBlockScaledMma(.{ .n = 16, .cta_group = group }, sf_ptr); + _ = b.makeBlockScaledMma(.{ .n = 128, .cta_group = group }, sf_ptr); + } + _ = b.makeBlockScaledMma(.{ .m = 256, .n = 16, .cta_group = .two }, sf_ptr); + _ = b.makeBlockScaledMma(.{ .m = 256, .n = 128, .cta_group = .two }, sf_ptr); const nvfp4_sf_ptr = b.intToPtr(0, .f8e4m3fn, .tmem, 1); _ = b.makeBlockScaledMma(.{ .kind = .nvfp4_nvfp4, .n = 16 }, nvfp4_sf_ptr); const ir = try b.finish(.{ 192, 1, 1 }); defer std.testing.allocator.free(ir); - try std.testing.expect(std.mem.indexOf(u8, ir, "sm100.mma_bs<128x16x32") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "sm100.mma_bs<128x128x32") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "sm100.mma_bs<128x16x64, num_cta = 1, ab_major = (k, k), elem_type = (f4E2M1FN, f4E2M1FN, f32), sf_type = f8E4M3FN, frag_kind = ss, vec_size = 16>") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "cute.make_tiled_mma") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "cute_nvgpu.grid_constant") != null); + try expectContains(ir, "sm100.mma_bs<128x16x32, num_cta = 1"); + try expectContains(ir, "sm100.mma_bs<128x128x32, num_cta = 1"); + try expectContains(ir, "sm100.mma_bs<128x16x32, num_cta = 2"); + try expectContains(ir, "sm100.mma_bs<128x128x32, num_cta = 2"); + try expectContains(ir, "sm100.mma_bs<256x16x32, num_cta = 2"); + try expectContains(ir, "sm100.mma_bs<256x128x32, num_cta = 2"); + try expectContains(ir, "sm100.mma_bs<128x16x64, num_cta = 1, ab_major = (k, k), elem_type = (f4E2M1FN, f4E2M1FN, f32), sf_type = f8E4M3FN, frag_kind = ss, vec_size = 16>"); + try expectContains(ir, "non_exec_tiled_tma_load"); + try expectContains(ir, "value = 3 : i32"); + try expectContains(ir, "value = 32 : i32"); + try expectContains(ir, "!llvm.ptr<7>"); + try expectContains(ir, "#nvvm.mem_scope"); + try expectContains(ir, "#nvvm.mem_scope"); + + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "mbarrier waits preserve acquire scope through polling and blocking retries" { + const ctx = try testContext(); + defer ctx.deinit(); + for ([_]MBarrierScope{ .cta, .cluster }) |scope| { + var b = try Builder.open(std.testing.allocator, ctx, "barrier_waits"); + defer b.deinit(); + _ = try b.declareArgs(.{}); + + const barrier = b.allocSmem(.i64, 1, 8); + const phase = b.constant(.i32, 0); + _ = b.mbarrierTestWait(barrier, phase, .{ .scope = scope }); + _ = b.mbarrierTryWait(barrier, phase, .{ .scope = scope }); + b.mbarrierWait(barrier, phase, .{ .scope = scope, .time_limit = 1234, .test_first = false }); + b.mbarrierWait(barrier, phase, .{ .scope = scope }); + + const ir = try b.finish(.{ 32, 1, 1 }); + defer std.testing.allocator.free(ir); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "#nvvm.mbar_wait")); + try std.testing.expectEqual(@as(usize, 1), std.mem.count(u8, ir, "#nvvm.mbar_wait")); + try expectContains(ir, "value = 1234 : i32"); + try expectContains(ir, "\"scf.if\""); + switch (scope) { + .cta => { + try std.testing.expectEqual(@as(usize, 3), std.mem.count(u8, ir, "#nvvm.mbar_scope")); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "\"nvvm.mbarrier.try_wait.parity\"")); + try std.testing.expectEqual(@as(usize, 0), std.mem.count(u8, ir, "\"scf.while\"")); + }, + .cluster => { + try std.testing.expectEqual(@as(usize, 5), std.mem.count(u8, ir, "#nvvm.mbar_scope")); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "#nvvm.mem_order")); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "\"nvvm.mbarrier.try_wait.parity.timelimit\"")); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "\"scf.while\"")); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "\"scf.condition\"")); + try std.testing.expectEqual(@as(usize, 0), std.mem.count(u8, ir, "#nvvm.mbar_scope")); + }, + } + + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); + } +} + +test "SM100 raw primitives preserve operand types and round-trip through MLIR" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "sm100_raw_primitives"); + defer b.deinit(); + const cfg: TmaConfig = .{ + .dtype = .f8e4m3fn, + .copy_bits = 1024, + .global_basis = b.layoutSpec(.{ 128, 1 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}) }), + .coordinate_layout = b.layoutSpec(.{ 1024, 16 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}) }), + .coordinate_rank = 2, + }; + const a = try b.declareArgs(.{ + .gmem = .{ .ptr = DType.i32 }, + .tma = .{ .mlir_type = .{ .type_ = b.tmaLoadAtomType(cfg), .grid_constant = true } }, + }); + const atom = b.makeExecTma(.{ .inner = a.tma.inner, .kernel = &b }, b.tmaLoadExecAtomType(cfg)); + const desc = b.tmaDescriptorAddress(atom); + const dst = b.constant(.i32, 1024); + const bar = b.constant(.i32, 0); + const zero = b.constant(.i32, 0); + const gmem = b.ptrToInt(a.gmem, .i64); + const pred = b.threadIdx().x.eq(0); + b.tmaGather4(desc, dst, bar, zero, .{ zero, zero.add(1), zero.add(2), zero.add(3) }, .evict_last); + b.cpAsyncBulkPrefetch(gmem, b.constant(.i32, 128), .{}); + b.prefetchL1(gmem); + b.prefetchTensorMap(desc); + const swizzled = Tcgen05SmemDesc.build(&b, dst, .{ .leading_byte_offset = 16, .stride_byte_offset = 1024, .layout = .swizzle_128b }); + const interleaved = Tcgen05SmemDesc.build(&b, dst, .{ .leading_byte_offset = 1024, .stride_byte_offset = 128 }); + b.tcgen05MmaWs(.f16, zero, .{ .tmem = dst }, swizzled, Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 64, .n_dim = 128, .a_major = .k, .b_major = .k }), pred, .{}); + b.tcgen05MmaWs(.f16, zero, .{ .smem = interleaved }, swizzled, Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 64, .n_dim = 256, .a_major = .k, .b_major = .mn }), pred, .{}); + b.setMaxRegisterIncrease(128); + b.barrierCtaSyncAligned(0, 512); + b.setMaxRegisterDecrease(128); + b.barrierCtaSyncAligned(b.threadIdx().x.div(128), 128); + b.mbarrierSpin(bar, zero, .cta); + b.mbarrierSpin(bar, zero, .cluster); + const state = b.mbarrierArrive(bar, .{}).?; + const remote = b.mapDsmemPtr(dst, zero); + _ = remote; + _ = b.mbarrierArrive(bar, .{ .peer_cta_rank_in_cluster = zero, .arrive_count = b.constant(.i32, 1), .scope = .cluster }); + b.storeAsyncDsmem(dst, .{ zero, zero }, bar, zero); + b.storeAsyncDsmem(dst, .{ zero, zero, zero, zero }, bar, zero); + b.barrierClusterArriveRelaxedAligned(); + b.barrierClusterWaitAligned(); + const any = b.voteAnySync(pred, null); + const mask = b.voteBallotSync(pred, null); + const x = b.constant(.f32, 1.0); + const y = b.constant(.f32, 2.0); + const maximum = b.fmax(x, y, .{}); + const sum = b.addPackedF32x2(.{ x, y }, .{ y, x }, .{}); + const product = b.mulPackedF32x2(sum, .{ x, y }, .{}); + const fused = b.fmaPackedF32x2(product, sum, .{ x, y }, .{}); + const scalar = b.fma(maximum, fused[0], fused[1]); + b.store(gmem, .{ scalar.bitCast(.i32), maximum.bitCast(.i32), mask, any.to(.i32), state.to(.i32), b.clock(), b.globalTimer().to(.i32), zero }, .{ .level1_eviction_priority = .evict_no_allocate }); + + const ir = try b.finish(.{ 512, 1, 1 }); + defer std.testing.allocator.free(ir); + for ([_][]const u8{ + "cute_nvgpu.get_tma_desc_addr", "tile::gather4", "nvvm.cp.async.bulk.prefetch", + "prefetch.global.L1", "prefetch.tensormap", "nvvm.tcgen05.mma.ws", + "nvvm.tcgen05.mma.ws", "setmaxnreg.inc", "setmaxnreg.dec", + "llvm.nvvm.barrier.cta.sync.aligned.count", "mbarrier.test_wait.parity.acquire.cta", "mbarrier.test_wait.parity.acquire.cluster", + "nvvm.mapa", "#nvvm.mem_scope", "nvvm.store.async", + "vector<4xi32>", "barrier.cluster.arrive.relaxed.aligned", "nvvm.cluster.wait", + "llvm.nvvm.vote.any.sync", "llvm.nvvm.vote.ballot.sync", "llvm.nvvm.fmax.f", + "nvvm.add.packed.f32x2", "nvvm.mul.packed.f32x2", "nvvm.fma.packed.f32x2", + "llvm.fma.f32", "st.global.L1::no_allocate.v8.b32", "%clock", + "%globaltimer", + }) |op| try expectContains(ir, op); + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "SM100 BF16 weight-stationary descriptor encodes supported shapes and major modes" { + try std.testing.expectEqual(@as(u32, 0x04200490), Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 64, .n_dim = 128, .a_major = .k, .b_major = .k })); + try std.testing.expectEqual(@as(u32, 0x04410490), Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 64, .n_dim = 256, .a_major = .k, .b_major = .mn })); + try std.testing.expectEqual(@as(u32, 0x08118490), Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 128, .n_dim = 64, .a_major = .mn, .b_major = .mn })); + try std.testing.expectEqual(@as(u32, 0x02100490), Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 32, .n_dim = 64, .a_major = .k, .b_major = .k })); +} + +test "SM100 shared descriptors preserve nonzero swizzle bases and boundary offsets" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "smem_descriptors"); + defer b.deinit(); + const args = try b.declareArgs(.{ .out = .{ .ptr = DType.i64 } }); + const addr = b.constant(.i32, 128); + const fields = [_]u64{ + 0x4000404000010000, // sw128, base 0, leading 16 bytes, stride 1024 bytes + 0x4002404000010000, // sw128, pattern at byte 128: base 1 + 0x400e404000010000, // sw128, pattern at byte 896: base 7 + 0x00007fff3fff0000, // unswizzled, largest representable offsets + }; + const descs = [_]Value{ + Tcgen05SmemDesc.build(&b, addr, .{ .leading_byte_offset = 16, .stride_byte_offset = 1024, .base_offset = 0, .layout = .swizzle_128b }), + Tcgen05SmemDesc.build(&b, addr, .{ .leading_byte_offset = 16, .stride_byte_offset = 1024, .base_offset = 1, .layout = .swizzle_128b }), + Tcgen05SmemDesc.build(&b, addr, .{ .leading_byte_offset = 16, .stride_byte_offset = 1024, .base_offset = 7, .layout = .swizzle_128b }), + Tcgen05SmemDesc.build(&b, addr, .{ .leading_byte_offset = 0x3fff0, .stride_byte_offset = 0x3fff0 }), + }; + const out = b.ptrToInt(args.out, .i64); + for (descs, 0..) |desc, i| b.storeWords(2, out.add(@as(i64, @intCast(i * 8))), .global, .{ desc.to(.i32), desc.shrLogical(32).to(.i32) }); + const ir = try b.finish(.{ 32, 1, 1 }); + defer std.testing.allocator.free(ir); + for (fields) |bits| { + const expected = try std.fmt.allocPrint(std.testing.allocator, "value = {d} : i64", .{bits}); + defer std.testing.allocator.free(expected); + try expectContains(ir, expected); + } + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "makeLayout preserves dynamic singleton and nested shape and stride leaves" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "mixed_layout"); + defer b.deinit(); + _ = try b.declareArgs(.{}); + const n = b.blockIdx().x.add(1); + const singleton = b.constant(.i32, 1); + const compact = b.makeLayout(.{ .{ 4, n }, singleton }, .{}); + const explicit = b.makeLayout(.{ n, 16 }, .{ .stride = .{ 1, n } }); + const static_nested = b.makeLayout(.{ .{ 4, 8 }, 16 }, .{}); + _ = compact; + _ = explicit; + _ = static_nested; + const ir = try b.finish(.{ 32, 1, 1 }); + defer std.testing.allocator.free(ir); + try expectContains(ir, "((4,?),?):((1,4),?)"); + try expectContains(ir, "(?,16):(1,?)"); + try expectContains(ir, "((4,8),16):((1,4),32)"); + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "makeLayout preserves 64-bit shape and stride values" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "wide_layout"); + defer b.deinit(); + const args = try b.declareArgs(.{ .extent = .{ .mlir_type = .{ .type_ = DType.i64.toMlir(ctx) } } }); + const wide = args.extent.add(@as(i64, 1) << 32); + const explicit = b.makeLayout(.{ 8, 2 }, .{ .stride = .{ 1, wide } }); + const wide_leaf: AlgebraExpr = .{ .dynamic_int = .{ .width = 64 } }; + try std.testing.expect(explicit.type_().eql(b.layoutType(.{ + .shape = b.expr(.{ 8, 2 }), + .stride = .{ .tuple = &.{ .{ .static = 1 }, wide_leaf } }, + }))); + _ = b.makeLayout(.{ .{ 4, wide }, 1, 8 }, .{ .stride = null }); + const ir = try b.finish(.{ 32, 1, 1 }); + defer std.testing.allocator.free(ir); + try std.testing.expect(std.mem.indexOf(u8, ir, "arith.trunci") == null); + try expectContains(ir, "((4,?{i64}),1,8):((1,4),0,?{i64})"); + try expectContains(ir, "value = 4294967296 : i64"); + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "TMA constructors normalize tilers and staged layouts and copy executes store atoms" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "tma_stores"); + defer b.deinit(); + const cfg: TmaConfig = .{ + .dtype = .f32, + .copy_bits = 2048, + .global_basis = b.layoutSpec(.{ 16, 4 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}) }), + .coordinate_layout = b.layoutSpec(.{ 128, 64 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}) }), + .coordinate_rank = 2, + }; + const args = try b.declareArgs(.{ + .data = .{ .tensor = .{ .dtype = .f32, .shape = &.{ 128, 64 }, .stride = &.{ 1, 128 } } }, + }); + const staged = b.makeLayout(.{ 16, 4, 2 }, .{ .stride = .{ 1, 16, 64 } }); + const store = b.makeTiledTmaAtom(CopyBulkTensorTileS2GOp{}, args.data, staged, .{ 16, 4 }, cfg, .{}); + const smem = b.makeTensorView(b.intToPtr(0, .f32, .smem, 128), b.makeLayout(.{ 16, 4 }, .{})); + const tile = b.makeCoordTensor(.{ 0, 0 }, b.staticLayout(cfg.global_basis)); + b.copy(store.atom, .{smem}, .{tile}, .{}); + b.copy(store.atom, .{smem}, .{tile}, .{ .cache_policy = b.constant(.i64, CacheHintSm90.evict_last) }); + const swizzled = b.makeComposedLayout(.{ .bits = 1, .base = 4, .shift = 3 }, 0, b.layoutSpec(.{ 16, 4, 2 }, .{ 1, 16, 64 })); + _ = b.makeTiledTmaAtom(CopyBulkTensorTileG2SOp{}, args.data, swizzled, b.makeTile(&.{ 16, 4 }), cfg, .{}); + _ = b.makeTiledTmaAtom(CopyBulkTensorTileG2SOp{}, args.data, staged, b.makeLayout(.{ 16, 4 }, .{ .stride = .{ 1, 128 } }), cfg, .{}); + const ir = try b.finish(.{ 32, 1, 1 }); + defer std.testing.allocator.free(ir); + try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, ir, "\"cute_nvgpu.atom.make_exec_tma\"")); + try std.testing.expectEqual(@as(usize, 3), std.mem.count(u8, ir, "\"cute.select\"")); + try expectContains(ir, "(16,4):(1@0,1@1)"); + try expectContains(ir, "tma_store"); + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "CuTe options preserve runtime masks, timed try waits, descriptors and store widths" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "cute_options"); + defer b.deinit(); + const a = try b.declareArgs(.{ .out = .{ .ptr = DType.i32 } }); + const tid = b.threadIdx().x; + const mbar = b.allocSmem(.i64, 1, 8); + b.barrier(.{}); + b.barrier(.{ .barrier_id = tid.div(128), .number_of_threads = b.blockDim().x }); + b.tcgen05Commit(mbar, .{ .cta_group = .two, .multicast_mask = b.clusterCtaRank().shl(1).bitOr(1) }); + _ = b.mbarrierTryWait(mbar, b.constant(.i32, 0), .{ .time_limit = tid.add(100), .scope = .cluster }); + const state = b.mbarrierArrive(mbar, .{ .relaxed = true }).?; + b.storeAsyncDsmem(b.constant(.i32, 128), tid, mbar, b.clusterCtaRank().bitXor(1)); + b.clusterArriveRelaxed(.{}); + b.clusterWait(); + b.clusterArriveRelaxed(.{ .aligned = true }); + b.barrierClusterWaitAligned(); + _ = b.voteAnySync(tid.eq(0), tid.bitOr(1)); + _ = b.voteBallotSync(tid.eq(0), 0xff); + const x = tid.to(.f32); + const y = b.fmax(x, b.constant(.f32, 0), .{ .nan = true, .ftz = true, .abs = true }); + const pair = b.addPackedF32x2(.{ x, y }, .{ y, x }, .{ .rnd = .rz, .ftz = true }); + b.store(a.out, state.to(.i32), .{}); + b.store(a.out, .{ pair[0].bitCast(.i32), pair[1].bitCast(.i32) }, .{}); + const desc = b.makeUmmaSmemDesc(b.intToPtr(0, .bf16, .smem, 16), b.layoutSpec(.{ 64, 16 }, .{ 16, 1 }), .k, null); + const raw_desc = b.smemDescriptorToInt(desc); + _ = b.intToSmemDescriptor(raw_desc); + const mapped = b.mapDsmemPtr(b.intToPtr(128, .i32, .smem, 16), b.clusterCtaRank()); + const mapped_ty = mapped.type_().isA(cute_ir.PtrType).?; + try std.testing.expectEqual(MemorySpace.dsmem, mapped_ty.addressSpace()); + try std.testing.expectEqual(@as(u64, 16), mapped_ty.alignment()); + b.store(mapped, tid, .{}); + const swizzles = [_]Tcgen05SmemSwizzle{ .none, .swizzle_32b, .swizzle_64b, .swizzle_128b, .swizzle_128b_atom_32b }; + for (swizzles) |layout| _ = Tcgen05SmemDesc.build(&b, tid.mul(16), .{ .layout = layout }); + const idesc = b.constant(.i32, Tcgen05InstrDesc.build(.{ .a_dtype = .bf16, .b_dtype = .bf16, .m_dim = 64, .n_dim = 128 })); + b.tcgen05MmaWs(.f16, tid, .{ .tmem = tid }, desc, idesc, tid.ne(0), .{ .collector_b_buffer = .b2, .collector_op = .use, .col_b_zero_mask = state }); + b.cpAsyncBulkPrefetch(a.out, 128, .{ .l2_cache_hint = state }); + const ir = try b.finish(.{ 128, 1, 1 }); + defer std.testing.allocator.free(ir); + try expectContains(ir, "nvvm.mbarrier.try_wait.parity.timelimit"); + try std.testing.expectEqual(@as(usize, 0), std.mem.count(u8, ir, "scf.while")); + try expectContains(ir, "nvvm.store.ext"); + try expectContains(ir, "vector<2xi32>"); + try expectContains(ir, "#nvvm.fp_rnd_mode"); + try expectContains(ir, "cute_nvgpu.make_umma_smem_desc"); + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} + +test "gather4 uses CuTe TMA atom construction, partitioning and copy options" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.openProgram(std.testing.allocator, ctx, "gather4_copy"); + defer b.deinit(); + const cfg: TmaConfig = .{ + .dtype = .f32, + .copy_bits = 2048, + .global_basis = b.layoutSpec(.{ 16, 1 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}) }), + .coordinate_layout = b.layoutSpec(.{ 128, 64 }, .{ b.basis(1, 1, .{0}), b.basis(1, 1, .{1}) }), + .coordinate_rank = 2, + .mode = .gather4, + }; + var pair = cfg; + pair.cta_group = .two; + const args = try b.declareArgs(.{ + .indices = .{ .tensor = .{ .dtype = .i32, .shape = &.{ 128, 64 }, .stride = &.{ 0, 1 } } }, + .atom = .{ .mlir_type = .{ .type_ = b.tmaLoadAtomType(cfg), .grid_constant = true } }, + .pair = .{ .mlir_type = .{ .type_ = b.tmaLoadAtomType(pair), .grid_constant = true } }, + }); + const smem_layout = b.makeLayout(.{ 16, 4 }, .{ .stride = .{ 1, 16 } }); + const smem = b.makeTensorView(b.intToPtr(0, .f32, .smem, 128), smem_layout); + const coords = b.makeCoordTensor(.{ 0, 0 }, b.staticLayout(cfg.coordinate_layout)); + const atom: Atom = .{ .inner = args.atom.inner, .kernel = &b }; + const partitioned = b.tmaPartition(atom, 0, b.makeLayout(1, .{}), smem, .{ coords, args.indices.view() }, smem.type_(), &.{ coords.type_(), args.indices.inner.type_() }); + const mbar = b.allocSmem(.i64, 1, 8); + b.copy(atom, .{ partitioned.targets[0], partitioned.targets[1] }, .{partitioned.smem}, .{ + .tma_bar_ptr = mbar, + .cache_policy = b.constant(.i64, CacheHintSm100.evict_last), + }); + b.copy(.{ .inner = args.pair.inner, .kernel = &b }, .{ partitioned.targets[0], partitioned.targets[1] }, .{partitioned.smem}, .{ .tma_bar_ptr = mbar }); + b.endFunction(.{ 32, 1, 1 }); + + b.beginFunction("gather4_copy", .host); + const host = try b.declareArgs(.{ + .data = .{ .tensor = .{ .dtype = .f32, .shape = &.{ 128, 64 }, .stride = &.{ 1, 128 } } }, + .indices = .{ .tensor = .{ .dtype = .i32, .shape = &.{ 128, 64 }, .stride = &.{ 0, 1 } } }, + }); + const host_layout = b.makeLayout(.{ 16, 4 }, .{ .stride = .{ 1, 16 } }); + const tiler = b.makeTile(&.{ 16, 4 }); + _ = b.makeTiledTmaAtom(CopyBulkTensor2DGather4G2SOp{}, host.data, host_layout, tiler, cfg, .{ .gmem_coord_tensor = host.indices.view() }); + _ = b.makeTiledTmaAtom(CopyBulkTensor2DGather4G2SOp{ .cta_group = .two }, host.data, host_layout, tiler, pair, .{ .gmem_coord_tensor = host.indices.view() }); + b.returnHostStatus(b.constant(.i32, 0)); + b.endFunction(null); + const ir = try b.finishProgram(); + defer std.testing.allocator.free(ir); + try expectContains(ir, "atom.make_non_exec_2d_gather4_tma_load"); + try expectContains(ir, "atom.tma_partition"); + try expectContains(ir, "mode = gather4, num_cta = 1"); + try expectContains(ir, "mode = gather4, num_cta = 2"); + try expectContains(ir, "non_exec_2d_gather4_tma_load", .{structFields(b, n, elem)}) catch @panic("OOM")) catch @panic("invalid inline assembly result type")); + const packed_ = b.emit(llvm.inline_asm(b.ctx, text, inner, .{ .result_type = struct_type, .constraints = constraints, .has_side_effects = side_effects }, b.loc())); + const out = b.arena.allocator().alloc(Value, n) catch @panic("OOM"); + for (out, 0..) |*o, i| { + o.* = b.emit(mlir.Operation.make(b.ctx, "llvm.extractvalue", .{ + .operands = .{ .flat = &.{packed_.inner} }, + .results = .{ .flat = &.{elem} }, + .attributes = &.{.named(b.ctx, "position", .denseArray(b.ctx, .i64, &[_]i64{@intCast(i)}))}, + .location = b.loc(), + })); + } + return out; +} + +pub fn structFields(b: *Builder, n: usize, elem: *const mlir.Type) []const u8 { + var w: std.Io.Writer.Allocating = .init(b.arena.allocator()); + for (0..n) |i| w.writer.print("{s}{f}", .{ if (i == 0) "" else ", ", elem }) catch @panic("OOM"); + return w.written(); +} + +pub fn ldAcq(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.acquire.gpu.global.b32 $0, [$1];", &.{ptr}, .i32, "=r,l,~{memory}", true); +} + +pub fn ldAcqSysU64(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.acquire.sys.global.b64 $0, [$1];", &.{ptr}, .i64, "=l,l,~{memory}", true); +} + +pub fn ldAcqSysS32(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.acquire.sys.global.s32 $0, [$1];", &.{ptr}, .i32, "=r,l,~{memory}", true); +} + +pub fn ldAcqSysU32(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.acquire.sys.global.u32 $0, [$1];", &.{ptr}, .i32, "=r,l,~{memory}", true); +} + +pub fn ldAcqGpuU64(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.acquire.gpu.global.u64 $0, [$1];", &.{ptr}, .i64, "=l,l,~{memory}", true); +} + +pub fn atomicAddU32(b: *Builder, ptr: Value, value: anytype) Value { + return asmValue(b, "atom.global.add.u32 $0, [$1], $2;", &.{ ptr, liftAs(b, value, .i32) }, .i32, "=r,l,r", true); +} + +pub fn atomicAddU64(b: *Builder, ptr: Value, value: Value) Value { + return asmValue(b, "atom.global.add.u64 $0, [$1], $2;", &.{ ptr, value }, .i64, "=l,l,l", true); +} + +pub fn atomicAddSys(b: *Builder, ptr: Value, value: Value) Value { + return asmValue(b, "atom.sys.global.add.u64 $0, [$1], $2;", &.{ ptr, value }, .i64, "=l,l,l", true); +} + +pub fn atomicAddRel(b: *Builder, ptr: Value, value: anytype) Value { + return asmValue(b, "atom.release.gpu.global.add.u32 $0, [$1], $2;", &.{ ptr, liftAs(b, value, .i32) }, .i32, "=r,l,r,~{memory}", true); +} + +pub fn stRelSys(b: *Builder, ptr: Value, value: Value) void { + asmVoid(b, "st.release.sys.global.u64 [$0], $1;", &.{ ptr, value }, "l,l,~{memory}"); +} + +pub fn redAddS32(b: *Builder, ptr: Value, value: anytype) void { + asmVoid(b, "red.gpu.global.add.s32 [$0], $1;", &.{ ptr, liftAs(b, value, .i32) }, "l,r"); +} + +pub fn redAddU32(b: *Builder, ptr: Value, value: anytype) void { + asmVoid(b, "red.gpu.global.add.u32 [$0], $1;", &.{ ptr, liftAs(b, value, .i32) }, "l,r"); +} + +pub fn redXorRel(b: *Builder, ptr: Value, value: Value) void { + asmVoid(b, "red.release.gpu.global.xor.b64 [$0], $1;", &.{ ptr, value }, "l,l,~{memory}"); +} + +pub fn redAddRel(b: *Builder, ptr: Value, value: anytype) void { + asmVoid(b, "red.release.gpu.global.add.u32 [$0], $1;", &.{ ptr, liftAs(b, value, .i32) }, "l,r,~{memory}"); +} + +pub fn redAddRelSys(b: *Builder, ptr: Value, value: anytype) void { + asmVoid(b, "red.release.sys.global.add.s32 [$0], $1;", &.{ ptr, liftAs(b, value, .i32) }, "l,r,~{memory}"); +} + +pub fn atomicAddBlockShared(b: *Builder, smem: Value, value: anytype) Value { + return asmValue(b, "atom.shared.cta.add.u32 $0, [$1], $2;", &.{ smem, liftAs(b, value, .i32) }, .i32, "=r,r,r,~{memory}", true); +} + +pub fn fenceAcqRelCta(b: *Builder) void { + asmVoid(b, "fence.acq_rel.cta;", &.{}, "~{memory}"); +} + +pub fn liftAs(b: *Builder, value: anytype, dt: DType) Value { + if (@TypeOf(value) == Value) return value; + return b.constant(dt, value); +} + +/// `cvta`-style widening of an `i32` shared-memory address to the 64-bit operand +/// `__cvta_generic_to_shared` gives DeepGEMM's `"l"` constraints. +pub fn zext64(b: *Builder, v: Value) Value { + if (v.dtype() == .i64) return v; + return b.emit(arith.extui(b.ctx, v.inner, DType.i64.toMlir(b.ctx), b.loc())); +} + +pub fn i16Of(b: *Builder, v: Value) Value { + return if (v.dtype() == .i16) v else b.emit(arith.trunci(b.ctx, v.inner, DType.i16.toMlir(b.ctx), b.loc())); +} + +pub fn cst32(b: *Builder, v: anytype) Value { + return b.constant(.i32, @as(i32, @bitCast(@as(u32, @truncate(@as(u64, @bitCast(@as(i64, v)))))))); +} + +pub fn cst64(b: *Builder, v: u64) Value { + return b.constant(.i64, @as(i64, @bitCast(v))); +} + +pub fn getSmIdx(b: *Builder) Value { + return asmValue(b, "mov.u32 $0, %smid;", &.{}, .i32, "=r", false); +} + +pub fn getLaneIdx(b: *Builder) Value { + return asmValue(b, "mov.u32 $0, %laneid;", &.{}, .i32, "=r", false); +} + +/// `ptx::get_grid_idx`: the 64-bit `%gridid`, unique per launch. +pub fn getGridIdx(b: *Builder) Value { + return asmValue(b, "mov.u64 $0, %gridid;", &.{}, .i64, "=l", true); +} + +pub fn blockRankInCluster(b: *Builder) Value { + return asmValue(b, "mov.u32 $0, %cluster_ctarank;\n", &.{}, .i32, "=r", true); +} + +/// `clock64()` (`mov.u64 %rd, %clock64;`). +pub fn clock64(b: *Builder) Value { + return asmValue(b, "mov.u64 $0, %clock64;", &.{}, .i64, "=l", true); +} + +/// Keep the original election sequence until typed election matches its small-batch performance. +pub fn electOneSync(b: *Builder) Value { + const pred = asmValue(b, + \\{ + \\.reg .b32 %rx; + \\.reg .pred %px; + \\ mov.u32 $0, 0; + \\ elect.sync %rx|%px, $1; + \\@%px mov.s32 $0, 1; + \\} + , &.{cst32(b, 0xFFFFFFFF)}, .i32, "=r,r", true); + return pred.ne(0); +} + +/// `asm volatile("trap;")` (`DG_TRAP_ONLY_DEVICE_ASSERT`, and this port's timeout). +pub fn trap(b: *Builder) void { + asmVoid(b, "trap;", &.{}, "~{memory}"); +} + +/// `ptx::sync_unaligned(num_threads, barrier_idx)`: `barrier.sync id, n` (threads of a warp may +/// reach it at different program points). +pub fn syncUnaligned(b: *Builder, num_threads: anytype, barrier_idx: anytype) void { + asmVoid(b, "barrier.sync $0, $1;", &.{ liftAs(b, barrier_idx, .i32), liftAs(b, num_threads, .i32) }, "r,r,~{memory}"); +} + +/// `cute::cluster_arrive`. +pub fn clusterArrive(b: *Builder) void { + asmVoid(b, "barrier.cluster.arrive.aligned;\n", &.{}, "~{memory}"); +} + +/// `cudaGridDependencySynchronize()` (`griddepcontrol.wait;`, PDL). +pub fn gridDependencySynchronize(b: *Builder) void { + asmVoid(b, "griddepcontrol.wait;", &.{}, "~{memory}"); +} + +/// `cutlass::arch::warpgroup_reg_alloc`. +pub fn warpgroupRegAlloc(b: *Builder, reg_count: u32) void { + asmVoid(b, std.fmt.allocPrint(b.arena.allocator(), "setmaxnreg.inc.sync.aligned.u32 {d};\n", .{reg_count}) catch @panic("OOM"), &.{}, ""); +} + +/// `cutlass::arch::warpgroup_reg_dealloc`. +pub fn warpgroupRegDealloc(b: *Builder, reg_count: u32) void { + asmVoid(b, std.fmt.allocPrint(b.arena.allocator(), "setmaxnreg.dec.sync.aligned.u32 {d};\n", .{reg_count}) catch @panic("OOM"), &.{}, ""); +} + +/// `ptx::ld_volatile(const uint64_t*)`. +pub fn ldVolatileU64(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.volatile.global.b64 $0, [$1];", &.{ptr}, .i64, "=l,l", true); +} + +/// `ptx::ld_volatile(const uint32_t*)`. +pub fn ldVolatileU32(b: *Builder, ptr: Value) Value { + return asmValue(b, "ld.volatile.global.b32 $0, [$1];", &.{ptr}, .i32, "=r,l", true); +} + +/// `ptx::ld_shared(const uint32_t*)`. +pub fn ldSharedU32(b: *Builder, smem: Value) Value { + return asmValue(b, "ld.shared.u32 $0, [$1];", &.{zext64(b, smem)}, .i32, "=r,l", true); +} + +/// `ptx::ld_shared(const float2*)`. +pub fn ldSharedF32x2(b: *Builder, smem: Value) [2]Value { + return asmValues(b, 2, "ld.shared.v2.f32 {$0, $1}, [$2];", &.{zext64(b, smem)}, .f32, "=f,=f,l", true); +} + +/// `ptx::ld_shared(const float4*)`. +pub fn ldSharedF32x4(b: *Builder, smem: Value) [4]Value { + return asmValues(b, 4, "ld.shared.v4.f32 {$0, $1, $2, $3}, [$4];", &.{zext64(b, smem)}, .f32, "=f,=f,=f,=f,l", true); +} + +/// `ptx::ld_shared(const uint4*)`. +pub fn ldSharedU32x4(b: *Builder, smem: Value) [4]Value { + return asmValues(b, 4, "ld.shared.v4.u32 {$0, $1, $2, $3}, [$4];", &.{zext64(b, smem)}, .i32, "=r,=r,=r,=r,l", true); +} + +/// `ptx::st_shared(const void*, x, y, z, w)`. +pub fn stSharedU32x4(b: *Builder, smem: Value, w: [4]Value) void { + asmVoid(b, "st.shared.v4.u32 [$0], {$1, $2, $3, $4};", &.{ zext64(b, smem), w[0], w[1], w[2], w[3] }, "l,r,r,r,r"); +} + +/// `ptx::st_shared(const uint32_t*, uint32_t)`. +pub fn stSharedU32(b: *Builder, smem: Value, value: Value) void { + asmVoid(b, "st.shared.u32 [$0], $1;", &.{ zext64(b, smem), value }, "l,r"); +} + +/// `ptx::st_shared(const float2*, float2)`. +pub fn stSharedF32x2(b: *Builder, smem: Value, v: [2]Value) void { + asmVoid(b, "st.shared.v2.f32 [$0], {$1, $2};", &.{ zext64(b, smem), v[0], v[1] }, "l,f,f"); +} + +/// `ptx::st_shared_bulk(smem, num_bytes)`: zero `num_bytes` (a multiple of 8) of shared memory. +pub fn stSharedBulk(b: *Builder, smem: Value, num_bytes: u32) void { + std.debug.assert(num_bytes % 8 == 0); + asmVoid(b, "st.bulk.weak.shared::cta [$0], $1, 0;", &.{ zext64(b, smem), b.constant(.i64, num_bytes) }, "l,l,~{memory}"); +} + +/// `ptx::mapa_shared(ptr, dst_cta_idx)`: this CTA's shared address in cluster CTA `cta`. +pub fn mapaShared(b: *Builder, smem: Value, dst_cta_idx: Value) Value { + return asmValue(b, "mapa.shared::cluster.u32 $0, $1, $2;", &.{ smem, dst_cta_idx }, .i32, "=r,r,r", true); +} + +/// `ptx::st_async_cluster(dst, src, dst_cta_idx, barrier)` for `words.len / 4` 16-byte +/// chunks: `st.async` into CTA `dst_cta_idx`'s shared memory, completing the bytes on its +/// mbarrier at `barrier` (both local addresses, mapped with `mapa`). +pub fn stAsyncCluster(b: *Builder, dst: Value, words: []const Value, dst_cta_idx: Value, barrier: Value) void { + std.debug.assert(words.len % 4 == 0); + const mapped_barrier = mapaShared(b, barrier, dst_cta_idx); + const mapped_dst = mapaShared(b, dst, dst_cta_idx); + var i: usize = 0; + while (i < words.len) : (i += 4) { + asmVoid( + b, + "st.async.shared::cluster.mbarrier::complete_tx::bytes.u32.v4 [$0], {$1, $2, $3, $4}, [$5];", + &.{ mapped_dst.add(@as(i32, @intCast(i * 4))), words[i], words[i + 1], words[i + 2], words[i + 3], mapped_barrier }, + "r,r,r,r,r,r,~{memory}", + ); + } +} + +/// `ptx::SM100_U8x4_STSM_T::copy(src, smem)`: `stmatrix.m16n8.x1.trans.b8`. +pub fn stmatrixU8x4T(b: *Builder, src: Value, smem: Value) void { + asmVoid(b, "stmatrix.sync.aligned.m16n8.x1.trans.shared.b8 [$0], {$1};\n", &.{ zext64(b, smem), src }, "l,r"); +} + +/// `ptx::SM90_U32x4_STSM_T::copy(s0, s1, s2, s3, smem)`: `stmatrix.x4.m8n8.trans.b16`. +pub fn stmatrixU32x4T(b: *Builder, src: [4]Value, smem: Value) void { + asmVoid(b, "stmatrix.sync.aligned.x4.m8n8.shared.b16.trans [$0], {$1, $2, $3, $4};\n", &.{ zext64(b, smem), src[0], src[1], src[2], src[3] }, "l,r,r,r,r"); +} + +/// `ClusterBarrier::arrive()` on this CTA's barrier. +pub fn mbarrierArrive(b: *Builder, bar: Value) void { + asmVoid(b, "{\n\tmbarrier.arrive.shared::cta.b64 _, [$0];\n\t}", &.{bar}, "r,~{memory}"); +} + +/// `ClusterBarrier::arrive(cta_id, pred = true)`: arrive on the barrier at the same offset of +/// cluster CTA `cta_id`. +pub fn mbarrierArriveCluster(b: *Builder, bar: Value, cta_id: Value) void { + asmVoid( + b, + "{\n\t.reg .b32 remAddr32;\n\tmapa.shared::cluster.u32 remAddr32, $0, $1;\n\tmbarrier.arrive.shared::cluster.b64 _, [remAddr32];\n\t}", + &.{ bar, cta_id }, + "r,r,~{memory}", + ); +} + +/// `ClusterTransactionBarrier::arrive_and_expect_tx(bytes, cta_id, pred)`: on cluster CTA +/// `cta_id`'s barrier, only when `pred == 1`. +pub fn mbarrierArriveAndExpectTxCluster(b: *Builder, bar: Value, transaction_bytes: anytype, cta_id: Value, pred: Value) void { + asmVoid( + b, + "{\n\t.reg .pred p;\n\t.reg .b32 remAddr32;\n\tsetp.eq.u32 p, $2, 1;\n\t@p mapa.shared::cluster.u32 remAddr32, $0, $1;\n\t@p mbarrier.arrive.expect_tx.shared::cluster.b64 _, [remAddr32], $3;\n\t}", + &.{ bar, cta_id, pred, liftAs(b, transaction_bytes, .i32) }, + "r,r,r,r,~{memory}", + ); +} + +/// `cute::prefetch_tma_descriptor(desc)` (generic descriptor address). +pub fn prefetchTmaDescriptor(b: *Builder, desc: Value) void { + asmVoid(b, "prefetch.tensormap [$0];", &.{desc}, "l,~{memory}"); +} + +/// `SM100_TMA_2SM_LOAD_2D::copy(desc, mbar, cache_hint, smem, c0, c1)`: issued by both CTAs of +/// the pair; the barrier address is masked so the bytes complete on the leader's barrier. +pub fn tma2SmLoad2d(b: *Builder, desc: Value, mbar: Value, cache_hint: u64, smem: Value, c0: Value, c1: Value) void { + asmVoid( + b, + "cp.async.bulk.tensor.2d.cta_group::2.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [$0], [$1, {$3, $4}], [$2], $5;", + &.{ smem, desc, mbar.bitAnd(cst32(b, sm100_mma_peer_bit_mask)), c0, c1, cst64(b, cache_hint) }, + "r,l,r,r,r,l,~{memory}", + ); +} + +/// `SM100_TMA_2SM_LOAD_4D::copy(desc, mbar, cache_hint, smem, c0, c1, c2, c3)`. +pub fn tma2SmLoad4d(b: *Builder, desc: Value, mbar: Value, cache_hint: u64, smem: Value, c: [4]Value) void { + asmVoid( + b, + "cp.async.bulk.tensor.4d.cta_group::2.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [$0], [$1, {$3, $4, $5, $6}], [$2], $7;", + &.{ smem, desc, mbar.bitAnd(cst32(b, sm100_mma_peer_bit_mask)), c[0], c[1], c[2], c[3], cst64(b, cache_hint) }, + "r,l,r,r,r,r,r,l,~{memory}", + ); +} + +/// `ptx::tma_load_1d(dst, src, mbar, num_bytes, hint = EVICT_FIRST)`: non-tensor bulk copy of +/// `num_bytes` from a global (possibly peer) address into this CTA's shared memory. +pub fn tmaLoad1d(b: *Builder, dst: Value, src: Value, mbar: Value, num_bytes: anytype, cache_hint: u64) void { + asmVoid( + b, + "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [$0], [$1], $2, [$3], $4;\n", + &.{ dst, src, liftAs(b, num_bytes, .i32), mbar, cst64(b, cache_hint) }, + "r,l,r,r,l,~{memory}", + ); +} + +/// `ptx::tma_store_1d(dst, src, num_bytes, hint = EVICT_NORMAL)`. +pub fn tmaStore1d(b: *Builder, dst: Value, src: Value, num_bytes: anytype, cache_hint: u64) void { + asmVoid( + b, + "cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint [$0], [$1], $2, $3;\n", + &.{ dst, src, liftAs(b, num_bytes, .i32), cst64(b, cache_hint) }, + "l,r,r,l,~{memory}", + ); +} + +/// `cute::SM90_TMA_STORE_2D::copy(desc, smem, c0, c1)`. +pub fn tmaStore2d(b: *Builder, desc: Value, smem: Value, c0: Value, c1: Value) void { + asmVoid(b, "cp.async.bulk.tensor.2d.global.shared::cta.bulk_group [$0, {$2, $3}], [$1];", &.{ desc, smem, c0, c1 }, "l,r,r,r,~{memory}"); +} + +/// `cute::TMEM::Allocator2Sm::allocate(num_columns, dst)`: write the base column to shared `dst`. +pub fn tmemAllocate2Sm(b: *Builder, num_columns: u32, dst: Value) void { + asmVoid(b, "tcgen05.alloc.cta_group::2.sync.aligned.shared::cta.b32 [$0], $1;", &.{ dst, cst32(b, num_columns) }, "r,r,~{memory}"); +} + +/// `cute::TMEM::Allocator2Sm::free(tmem_ptr, num_columns)`. +pub fn tmemFree2Sm(b: *Builder, tmem_ptr: Value, num_columns: u32) void { + asmVoid(b, "{\n\ttcgen05.dealloc.cta_group::2.sync.aligned.b32 $0, $1; \n\t}", &.{ tmem_ptr, cst32(b, num_columns) }, "r,r,~{memory}"); +} + +/// `ptx::tcgen05_before_thread_sync`. +pub fn tcgen05BeforeThreadSync(b: *Builder) void { + asmVoid(b, "tcgen05.fence::before_thread_sync;", &.{}, "~{memory}"); +} + +/// `ptx::tcgen05_after_thread_sync`. +pub fn tcgen05AfterThreadSync(b: *Builder) void { + asmVoid(b, "tcgen05.fence::after_thread_sync;", &.{}, "~{memory}"); +} + +/// `cutlass::arch::fence_view_async_tmem_load` (`tcgen05.wait::ld`). +pub fn fenceViewAsyncTmemLoad(b: *Builder) void { + asmVoid(b, "{\n\ttcgen05.wait::ld.sync.aligned; \n}", &.{}, "~{memory}"); +} + +/// `fence_view_async_tmem_load` with the values of earlier TMEM loads threaded through it, so +/// LLVM cannot move their uses above the wait (the PTX is the same single instruction). +pub fn fenceViewAsyncTmemLoadValues(b: *Builder, comptime n: usize, values: [n]Value) [n]Value { + comptime var constraints: []const u8 = ""; + inline for (0..n) |i| constraints = constraints ++ (if (i == 0) "" else ",") ++ "=r"; + inline for (0..n) |i| constraints = constraints ++ std.fmt.comptimePrint(",{d}", .{i}); + constraints = constraints ++ ",~{memory}"; + return asmValues(b, n, "{\n\ttcgen05.wait::ld.sync.aligned; \n}", &values, .i32, constraints, true); +} + +/// `fenceViewAsyncTmemLoadValues` for a run-time number of values (the epilogue's store blocks +/// hold `STORE_BLOCK_M / ATOM_M * 8` words): the same single `tcgen05.wait::ld`, every value +/// tied through it. +pub fn fenceViewAsyncTmemLoadSlice(b: *Builder, values: []const Value) []Value { + const n = values.len; + const a = b.arena.allocator(); + var w: std.Io.Writer.Allocating = .init(a); + for (0..n) |i| w.writer.print("{s}=r", .{if (i == 0) "" else ","}) catch @panic("OOM"); + for (0..n) |i| w.writer.print(",{d}", .{i}) catch @panic("OOM"); + w.writer.writeAll(",~{memory}") catch @panic("OOM"); + return asmValuesSlice(b, n, "{\n\ttcgen05.wait::ld.sync.aligned; \n}", values, .i32, w.written(), true); +} + +/// `cute::SM100_TMEM_LOAD_16dp256b1x::copy(src, d0, d1, d2, d3)`: lane t of the warp gets +/// rows `base + t/4` (d0, d1) and `base + 8 + t/4` (d2, d3), columns `2(t%4)`, `2(t%4)+1`. +pub fn tmemLoad16dp256b1x(b: *Builder, taddr: Value) [4]Value { + return asmValues(b, 4, "tcgen05.ld.sync.aligned.16x256b.x1.b32{$0, $1, $2, $3},[$4];\n", &.{taddr}, .i32, "=r,=r,=r,=r,r", true); +} + +/// `cute::SM100_UTCCP_4x32dp128bit_2cta::copy(src_desc, dst_taddr)`. +pub fn utccp4x32dp128bit2cta(b: *Builder, src_desc: Value, dst_taddr: Value) void { + asmVoid(b, "tcgen05.cp.cta_group::2.32x128b.warpx4 [$0], $1;", &.{ dst_taddr, src_desc }, "r,l"); +} + +/// `ptx::SM100_MMA_MXF8F6F4_2x1SM_SS::fma(desc_a, desc_b, tmem_c, scale_c, desc, tmem_sfa, +/// tmem_sfb)`: `desc` is the 64-bit runtime descriptor (`instr << 32`), `scale_c` (i32) +/// nonzero accumulates. +pub fn mmaMxf8f6f42x1SmSs(b: *Builder, desc_a: Value, desc_b: Value, tmem_c: Value, scale_c: Value, desc: Value, tmem_sfa: Value, tmem_sfb: Value) void { + const idesc = b.emit(arith.trunci(b.ctx, desc.shrLogical(32).inner, DType.i32.toMlir(b.ctx), b.loc())); + asmVoid( + b, + "{\n\t.reg .pred p;\n\tsetp.ne.b32 p, $4, 0;\n\ttcgen05.mma.cta_group::2.kind::mxf8f6f4.block_scale [$0], $1, $2, $3, [$5], [$6], p; \n\t}\n", + &.{ tmem_c, desc_a, desc_b, idesc, scale_c, tmem_sfa, tmem_sfb }, + "r,l,l,r,r,r,r", + ); +} + +/// `cutlass::arch::umma_arrive_multicast_2x1SM(barrier, cta_mask)`: inside `elect_one_sync`, +/// `tcgen05.commit` arriving on `barrier` in every CTA of `cta_mask`. +pub fn ummaArriveMulticast2x1Sm(b: *Builder, bar: Value, cta_mask: u16) void { + var elected = b.openIf(electOneSync(b)); + { + asmVoid( + b, + "{\n\ttcgen05.commit.cta_group::2.mbarrier::arrive::one.shared::cluster.multicast::cluster.b64 [$0], $1; \n\t}", + &.{ bar, b.constant(.i16, @as(i16, @bitCast(cta_mask))) }, + "r,h,~{memory}", + ); + elected.close(); + } +} + +pub fn bits32(v: Value) Value { + return if (v.dtype() == .i32) v else v.bitCast(.i32); +} + +pub fn fromBits32(v: Value, dt: DType) Value { + return if (dt == .i32) v else v.bitCast(dt); +} + +/// `__shfl_sync(0xffffffff, v, src_lane)` / `ptx::exchange` (32-bit). +pub fn shflSync(b: *Builder, v: Value, src_lane: anytype) Value { + const r = asmValue(b, "shfl.sync.idx.b32 $0, $1, $2, 31, -1;", &.{ bits32(v), liftAs(b, src_lane, .i32) }, .i32, "=r,r,r", true); + return fromBits32(r, v.dtype()); +} + +/// `__shfl_xor_sync(0xffffffff, v, lane_mask)`. +pub fn shflXorSync(b: *Builder, v: Value, lane_mask: anytype) Value { + const r = asmValue(b, "shfl.sync.bfly.b32 $0, $1, $2, 31, -1;", &.{ bits32(v), liftAs(b, lane_mask, .i32) }, .i32, "=r,r,r", true); + return fromBits32(r, v.dtype()); +} + +/// `__shfl_up_sync(0xffffffff, v, delta)`. +pub fn shflUpSync(b: *Builder, v: Value, delta: anytype) Value { + const r = asmValue(b, "shfl.sync.up.b32 $0, $1, $2, 0, -1;", &.{ bits32(v), liftAs(b, delta, .i32) }, .i32, "=r,r,r", true); + return fromBits32(r, v.dtype()); +} + +/// `__reduce_add_sync(0xffffffff, v)` (u32). +pub fn reduceAddSync(b: *Builder, v: Value) Value { + return asmValue(b, "redux.sync.add.u32 $0, $1, $2;", &.{ v, cst32(b, 0xFFFFFFFF) }, .i32, "=r,r,r", true); +} + +/// `__reduce_min_sync(0xffffffff, v)` (u32). +pub fn reduceMinSync(b: *Builder, v: Value) Value { + return asmValue(b, "redux.sync.min.u32 $0, $1, $2;", &.{ v, cst32(b, 0xFFFFFFFF) }, .i32, "=r,r,r", true); +} + +/// `__all_sync(0xffffffff, pred)` (`pred` i1; returns i1). +pub fn allSync(b: *Builder, pred: Value) Value { + const r = asmValue(b, "{ .reg .pred p, q; setp.ne.u32 p, $1, 0; vote.sync.all.pred q, p, $2; selp.u32 $0, 1, 0, q; }", &.{ pred.to(.i32), cst32(b, 0xFFFFFFFF) }, .i32, "=r,r,r", true); + return r.ne(0); +} + +pub fn popc(b: *Builder, v: Value) Value { + return b.callIntrinsic("llvm.ctpop.i32", &.{v}, DType.i32.toMlir(b.ctx)).?; +} + +/// `__ffs(v)`: 1-based index of the lowest set bit, 0 for 0 (nvcc: `brev` + `bfind.shiftamt`). +pub fn ffs(b: *Builder, v: Value) Value { + return asmValue(b, "{ .reg .b32 r; brev.b32 r, $1; bfind.shiftamt.u32 $0, r; }", &.{v}, .i32, "=r,r", false).add(1); +} + +/// `__fns(mask, base, offset)`: position of the `offset`-th set bit of `mask` from `base`. +pub fn fns(b: *Builder, mask: Value, base: anytype, offset: Value) Value { + return asmValue(b, "fns.b32 $0, $1, $2, $3;", &.{ mask, liftAs(b, base, .i32), offset }, .i32, "=r,r,r,r", false); +} + +/// `fminf(a, b)` (`min.f32`, NaN-ignoring; `kClampAfterSilu`). +pub fn fminf(b: *Builder, x: Value, y: Value) Value { + return asmValue(b, "min.f32 $0, $1, $2;", &.{ x, y }, .f32, "=f,f,f", false); +} + +/// `fabsf(x)` (`abs.f32`). +pub fn fabsf(b: *Builder, x: Value) Value { + return asmValue(b, "abs.f32 $0, $1;", &.{x}, .f32, "=f,f", false); +} + +/// `math::fast_rcp(x)`. +pub fn fastRcp(b: *Builder, x: Value) Value { + return asmValue(b, "rcp.approx.ftz.f32 $0, $1;", &.{x}, .f32, "=f,f", true); +} + +/// `__expf(x)` without `--use_fast_math`: `ex2.approx.f32(x * log2(e))` (non-FTZ). +pub fn expfFast(b: *Builder, x: Value) Value { + return asmValue(b, "{ .reg .f32 t; mul.f32 t, $1, 0f3FB8AA3B; ex2.approx.f32 $0, t; }", &.{x}, .f32, "=f,f", false); +} + +/// `__expf(-x)`: nvcc folds the negation into the constant (`mul.f32 t, x, -log2(e)`). +pub fn expfFastNeg(b: *Builder, x: Value) Value { + return asmValue(b, "{ .reg .f32 t; mul.f32 t, $1, 0fBFB8AA3B; ex2.approx.f32 $0, t; }", &.{x}, .f32, "=f,f", false); +} + +/// `__float22bfloat162_rn({lo, hi})`: bf16x2 word, `lo` in the low half. +pub fn float22bfloat162Rn(b: *Builder, lo: Value, hi: Value) Value { + return asmValue(b, "{ cvt.rn.bf16x2.f32 $0, $2, $1;}\n", &.{ lo, hi }, .i32, "=r,f,f", false); +} + +/// `__hmin2(a, b)` on bf16x2 words. +pub fn hmin2Bf16(b: *Builder, x: Value, y: Value) Value { + return asmValue(b, "{ min.bf16x2 $0,$1,$2;\n}", &.{ x, y }, .i32, "=r,r,r", false); +} + +/// `__hmax2(a, b)` on bf16x2 words. +pub fn hmax2Bf16(b: *Builder, x: Value, y: Value) Value { + return asmValue(b, "{ max.bf16x2 $0,$1,$2;\n}", &.{ x, y }, .i32, "=r,r,r", false); +} + +/// `__bfloat1622float2(word)`: `{low, high}` f32 (exact). +pub fn bfloat1622float2(b: *Builder, word: Value) [2]Value { + return asmValues(b, 2, "{ .reg .b16 l, h; mov.b32 {l, h}, $2; cvt.f32.bf16 $0, l; cvt.f32.bf16 $1, h; }", &.{word}, .f32, "=f,=f,r", false); +} + +/// `ptx::accumulate(float2& a, nv_bfloat162 b)`: `a.x += bf16(b.x)`, `a.y += bf16(b.y)` with +/// `add.rn.f32.bf16` (one rounding, the bf16 widened exactly). +pub fn accumulateBf16x2(b: *Builder, acc: [2]Value, bf16x2: Value) [2]Value { + const lo = i16Of(b, bf16x2); + const hi = i16Of(b, bf16x2.shrLogical(16)); + return .{ + asmValue(b, "add.rn.f32.bf16 $0, $1, $0;\n", &.{ lo, acc[0] }, .f32, "=f,h,0", false), + asmValue(b, "add.rn.f32.bf16 $0, $1, $0;\n", &.{ hi, acc[1] }, .f32, "=f,h,0", false), + }; +} +/// IEEE operations used when staging quantized inputs; no FTZ or contraction. +pub fn divRn(b: *Builder, x: Value, y: Value) Value { + return asmValue(b, "div.rn.f32 $0, $1, $2;", &.{ x, y }, .f32, "=f,f,f", false); +} +pub fn mulRn(b: *Builder, x: Value, y: Value) Value { + return asmValue(b, "mul.rn.f32 $0, $1, $2;", &.{ x, y }, .f32, "=f,f,f", false); +} + +test "canonical primitives preserve aligned barriers and runtime transaction counts" { + const ctx = try testContext(); + defer ctx.deinit(); + var b = try Builder.open(std.testing.allocator, ctx, "compatibility_primitives"); + defer b.deinit(); + _ = try b.declareArgs(.{}); + const bar = b.constant(.i32, 0); + const phase = b.constant(.i32, 0); + const bytes = b.threadIdx().x.add(16); + b.barrierCtaSyncAligned(1, 32); + b.barrierClusterArriveRelaxedAligned(); + b.barrierClusterWaitAligned(); + b.mbarrierInit(bar, 1); + b.fenceMbarrierInit(); + b.mbarrierArriveExpectTx(bar, bytes); + b.mbarrierArriveExpectTx(bar, 16); + b.mbarrierWait(bar, phase, .{ .test_first = false }); + _ = phase.bitXor(1); + b.fenceProxyAsyncShared(); + b.fenceProxyAsyncShared(); + b.cpAsyncBulkCommitGroup(); + b.cpAsyncBulkWaitGroup(0, false); + _ = b.voteBallotSync(bytes.ne(0), null); + _ = b.fmax(b.constant(.f32, 1), b.constant(.f32, 2), .{}); + const ir = try b.finish(.{ 32, 1, 1 }); + defer std.testing.allocator.free(ir); + try expectContains(ir, "llvm.nvvm.barrier.cta.sync.aligned.count"); + try expectContains(ir, "barrier.cluster.arrive.relaxed.aligned"); + try expectContains(ir, "nvvm.mbarrier.arrive.expect_tx"); + const parsed = try mlir.Module.parse(ctx, ir); + defer parsed.deinit(); + try std.testing.expect(parsed.operation().verify()); +} diff --git a/dsl/fly/README.md b/dsl/fly/README.md index b886988b..687c99dc 100644 --- a/dsl/fly/README.md +++ b/dsl/fly/README.md @@ -81,7 +81,7 @@ const c = VectorAdd.call(.{ .a = a, .b = bt }, .{ .c = a.shape() }, .{ ``` Complete vector-add and tiled-MMA kernels are the tests in `builder.zig` -and `zml/kernel.zig`. +and `zml/kernel/fly/test.zig`. Model kernels live beside their consumers: diff --git a/dsl/tcl/README.md b/dsl/tcl/README.md index ec1ae61f..b01b7189 100644 --- a/dsl/tcl/README.md +++ b/dsl/tcl/README.md @@ -104,8 +104,7 @@ def softmax_kernel(B: Axis = 2, D: Axis = 1024, S: Axis = 16, v0: [B, S, D]/bf16 across the partitions; other kernels stay per chip. Its arguments and results are DRAM tensors: `Dram{ .chip = &.{.{ .axis = C }} }` splits axis `C`, sized to the chip count, across chips (each partition holds its share), - the default `.broadcast` keeps a copy per chip. Call it inside - `zml.ops.manualComputation` so it sees each partition's shards. + the default `.broadcast` keeps a copy per chip. * **Convolutions** pad and slide the input in its read: a slide replaces its axis by `window_axis, frame_axis`, then `contract` with the filter. On its own (without a slide) the padding is dropped. diff --git a/examples/io/main.zig b/examples/io/main.zig index 1cec23b0..c9dc05ea 100644 --- a/examples/io/main.zig +++ b/examples/io/main.zig @@ -132,13 +132,9 @@ pub fn main(init: std.process.Init) !void { const root = try TensorNode.init(init.arena.allocator(), ""); - var registry_it = registry.iterator(); - while (registry_it.next()) |kv| { - const name = kv.key_ptr.*; - const tensor = kv.value_ptr.*; - + for (registry.tensors.values()) |tensor| { var current = root; - var parts = std.mem.tokenizeScalar(u8, name, '.'); + var parts = std.mem.tokenizeScalar(u8, tensor.name, '.'); while (parts.next()) |part| { const gop = try current.children.getOrPut(current.allocator, part); if (!gop.found_existing) { @@ -183,16 +179,16 @@ pub fn main(init: std.process.Init) !void { const tensors = try allocator.alloc(zml.Tensor, tensor_count); defer allocator.free(tensors); - var registry_it = registry.iterator(); var load_count: usize = 0; - while (registry_it.next()) |entry| : (load_count += 1) { + for (registry.tensors.keys(), registry.tensors.values()) |name, tensor| { tensors[load_count] = switch (mesh_type) { - .replicated => store.view().createReplicatedTensor(entry.key_ptr.*, null), - .sharded => if (entry.value_ptr.shape.rank() > 0) - store.view().createTensor(entry.key_ptr.*, null, .tp, tp_partitioning) + .replicated => store.view().createReplicatedTensor(name, null), + .sharded => if (tensor.shape.rank() > 0) + store.view().createTensor(name, null, .tp, tp_partitioning) else - store.view().createReplicatedTensor(entry.key_ptr.*, null), + store.view().createReplicatedTensor(name, null), }; + load_count += 1; } const model: AllTensorsModel = .{ .tensors = tensors }; diff --git a/examples/llm/models/lfm2/model.zig b/examples/llm/models/lfm2/model.zig index dd21c7f9..9f7cb909 100644 --- a/examples/llm/models/lfm2/model.zig +++ b/examples/llm/models/lfm2/model.zig @@ -571,6 +571,8 @@ pub const Attention = struct { v = new_kv_cache.values(cache_index); stdx.debug.assert(q.dim(.batch) == 1, "LFM attention currently expects batch size 1 for flash attention backend, got {}", .{q.dim(.batch)}); + const mesh = zml.Compiler.current().mesh(.model); + const partitioning: zml.attention.Partitioning = .head(.model); const attn = switch (attention_parameters) { .stablehlo => zml.attention.attention( q, @@ -579,6 +581,8 @@ pub const Attention = struct { tokens_position_offset, attention_metadata, attention_parameters, + mesh, + partitioning, ).merge(.{ .d = .{ .h, .hd } }), else => zml.attention.attention( q.squeeze(.batch), @@ -587,6 +591,8 @@ pub const Attention = struct { tokens_position_offset.squeeze(.batch), attention_metadata, attention_parameters, + mesh, + partitioning, ).merge(.{ .d = .{ .h, .hd } }) .rename(.{ .q = .seq }) .insertAxes(.seq, .{.batch}), diff --git a/examples/llm/models/llama/model.zig b/examples/llm/models/llama/model.zig index 05610dc1..8a6a09e9 100644 --- a/examples/llm/models/llama/model.zig +++ b/examples/llm/models/llama/model.zig @@ -329,7 +329,7 @@ pub const EmbedTokens = struct { const tokens = input.tokens.withPartialTags(.{.s}); return .{ .hidden = input.embedding.embed_tokens.forward(tokens) .withPartialTags(.{.d}) - .withPartitioning(.model, .{ .d = .replicated }) }; + .replicate() }; } }; @@ -587,6 +587,8 @@ const SelfAttn = struct { token_index, attention_metadata, attention_parameters, + zml.Compiler.current().mesh(.model), + .head(.model), ); const attn = attn_output.merge(.{ .d = .{ .h, .hd } }).rename(.{ .q = .s }); diff --git a/examples/llm/models/qwen3_5/model.zig b/examples/llm/models/qwen3_5/model.zig index 77d3884e..40174f5c 100644 --- a/examples/llm/models/qwen3_5/model.zig +++ b/examples/llm/models/qwen3_5/model.zig @@ -290,7 +290,7 @@ pub const EmbedTokens = struct { const tokens = input.tokens.withPartialTags(.{.s}); return .{ .hidden = input.embedding.embed_tokens.forward(tokens) .withPartialTags(.{.d}) - .withPartitioning(.model, .{ .d = .replicated }) }; + .replicate() }; } }; @@ -657,6 +657,10 @@ pub const SelfAttn = struct { k = partitionCachedKv(k, kv_head_sharding); v = partitionCachedKv(v, kv_head_sharding); + const head_partitioning: zml.attention.Partitioning = switch (kv_head_sharding) { + .sharded => .head(.model), + .replicated => .head(.replicated), + }; const attn_output = zml.attention.attention( q, k, @@ -664,6 +668,8 @@ pub const SelfAttn = struct { token_index, zml.attention.Metadata.init(.fromBackend(.stablehlo)), zml.attention.Parameters.init(.fromBackend(.stablehlo)), + zml.Compiler.current().mesh(.model), + head_partitioning, ).rename(.{ .q = .s }).merge(.{ .d_out_proj = .{ .h, .hd } }); const gated_output = attn_output.mul(gate.sigmoid()); diff --git a/examples/llm/models/qwen3_5_moe/model.zig b/examples/llm/models/qwen3_5_moe/model.zig index 11ce8151..2a3b0ba8 100644 --- a/examples/llm/models/qwen3_5_moe/model.zig +++ b/examples/llm/models/qwen3_5_moe/model.zig @@ -641,6 +641,10 @@ pub const SelfAttn = struct { k = partitionCachedKv(k, kv_head_sharding); v = partitionCachedKv(v, kv_head_sharding); + const head_partitioning: zml.attention.Partitioning = switch (kv_head_sharding) { + .sharded => .head(.model), + .replicated => .head(.replicated), + }; const attn_output = zml.attention.attention( q, k, @@ -648,7 +652,9 @@ pub const SelfAttn = struct { token_index, zml.attention.Metadata.init(.fromBackend(.stablehlo)), zml.attention.Parameters.init(.fromBackend(.stablehlo)), - ).withPartitioning(tp, .{ .q = .replicated, .h = .model, .hd = .replicated }).rename(.{ .q = .s }).merge(.{ .d_out_proj = .{ .h, .hd } }); + zml.Compiler.current().mesh(.model), + head_partitioning, + ).rename(.{ .q = .s }).merge(.{ .d_out_proj = .{ .h, .hd } }); const gated_output = attn_output.mul(gate.sigmoid()); const projected_output = self.o_proj.forward(gated_output.rename(.{ .d_out_proj = .d }), gated_output.dtype()).rename(.{ .dout = .d }).withPartitioning(tp, .{ .d = .replicated }); @@ -781,6 +787,8 @@ pub const Moe = struct { self.gate_up_proj, self.down_proj, moe_backend, + zml.Compiler.current().mesh(.experts), + .expert(.experts), .{ .activation = .{ .swiglu = .{} }, .quantize_input = false, diff --git a/examples/mnist/mnist_data_parallel.zig b/examples/mnist/mnist_data_parallel.zig index a539ae5a..b7805e49 100644 --- a/examples/mnist/mnist_data_parallel.zig +++ b/examples/mnist/mnist_data_parallel.zig @@ -2,9 +2,11 @@ //! Separated to kept `mnist.zig` minimal const std = @import("std"); + const zml = @import("zml"); const base = @import("mnist.zig"); + const log = std.log.scoped(.mnist); pub const std_options: std.Options = .{ @@ -53,14 +55,20 @@ const Mnist = struct { /// just two linear layers + relu activation pub fn forward(self: Mnist, input: zml.Tensor) zml.Tensor { var x = input.onMemory(.host_pinned).toMemory(.device).merge(.{ .d = .{ .x, .y } }).convert(.f32); + const layers: []const Layer = &.{ self.fc1, self.fc2 }; + const compiler = zml.Compiler.current(); + for (0.., layers) |i, layer| { compiler.pushLocationFmt(@src(), "layer[{d}]", .{i}); defer compiler.popLocation(); x = layer.forward(x); } - return x.argMax(.d).indices.convert(.u8).withPartitioning(input.shape()._sharding.mesh.?, .{ .b = .data }).toMemory(.host_pinned); + + return x.argMax(.d).indices.convert(.u8) + .withPartitioning(.dp_mp, .{ .b = .data, .d = .replicated }) + .toMemory(.host_pinned); } const Layer = struct { @@ -68,7 +76,7 @@ const Mnist = struct { bias: zml.Tensor, pub fn forward(self: Layer, input: zml.Tensor) zml.Tensor { - log.warn("layer(input={f}, w={f}, b={f}) -> {f}", .{ input, self.weight, self.bias, input.dot(self.weight, .d) }); + log.info("Layer(input={f}, w={f}, b={f}) -> {f}", .{ input, self.weight, self.bias, input.dot(self.weight, .d) }); const x = input.dot(self.weight, .d); return x.add(self.bias.broad(x.shape())).relu().rename(.{ .d_out = .d }); } @@ -100,18 +108,20 @@ pub fn main(init: std.process.Init) !void { &.{ .low_bandwidth, .high_bandwidth }, &.{ 2, -1 }, )); - log.info("topology used: {f}", .{dp_mp}); + log.info("{f}", .{dp_mp}); var profiler = try platform.profiler(allocator, io, .{ .session_id = "mnist" }); defer profiler.deinit(); var store: zml.io.TensorStore = .fromRegistry(allocator, ®istry, &.{dp_mp}); defer store.deinit(); + const mnist_model: Mnist = .init(store.view()); - // // Compile model + // Compile model const bs: u32 = 4; const input: zml.Tensor = .withPartitioning(.init(.{ .b = bs, .x = 28, .y = 28 }, .u8), dp_mp, .{ .b = .data }); + var exe = exe: { log.info("Compiling model....", .{}); const start: std.Io.Timestamp = .now(io, .awake); @@ -125,7 +135,7 @@ pub fn main(init: std.process.Init) !void { .{ .program_name = "mnist", .meshes = &.{dp_mp}, - .xla_dump_to = "/tmp/zml/mnist", + .xla_dump_to = "/tmp/zml/mnist_dp", }, ); }; diff --git a/examples/sharding/main.zig b/examples/sharding/main.zig index cc46a1bd..34a78bfe 100644 --- a/examples/sharding/main.zig +++ b/examples/sharding/main.zig @@ -30,14 +30,18 @@ const DemoModel = struct { } pub fn forward(self: DemoModel, input: zml.Tensor) zml.Tensor { - const x = input.convert(.f32); + const compiler = zml.Compiler.current(); + const x = input.convert(.f32); var y = x.dot(self.w, .feature); y = y.add(self.b.broad(y.shape())); y = y.withPartitioning(.dp_mp, .{ .batch = .data, .hidden = .model }); y.print("dense_out"); - const gate = y.scale(0.01).sigmoid().withPartitioning(.dp_mp, .{ .batch = .data, .hidden = .model }); + const gate = y.scale(0.01).sigmoid(); + + const spec: zml.Sharding.Partitioning.Logical = .parse(.{ .batch = .data, .hidden = .model }); + return zml.ops.manualComputation( (struct { y: zml.Tensor, @@ -50,6 +54,11 @@ const DemoModel = struct { }).body, .{ .y = y, .gate = gate }, y.shape(), + compiler.mesh(.dp_mp), + .{ + .in_specs = .{ .y = spec, .gate = spec }, + .out_specs = spec, + }, ); } }; @@ -60,7 +69,7 @@ fn usage() void { \\ \\Simple sharding demo: \\ - explicit partitioning on data/model axes - \\ - shard-local tensor print + \\ - tensor print \\ - manualComputation with multiple inputs and one output \\ \\Mesh modes: diff --git a/mlir/dialects/cute_ir/capi.cc b/mlir/dialects/cute_ir/capi.cc index fe1296db..6e7c296d 100644 --- a/mlir/dialects/cute_ir/capi.cc +++ b/mlir/dialects/cute_ir/capi.cc @@ -166,6 +166,10 @@ MlirStringRef mlirCuteShapeAttrGetValue(MlirAttribute attr) { return getAlgebraValue(attr); } +intptr_t mlirCuteShapeAttrGetRank(MlirAttribute attr) { + return cutegen::rank(llvm::cast(unwrap(attr)).getRef()); +} + MlirAttribute mlirCuteStrideAttrStaticGet(MlirContext context, int64_t value) { return getStaticAttr(context, value); } diff --git a/mlir/dialects/cute_ir/capi.h b/mlir/dialects/cute_ir/capi.h index e9e285d2..002f4885 100644 --- a/mlir/dialects/cute_ir/capi.h +++ b/mlir/dialects/cute_ir/capi.h @@ -54,6 +54,7 @@ MLIR_CAPI_EXPORTED MlirAttribute mlirCuteShapeAttrTupleGet( MlirContext context, intptr_t numElements, MlirAttribute const *elements); MLIR_CAPI_EXPORTED bool mlirAttributeIsACuteShape(MlirAttribute attr); MLIR_CAPI_EXPORTED MlirStringRef mlirCuteShapeAttrGetValue(MlirAttribute attr); +MLIR_CAPI_EXPORTED intptr_t mlirCuteShapeAttrGetRank(MlirAttribute attr); MLIR_CAPI_EXPORTED MlirAttribute mlirCuteStrideAttrStaticGet(MlirContext context, int64_t value); diff --git a/mlir/dialects/cute_ir/cute_ir.zig b/mlir/dialects/cute_ir/cute_ir.zig index 8a031d0b..b277d061 100644 --- a/mlir/dialects/cute_ir/cute_ir.zig +++ b/mlir/dialects/cute_ir/cute_ir.zig @@ -145,6 +145,10 @@ pub const ShapeAttr = opaque { return mlir.string(c.mlirCuteShapeAttrGetValue(self.ptr())); } + pub fn rank(self: *const ShapeAttr) usize { + return @intCast(c.mlirCuteShapeAttrGetRank(self.ptr())); + } + pub fn attribute(self: *const ShapeAttr) *const mlir.Attribute { return @ptrCast(self); } diff --git a/mlir/dialects/fly/REGENERATING.md b/mlir/dialects/fly/REGENERATING.md index b66b7abe..9673870f 100644 --- a/mlir/dialects/fly/REGENERATING.md +++ b/mlir/dialects/fly/REGENERATING.md @@ -71,7 +71,7 @@ dialect today; when one appears, extend the generator and this file. `test.zig`. - Callers that no longer compile are the point of generating: an operation whose signature changed in FlyDSL now fails at the call site. Fix them in - the same change (`dsl/fly/builder.zig`, `zml/kernel.zig`, `zml/kernels/*/fly/`, + the same change (`dsl/fly/builder.zig`, `zml/kernel/fly/fly.zig`, `zml/kernels/*/fly/`, and downstream llmd's `llmd/kernels/*/fly/`). - Emit the IR of the fly kernels before and after (for instance llmd's `sparse_mla.Main2D.emit` and the `zml/kernels/moe/fly/` kernels with diff --git a/mlir/dialects/llvm/llvm.zig b/mlir/dialects/llvm/llvm.zig index 78524e0a..d1197317 100644 --- a/mlir/dialects/llvm/llvm.zig +++ b/mlir/dialects/llvm/llvm.zig @@ -22,6 +22,7 @@ pub const elementtype = ops.elementtype; pub const AtomicBinOp = ops.AtomicBinOp; pub const AtomicOrdering = ops.AtomicOrdering; pub const inttoptr = ops.inttoptr; +pub const ptrtoint = ops.ptrtoint; pub const bitcast = ops.bitcast; pub const fmul = ops.fmul; pub const extractvalue = ops.extractvalue; diff --git a/mlir/dialects/llvm/ops.zig b/mlir/dialects/llvm/ops.zig index 2102f68a..d1fb5771 100644 --- a/mlir/dialects/llvm/ops.zig +++ b/mlir/dialects/llvm/ops.zig @@ -234,6 +234,15 @@ pub fn inttoptr(ctx: *mlir.Context, address: *const mlir.Value, result_type: *co }); } +/// llvm.ptrtoint — the integer address of an LLVM pointer. +pub fn ptrtoint(ctx: *mlir.Context, pointer: *const mlir.Value, result_type: *const mlir.Type, location: *const mlir.Location) *mlir.Operation { + return mlir.Operation.make(ctx, "llvm.ptrtoint", .{ + .operands = .{ .flat = &.{pointer} }, + .results = .{ .flat = &.{result_type} }, + .location = location, + }); +} + /// llvm.bitcast — reinterpret `value` as `result_type` (same bit width). pub fn bitcast(ctx: *mlir.Context, value: *const mlir.Value, result_type: *const mlir.Type, location: *const mlir.Location) *mlir.Operation { return mlir.Operation.make(ctx, "llvm.bitcast", .{ diff --git a/mlir/dialects/llvm/test.zig b/mlir/dialects/llvm/test.zig index 28541d44..6ba84b4d 100644 --- a/mlir/dialects/llvm/test.zig +++ b/mlir/dialects/llvm/test.zig @@ -20,6 +20,7 @@ const operation_names: []const []const u8 = &.{ "llvm.intr.lifetime.end", "llvm.intr.lifetime.start", "llvm.inttoptr", + "llvm.ptrtoint", "llvm.load", "llvm.mlir.poison", "llvm.mlir.undef", diff --git a/mlir/dialects/nvvm/nvvm.zig b/mlir/dialects/nvvm/nvvm.zig index d46f3a4f..14a0e8e2 100644 --- a/mlir/dialects/nvvm/nvvm.zig +++ b/mlir/dialects/nvvm/nvvm.zig @@ -136,7 +136,8 @@ pub fn mbarrier_wait_parity( } /// nvvm.mbarrier.try_wait.parity — wait (looping in PTX) until phase `parity` (i32) -/// of the mbarrier completes, suspending up to `suspend_time` (i32) cycles per try. +/// of the mbarrier completes, with CTA acquire semantics. `suspend_time` (i32) +/// is a suspend-time hint in nanoseconds per try. pub fn mbarrier_try_wait_parity( ctx: *mlir.Context, ptr: *const mlir.Value, @@ -147,6 +148,27 @@ pub fn mbarrier_try_wait_parity( return ops.mbarrier_try_wait_parity(ctx, ptr, parity, suspend_time, .{ .useIntrinsic = .boolean(ctx, false) }, location); } +/// One potentially suspending parity test with acquire semantics. A false +/// result means the timeout expired; callers that need completion must retry. +pub fn mbarrier_try_wait_parity_timelimit( + ctx: *mlir.Context, + ptr: *const mlir.Value, + parity: *const mlir.Value, + suspend_time: *const mlir.Value, + scope: MBarrierScope, + location: *const mlir.Location, +) *mlir.Operation { + return ops.mbarrier_try_wait_parity_timelimit( + ctx, + ptr, + parity, + suspend_time, + .int(ctx, .i1), + .{ .scope = enumAttribute(ops.MBarrierScopeKindAttr, ctx, scope), .order = enumAttribute(ops.MemOrderKindAttr, ctx, .acquire) }, + location, + ); +} + /// nvvm.fence.mbarrier.init — make mbarrier initialization visible to the cluster. pub fn fence_mbarrier_init(ctx: *mlir.Context, location: *const mlir.Location) *mlir.Operation { return ops.fence_mbarrier_init(ctx, location); @@ -208,10 +230,12 @@ pub fn tcgen05_wait(ctx: *mlir.Context, kind: Tcgen05WaitKind, location: *const return ops.tcgen05_wait(ctx, enumAttribute(ops.Tcgen05WaitKindAttr, ctx, kind), location); } -/// nvvm.tcgen05.commit — arrive on the mbarrier at `barrier_ptr` when the issued -/// tcgen05 operations complete. -pub fn tcgen05_commit(ctx: *mlir.Context, barrier_ptr: *const mlir.Value, group: CtaGroup, location: *const mlir.Location) *mlir.Operation { - return ops.tcgen05_commit(ctx, barrier_ptr, .{ .group = enumAttribute(ops.CTAGroupKindAttr, ctx, group) }, location); +/// nvvm.tcgen05.commit — arrive on the mbarrier at `barrier_ptr` when the +/// tcgen05 MMA/copy operations issued by the executing thread with `group` +/// complete. An optional mask selects CTAs in the cluster whose barriers at +/// the same shared-memory offset receive one arrival each. +pub fn tcgen05_commit(ctx: *mlir.Context, barrier_ptr: *const mlir.Value, group: CtaGroup, multicast_mask: ?*const mlir.Value, location: *const mlir.Location) *mlir.Operation { + return ops.tcgen05_commit(ctx, barrier_ptr, .{ .multicastMask = multicast_mask, .group = enumAttribute(ops.CTAGroupKindAttr, ctx, group) }, location); } /// nvvm.tcgen05.mma — D (tensor memory, `!llvm.ptr<6>`) = A (shared-memory diff --git a/platforms/cpu/cpu.bzl b/platforms/cpu/cpu.bzl index 1105f776..08817afe 100644 --- a/platforms/cpu/cpu.bzl +++ b/platforms/cpu/cpu.bzl @@ -43,21 +43,21 @@ def _cpu_plugin_impl(mctx): name = "libzml_cpu_linux_amd64", build_file_content = _BUILD_FILE_DEFAULT_VISIBILITY + _BUILD_LINUX, sha256 = "3d6a6d1b047a0943b50fd9e63e5f6dff6b7cf184b186c0cc1052de40f0dbec92", - url = "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-cpu-linux-amd64.tar.zst", + url = "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-cpu-linux-amd64.tar.zst", ) http_archive( name = "libzml_cpu_darwin_amd64", build_file_content = _BUILD_FILE_DEFAULT_VISIBILITY + _BUILD_DARWIN, sha256 = "d6b4b15b13a2d1a3e0b84f143f568cd572bef5db91784f9f7a826173ed626c7e", - url = "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-cpu-darwin-amd64.tar.zst", + url = "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-cpu-darwin-amd64.tar.zst", ) http_archive( name = "libzml_cpu_darwin_arm64", build_file_content = _BUILD_FILE_DEFAULT_VISIBILITY + _BUILD_DARWIN, sha256 = "c06babc001edeab313bc63daab3aa3d40e087266a3a29634d182628fb65fad14", - url = "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-cpu-darwin-arm64.tar.zst", + url = "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-cpu-darwin-arm64.tar.zst", ) return mctx.extension_metadata( diff --git a/platforms/cuda/cuda.bzl b/platforms/cuda/cuda.bzl index ced8ae98..8482e52a 100644 --- a/platforms/cuda/cuda.bzl +++ b/platforms/cuda/cuda.bzl @@ -224,12 +224,12 @@ _UBUNTU_PACKAGES = { _PLUGINS = { "amd64": { - "sha256": "77d2ddb19566604c827828fe3dcc53e938f3a1b16d431416fdd5c466266f5242", - "url": "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-cuda-linux-amd64.tar.zst", + "sha256": "72b1359384ac50e727b800e81c56d7134f9d6a061a982c4e36b7fef62cf7c331", + "url": "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-cuda-linux-amd64.tar.zst", }, "arm64": { - "sha256": "2cd18393c44816d3d01aa22fbd890632ba46eca3939c644bddef00d590578a80", - "url": "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-cuda-linux-arm64.tar.zst", + "sha256": "caeefd717960fa2acec6304cd8296299e262abdba669ecf4d487e3af76d73a7b", + "url": "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-cuda-linux-arm64.tar.zst", }, } diff --git a/platforms/furiosa/furiosa.bzl b/platforms/furiosa/furiosa.bzl index 0aa3905d..afb26864 100644 --- a/platforms/furiosa/furiosa.bzl +++ b/platforms/furiosa/furiosa.bzl @@ -6,8 +6,8 @@ def _furiosa_impl(mctx): http_archive( name = "libzml_furiosa", build_file = "libzml_furiosa.BUILD.bazel", - url = "https://mirror.zml.ai/plugins/202610221403.283.1.d9f4f68759e1/zml-furiosa-linux-amd64.tar.zst", - sha256 = "f8d6bb03fb4cd5dad9b37856b2f8f12b28ec1a8d46f59ca03c401bea668215e9", + url = "https://mirror.zml.ai/plugins/202611131503.373.1.be246d178d7d/zml-furiosa-linux-amd64.tar.zst", + sha256 = "c7cdf0120818589aee3b53d8a6813e2b5e244447701acc34eb697b30f46e22f9", ) return mctx.extension_metadata( diff --git a/platforms/furiosa/furiosa.zig b/platforms/furiosa/furiosa.zig index ebbd9e3c..ee33fd58 100644 --- a/platforms/furiosa/furiosa.zig +++ b/platforms/furiosa/furiosa.zig @@ -50,7 +50,6 @@ pub fn load(allocator: std.mem.Allocator, io: std.Io) !*const pjrt.Api { var lib_path_buf: [std.Io.Dir.max_path_bytes]u8 = undefined; try setEnv("FURIOSA_SDK_ROOT", try stdx.Io.Dir.path.bufJoinZ(&lib_path_buf, &.{root}), 1); - try setEnv("FURIOSA_VISIBLE_DEVICES", "0", 0); const library = try stdx.Io.Dir.path.bufJoinZ(&lib_path_buf, &.{ root, "lib", "libzml_furiosa.so" }); log.info("Loading Furiosa plugin: {s}", .{library}); diff --git a/platforms/metal/metal.bzl b/platforms/metal/metal.bzl index 03cdaf02..71e1858a 100644 --- a/platforms/metal/metal.bzl +++ b/platforms/metal/metal.bzl @@ -4,8 +4,8 @@ def _metal_impl(mctx): http_archive( name = "libzml_metal", build_file = "libzml_metal.BUILD.bazel", - sha256 = "6eee91776397d472ac590413cf44d4ff3f7b31bebb163aa96339c6becaf2defd", - url = "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-metal-darwin-arm64.tar.zst", + sha256 = "89881d1a1db336e3337a9373db2c4e94f71ddfb77619dee75c0c12e2903913ae", + url = "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-metal-darwin-arm64.tar.zst", ) return mctx.extension_metadata( diff --git a/platforms/oneapi/oneapi.bzl b/platforms/oneapi/oneapi.bzl index 6016e389..a3c968d3 100644 --- a/platforms/oneapi/oneapi.bzl +++ b/platforms/oneapi/oneapi.bzl @@ -1,8 +1,8 @@ load("@llvm//:http_bsdtar_archive.bzl", http_archive = "http_bsdtar_archive") load("//bazel:http_deb_archive.bzl", "http_deb_archive") -_PLUGIN_ONEAPI_URL = "https://mirror.zml.ai/plugins/202610011939.137.1.0d59bb389ee2/zml-oneapi-linux-amd64.tar.zst" -_PLUGIN_ONEAPI_SHA256 = "9ca5e2c377d938b3d855c46c685d193e90424e44b1c1ea8d0a8e1c924b0d11f1" +_PLUGIN_ONEAPI_URL = "https://mirror.zml.ai/plugins/202611132003.375.1.7d765f57a407/zml-oneapi-linux-amd64.tar.zst" +_PLUGIN_ONEAPI_SHA256 = "294722670832c543c8ac4c074c30f636f2a3b0854820f7c5a070fa82c0d6f3d5" ONEAPI_VERSION = "2026.1" @@ -248,6 +248,7 @@ filegroup( name = "mkl_core_runtime", srcs = [ "{ONEAPI_MKL_LIB}/libmkl_core.so.3", + "{ONEAPI_MKL_LIB}/libmkl_def.so.3", "{ONEAPI_MKL_LIB}/libmkl_intel_ilp64.so.3", "{ONEAPI_MKL_LIB}/libmkl_sequential.so.3", ], diff --git a/platforms/oneapi/packages.lock.json b/platforms/oneapi/packages.lock.json index a3e69ad4..140ce102 100644 --- a/platforms/oneapi/packages.lock.json +++ b/platforms/oneapi/packages.lock.json @@ -216,7 +216,7 @@ "name": "libigc2", "sha256": "c3e5f181eecd40dc8b656c3ff5b7e3d4f690e93181255726516416e61ade1d1f", "urls": [ - "https://ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/pool/main/i/intel-graphics-compiler/libigc2_2.41.15-1~24.04_amd64.deb" + "https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20261029T000000Z/pool/main/i/intel-graphics-compiler/libigc2_2.41.15-1~24.04_amd64.deb" ], "version": "2.41.15-1~24.04" }, @@ -227,7 +227,7 @@ "name": "libigdfcl2", "sha256": "8a2058fd9f5e7d0a4b05001ff861780d2f72529135fa70fb7fbba0f0ff8cd74f", "urls": [ - "https://ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/pool/main/i/intel-graphics-compiler/libigdfcl2_2.41.15-1~24.04_amd64.deb" + "https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20261029T000000Z/pool/main/i/intel-graphics-compiler/libigdfcl2_2.41.15-1~24.04_amd64.deb" ], "version": "2.41.15-1~24.04" }, @@ -249,7 +249,7 @@ "name": "libze1", "sha256": "d4001dc474d3cb74ded50a449b9f646cfdfbb31e69e70b8af38f2d18273dfce1", "urls": [ - "https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260930T000000Z/pool/main/l/level-zero-loader/libze1_1.32.0-1~24.04~ppa1_amd64.deb" + "https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20261029T000000Z/pool/main/l/level-zero-loader/libze1_1.32.0-1~24.04~ppa1_amd64.deb" ], "version": "1.32.0-1~24.04~ppa1" }, diff --git a/platforms/oneapi/packages.yaml b/platforms/oneapi/packages.yaml index 272bc02c..5a829ac5 100644 --- a/platforms/oneapi/packages.yaml +++ b/platforms/oneapi/packages.yaml @@ -15,9 +15,11 @@ sources: - channel: noble main url: https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260930T000000Z - # IGC 2.41.15 is newer than the graphics snapshot above. + # IGC 2.41.15 is newer than the graphics snapshot above; the other graphics + # packages are pinned below to keep their versions from it. Use snapshots, + # not the live PPA: the PPA deletes superseded packages from its pool. - channel: noble main - url: https://ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu + url: https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20261029T000000Z archs: - "amd64" @@ -43,9 +45,9 @@ packages: - "libnl-genl-3-200" - "libnl-route-3-200" - "libnuma1" - - "libigdgmm12" + - "libigdgmm12 (= 22.10.1-1~24.04~ppa1)" - "libigc2 (= 2.41.15-1~24.04)" - "libigdfcl2 (= 2.41.15-1~24.04)" - - "libze-intel-gpu1" + - "libze-intel-gpu1 (= 26.31.39395.14-1~24.04~ppa1)" - "libze1" - "zlib1g" diff --git a/platforms/rocm/rocm.bzl b/platforms/rocm/rocm.bzl index babd73ab..41b01b60 100644 --- a/platforms/rocm/rocm.bzl +++ b/platforms/rocm/rocm.bzl @@ -11,8 +11,8 @@ load("@rules_cc//cc:cc_library.bzl", "cc_library") _ROCM_VERSION = "10.1" _ROCM_STRIP_PREFIX = "./opt/rocm/core-" + _ROCM_VERSION -_PLUGIN_ROCM_URL = "https://mirror.zml.ai/plugins/202610290951.309.1.785550677e75/zml-rocm-linux-amd64.tar.zst" -_PLUGIN_ROCM_SHA256 = "cd6c3c8673ce690a0fbf89dab5a16e6893f0f501bb2b41997c657cf6effa7549" +_PLUGIN_ROCM_URL = "https://mirror.zml.ai/plugins/202611101651.363.1.d46c7471c419/zml-rocm-linux-amd64.tar.zst" +_PLUGIN_ROCM_SHA256 = "cdc4434f2023842f0e1a34c08aaed87f082df19f92c4ae21c9d4452a9739b604" def _rocm_package_name(name): return name + _ROCM_VERSION @@ -73,6 +73,7 @@ def _rocm_runfile_build_files(loaded_packages): for pkg_name in _family_runfile_package_names(loaded_packages, "dnn"): rocm_packages[pkg_name] = _glob_filegroup("runfiles", [ + ".kpack/**", "lib/libMIOpenCKGroupedConv_*.so", "share/miopen/db/**", ]) diff --git a/platforms/tpu/ragged_paged.zig b/platforms/tpu/ragged_paged.zig index 4c1c50ed..6a0fdfae 100644 --- a/platforms/tpu/ragged_paged.zig +++ b/platforms/tpu/ragged_paged.zig @@ -482,7 +482,7 @@ fn flashAttention( } /// Emit the kernel body into a pre-opened `mtt.Builder`. The factory in -/// `zml/kernel.zig` opens the builder, calls this `run`, and finalizes the +/// `zml/kernel/mosaic_tpu/mosaic_tpu.zig` opens the builder, calls this `run`, and finalizes the /// IR with `b.finishOpts(&.{}, .{ .canonicalize = true })`. pub fn run(b: *mtt.Builder, cfg: Cfg) mtt.FinishError!void { const q_transform: []const mtt.ArgSpec.TransformReturn = &.{ diff --git a/stdx/flags.zig b/stdx/flags.zig index 52a9e6a4..d0c4352d 100644 --- a/stdx/flags.zig +++ b/stdx/flags.zig @@ -591,25 +591,19 @@ fn fields_to_comma_list(comptime E: type) []const u8 { } } +/// Maps a field name to its CLI flag, e.g. `model_path` to `--model-path`. pub fn flag_name(comptime field_name: []const u8) []const u8 { - // TODO(Zig): Cleanup when this is fixed after Zig 0.11. - // Without comptime blk, the compiler thinks the result is a runtime slice returning a UAF. - return comptime blk: { - assert(!std.mem.eql(u8, field_name, "positional")); - - var result: []const u8 = "--"; - var index = 0; - while (std.mem.indexOf(u8, field_name[index..], "_")) |i| { - result = result ++ field_name[index..][0..i] ++ "-"; - index = index + i + 1; - } - result = result ++ field_name[index..]; - break :blk result; - }; + comptime assert(!std.mem.eql(u8, field_name, "positional")); + comptime var name = field_name[0..field_name.len].*; + comptime std.mem.replaceScalar(u8, &name, '_', '-'); + const dashed = name; + return "--" ++ dashed; } test flag_name { - try std.testing.expectEqualStrings(flag_name("statsd"), "--statsd"); + try std.testing.expectEqualStrings("--statsd", flag_name("statsd")); + try std.testing.expectEqualStrings("--model-path", flag_name("model_path")); + try std.testing.expectEqualStrings("--a-b-c", flag_name("a_b_c")); } fn flag_name_positional(comptime field_name: []const u8) []const u8 { diff --git a/third_party/zls/BUILD.bazel b/third_party/zls/BUILD.bazel index d53e20be..69cc6128 100644 --- a/third_party/zls/BUILD.bazel +++ b/third_party/zls/BUILD.bazel @@ -30,7 +30,7 @@ zls_toolchain( toolchain( name = "toolchain", - target_settings = ["@zig_toolchains//:0.17.0"], + target_settings = ["@zig_toolchains//:any_0.17.release"], toolchain = ":zls", toolchain_type = "@rules_zig//zig/zls:toolchain_type", ) diff --git a/tools/safetensors/BUILD.bazel b/tools/safetensors/BUILD.bazel new file mode 100644 index 00000000..aff4954d --- /dev/null +++ b/tools/safetensors/BUILD.bazel @@ -0,0 +1,13 @@ +load("@rules_zig//zig:defs.bzl", "zig_binary", "zig_test") + +zig_binary( + name = "safetensors", + main = "main.zig", + visibility = ["//visibility:public"], + zigopts = ["-lc"], + deps = [ + "//stdx", + "//vfs", + "//zml", + ], +) diff --git a/tools/safetensors/main.zig b/tools/safetensors/main.zig new file mode 100644 index 00000000..cf902eef --- /dev/null +++ b/tools/safetensors/main.zig @@ -0,0 +1,91 @@ +const std = @import("std"); + +const stdx = @import("stdx"); +const VFS = @import("vfs"); +const zml = @import("zml"); + +const log = std.log.scoped(.safetensors_convert); + +const Args = struct { + output: ?[]const u8 = null, + positional: struct { model: []const u8 }, + + pub const help = + \\Usage: [--output=] + \\ + \\Read safetensors headers and write model.safetensors.full_index.json in the working directory + \\Example: bazel run //tools/safetensors -- --output="model.safetensors.full_index.json" hf://deepseek-ai/DeepSeek-V4.1-Flash + ; +}; + +pub fn main(init: std.process.Init) !void { + const args = stdx.flags.parse(init.minimal.args, Args); + const allocator = init.gpa; + + const start = std.Io.Clock.now(.awake, init.io); + var client: std.http.Client = .{ .allocator = allocator, .io = init.io }; + defer client.deinit(); + try client.initDefaultProxies(allocator, init.environ_map); + + const input = args.positional.model; + + var file_vfs: VFS.File = .init(allocator, init.io, .{ .direct_io = false }); + defer file_vfs.deinit(); + var http_vfs: VFS.HTTP = try .init(allocator, init.io, &client, .http); + defer http_vfs.deinit(); + var https_vfs: VFS.HTTP = try .init(allocator, init.io, &client, .https); + defer https_vfs.deinit(); + var hf_vfs: ?VFS.HF = if (std.mem.startsWith(u8, input, "hf://")) + try .auto(allocator, init.io, &client, init.environ_map) + else + null; + defer if (hf_vfs) |*backend| backend.deinit(); + var s3_vfs: ?VFS.S3 = if (std.mem.startsWith(u8, input, "s3://")) + try .auto(allocator, init.io, &client, init.environ_map) + else + null; + defer if (s3_vfs) |*backend| backend.deinit(); + var gcs_vfs: ?VFS.GCS = if (std.mem.startsWith(u8, input, "gs://")) + try .auto(allocator, init.io, &client, init.environ_map) + else + null; + defer if (gcs_vfs) |*backend| backend.deinit(); + + var vfs: VFS = try .init(allocator, init.io); + const io = vfs.io(); + defer vfs.deinit(); + try vfs.register("file", file_vfs.io()); + try vfs.register("http", http_vfs.io()); + try vfs.register("https", https_vfs.io()); + if (hf_vfs) |*backend| try vfs.register("hf", backend.io()); + if (s3_vfs) |*backend| try vfs.register("s3", backend.io()); + if (gcs_vfs) |*backend| try vfs.register("gs", backend.io()); + + var arena: std.heap.ArenaAllocator = .init(allocator); + defer arena.deinit(); + + const repo = try zml.safetensors.resolveModelRepo(io, input); + var registry: zml.safetensors.TensorRegistry = try .fromRepo(allocator, io, repo); + defer registry.deinit(); + const loaded = std.Io.Clock.now(.awake, init.io); + + const index_json_file = if (args.output) |output| + try std.Io.Dir.cwd().createFile(io, output, .{}) + else + try repo.createFile(io, full_index_path, .{}); + const out_buffer = try allocator.alloc(u8, 4096); + defer allocator.free(out_buffer); + var out_writer = index_json_file.writer(io, out_buffer); + try registry.saveAsFullIndexJson(io, allocator, repo, &out_writer.interface); + const done = std.Io.Clock.now(.awake, init.io); + + log.info("Wrote {s}: {d} tensors; load {f}s, write {f}s, total {f}s", .{ + args.output orelse full_index_path, + registry.tensors.count(), + start.durationTo(loaded), + loaded.durationTo(done), + start.durationTo(done), + }); +} + +const full_index_path = "model.safetensors.full_index.json"; diff --git a/zml/BUILD.bazel b/zml/BUILD.bazel index 8188c743..f52ecdd5 100644 --- a/zml/BUILD.bazel +++ b/zml/BUILD.bazel @@ -78,6 +78,7 @@ zig_library( srcs = [ "Compiler.zig", "Sharding.zig", + "attention/Options.zig", "attention/attention.zig", "buffer.zig", "buffer/HostAccessible.zig", @@ -86,10 +87,13 @@ zig_library( "exe.zig", "floats.zig", "io.zig", - "kernel.zig", "kernels/flash_attention/flash_attention.zig", + "kernels/flash_attention/tests.zig", + "kernels/moe/api.zig", "kernels/moe/moe.zig", "kernels/moe/tests.zig", + "kernels/moe/weights.zig", + "kernels/paged_attention/options.zig", "kernels/paged_attention/paged_attention.zig", "logo.zig", "mem.zig", @@ -109,11 +113,12 @@ zig_library( "tensor.zig", "testing.zig", "zml.zig", + "//zml/kernel", "//zml/kernels/flash_attention/cute", "//zml/kernels/flash_attention/metal", "//zml/kernels/flash_attention/nki", "//zml/kernels/moe/cute", - "//zml/kernels/moe/cutlass_flashinfer", + "//zml/kernels/moe/flashinfer_cutlass", "//zml/kernels/moe/fly", "//zml/kernels/moe/metal", "//zml/kernels/moe/mosaic_tpu", diff --git a/zml/Compiler.zig b/zml/Compiler.zig index 3daba2ec..93eafa8d 100644 --- a/zml/Compiler.zig +++ b/zml/Compiler.zig @@ -23,6 +23,7 @@ const tracer = @import("profiling/tracer.zig"); const Shape = @import("shape.zig").Shape; const Sharding = @import("Sharding.zig"); const Tensor = @import("tensor.zig").Tensor; +const TensorStore = @import("io.zig").TensorStore; const Compiler = @This(); const log = std.log.scoped(.@"zml/Compiler"); @@ -31,6 +32,7 @@ allocator: std.mem.Allocator, io: std.Io, arena: std.heap.ArenaAllocator, +options: Options, mlir_registry: *mlir.DialectRegistry, mlir_ctx: *mlir.Context, mlir_pass_manager: *mlir.PassManager, @@ -77,6 +79,7 @@ pub const Options = struct { /// Incremented once after successful compilation. progress: std.Progress.Node = .none, meshes: []const *const Sharding.Mesh = &.{}, + store: ?*const TensorStore = null, // If null, will be initialized from the target partitioner: ?Sharding.Partitioner = null, /// HACK: Extra bytes added to the device memory size reported to XLA. @@ -100,7 +103,7 @@ pub const Error = std.mem.Allocator.Error || mlir.Error || upb.SerializeError || pjrtx.Client.CompileError || - error{MissingDeviceInTile}; + error{ MissingDeviceInTile, OpenShardingAtBoundaries }; pub const Donation = union(enum) { implicit: u32, @@ -194,6 +197,7 @@ pub fn init(allocator: std.mem.Allocator, io: std.Io, platform: *const Platform, .mlir_known_types = mlir_known_types, .module = module, .platform = platform, + .options = opts, .partitioner = opts.partitioner orelse .fromTarget(platform.target), .meshes = meshes, .location = unknown_location, @@ -288,6 +292,7 @@ test pushLocation { .mlir_pass_manager = undefined, .mlir_known_types = undefined, .module = undefined, + .options = .{}, .platform = undefined, .partitioner = undefined, .meshes = undefined, @@ -361,16 +366,13 @@ pub fn mesh(compiler: *const Compiler, name: @EnumLiteral()) *const Sharding.Mes /// Find a mesh that covers all the given axes pub fn resolveMesh(compiler: *const Compiler, logical_axes: anytype) *const Sharding.Mesh { if (@TypeOf(logical_axes) != []const Shape.Tag) { - const comp_tags = comptime Shape.parseTags(logical_axes); - const parsed_tags: []const Shape.Tag = comptime comp_tags.constSlice(); - return compiler.resolveMesh(parsed_tags); + return compiler.resolveMesh(comptime Shape.parseTags(logical_axes)); } var ok_meshe: ?*const Sharding.Mesh = null; for (compiler.meshes) |shd| { var covers_all: bool = true; - for (logical_axes) |ax| { - const input_axis = Shape.toTag(ax); + for (logical_axes) |input_axis| { var covers_this: bool = false; for (shd.logical.axes.slice()) |existing_axis| { if (std.mem.eql(u8, std.mem.span(existing_axis), std.mem.span(input_axis))) { @@ -462,10 +464,7 @@ pub fn compileInternal( var compiler: Compiler = .init(allocator, io, platform, opts); defer compiler.deinit(); - var result = compiler.emitMlir(func_, args) catch |err| switch (err) { - error.OutOfMemory => return error.OutOfMemory, - else => unreachable, - }; + var result = try compiler.emitMlir(func_, args); defer result.output_info.deinit(compiler.allocator); defer result.input_info.deinit(compiler.allocator); @@ -593,20 +592,21 @@ fn emitMlir(compiler: *Compiler, comptime func_: anytype, args: std.meta.ArgsTup var input_info = try createBlockArguments(compiler, fn_scope, &args); errdefer input_info.deinit(compiler.allocator); - var result = blk: { + var output_info = blk: { compiler.activate(); defer compiler.deactivate(); - break :blk @call(.auto, func_, args); - }; - var output_info = try collectOutputInfo(compiler, fn_scope, &result); + var result = @call(.auto, func_, args); + // Output collection can emit replication constraints, so it is part of tracing. + break :blk try collectOutputInfo(compiler, fn_scope, &result); + }; errdefer output_info.deinit(compiler.allocator); // Finalize in a separate function that doesn't depend on `func` type. return try finalizeMlirFunc(compiler, fn_scope, input_info, output_info); } -fn createBlockArguments(compiler: *Compiler, scope: *Scope, v: anytype) error{OutOfMemory}!std.MultiArrayList(TensorInfo) { +fn createBlockArguments(compiler: *Compiler, scope: *Scope, v: anytype) error{ OutOfMemory, OpenShardingAtBoundaries }!std.MultiArrayList(TensorInfo) { const CreateBlockArgumentsCb = struct { compiler: *Compiler, scope: *Scope, @@ -614,12 +614,23 @@ fn createBlockArguments(compiler: *Compiler, scope: *Scope, v: anytype) error{Ou current_argument_id: u32 = 0, infos: std.MultiArrayList(TensorInfo) = .empty, - fn cb(ctx: *@This(), tensor: *const Tensor) error{OutOfMemory}!void { + fn cb(ctx: *@This(), tensor: *const Tensor) error{ OutOfMemory, OpenShardingAtBoundaries }!void { // Declare the program argument as a packed u8 const og_shape = tensor._shape; + if (og_shape.containsPartitionSpec(.open)) { + log.err("Executable input {d} has open sharding: {f}. Specify a logical mesh axis or .replicated for every dimension.", .{ ctx.current_argument_id, og_shape }); + return error.OpenShardingAtBoundaries; + } const packed_shape = og_shape.packedShape(); - const mlir_type = mlirx.Type.rankedTensor(ctx.compiler.mlir_ctx, packed_shape); - var value = ctx.scope.block.addArgument(mlir_type, .unknown(ctx.compiler.mlir_ctx)); + const mlir_ctx = ctx.compiler.mlir_ctx; + const mlir_type = mlirx.Type.rankedTensor(mlir_ctx, packed_shape); + var location: *const mlir.Location = ctx.compiler.unknown_location; + if (ctx.compiler.options.store) |store| { + if (store.id_to_sources.get(tensor.id)) |sources| { + location = location.named(mlir_ctx, sources.tensors[0].name); + } + } + var value = ctx.scope.block.addArgument(mlir_type, location); // But immediately create the MLIR value corresponding to the fp4 data. // That way the user code only sees a proper fp4 tensor. @@ -650,7 +661,7 @@ fn createBlockArguments(compiler: *Compiler, scope: *Scope, v: anytype) error{Ou return context.infos; } -fn collectOutputInfo(compiler: *Compiler, scope: *Scope, v: anytype) error{OutOfMemory}!std.MultiArrayList(TensorInfo) { +fn collectOutputInfo(compiler: *Compiler, scope: *Scope, v: anytype) error{ OutOfMemory, OpenShardingAtBoundaries }!std.MultiArrayList(TensorInfo) { const CollectOutputInfoCb = struct { compiler: *Compiler, scope: *Scope, @@ -658,9 +669,16 @@ fn collectOutputInfo(compiler: *Compiler, scope: *Scope, v: anytype) error{OutOf infos: std.MultiArrayList(TensorInfo) = .empty, fn cb(ctx: *@This(), tensor: *const Tensor) !void { - const og_shape = tensor.shape(); - var value = ctx.scope.id_to_argument.get(tensor.id) orelse - tensor._value orelse + // Outputs with no declared sharding default to replication. Constrain the + // returned value so the partitioner produces a replicated output buffer. + const output = if (tensor.shape()._sharding.mesh != null or tensor.shape().isFullyReplicated()) tensor.* else tensor.replicate(); + const og_shape = output.shape(); + if (og_shape.containsPartitionSpec(.open)) { + log.err("Executable output {d} has open sharding: {f}. Specify a logical mesh axis or .replicated for every dimension.", .{ ctx.infos.len, og_shape }); + return error.OpenShardingAtBoundaries; + } + var value = ctx.scope.id_to_argument.get(output.id) orelse + output._value orelse @panic("no value found for output tensor"); if (tensor.dtype().bitSizeOf() < 8) { diff --git a/zml/Sharding.zig b/zml/Sharding.zig index 2637f209..040216bc 100644 --- a/zml/Sharding.zig +++ b/zml/Sharding.zig @@ -21,10 +21,22 @@ mesh: ?*const Mesh, partition: Partitioning, pub const MAX_MESH_RANK = 4; + +/// Full replication across participating devices. +/// For buffers, each device holds a complete copy. Applied as a graph constraint, +/// this closes every dimension to partitioning. `Shape.init` uses this sharding. +/// Unconstrained graph results use `open` instead. pub const replicated: Sharding = .{ .mesh = null, .partition = .replicated(Shape.MAX_RANK) }; +/// Unspecified sharding for graph results. During compilation, the partitioner +/// may infer sharding from surrounding constraints or leave the value replicated. +/// Executable inputs require closed sharding. An executable output with this +/// unspecified sharding defaults to replication. An explicitly declared output +/// sharding containing open dimensions is rejected. +pub const open: Sharding = .{ .mesh = null, .partition = .open(Shape.MAX_RANK) }; + pub fn get(sharding: Sharding, ax: usize) PartitionSpec { - return if (sharding.mesh == null) .replicated else sharding.partition.get(ax); + return sharding.partition.get(ax); } pub fn set(self: Sharding, ax: usize, spec: PartitionSpec) Sharding { @@ -52,7 +64,6 @@ pub fn toArray(self: Sharding) [Shape.MAX_RANK]PartitionSpec { } pub fn isFullyReplicated(self: Sharding) bool { - if (self.mesh == null) return true; for (self.partition.toArray()) |spec| { if (spec != .replicated and spec != .out_of_bound) return false; } @@ -98,8 +109,18 @@ test isCompatible { } } +/// Formats the declared sharding, not the partitioner's inferred sharding. +/// Without a mesh, prints `{replicated}` for full replication or `{open}` for +/// unspecified graph sharding. +/// +/// With a mesh, entries follow tensor dimension order, e.g. `{data,replicated,?}@dp`: +/// - `data`: split this dimension over the logical mesh axis named `data`. +/// - `replicated`: keep this dimension whole on each shard. +/// - `?`: leave this dimension to partitioner inference during compilation. +/// - `@dp`: the mesh name. +/// Trailing `.out_of_bound` entries are omitted. Any remaining ones print as `!`. pub fn format(self: Sharding, writer: *std.Io.Writer) !void { - const mesh = self.mesh orelse return writer.writeAll("{replicated}"); + const mesh = self.mesh orelse return writer.writeAll(if (self.isFullyReplicated()) "{replicated}" else "{open}"); const specs = self.partition.toArray(); var rank = specs.len; while (rank > 0 and specs[rank - 1] == .out_of_bound) rank -= 1; @@ -117,8 +138,10 @@ pub fn format(self: Sharding, writer: *std.Io.Writer) !void { try writer.print("}}@{s}", .{mesh.name}); } -/// Describes how a given Shape axis behaves inside a mesh. -/// Is it replicated ? sharded along a specific logical axis ? or open to replication ? +/// Describes how one tensor dimension is distributed across a logical mesh. +/// A mesh axis splits the dimension. `.replicated` keeps it whole on each shard. +/// In a graph constraint, `.open` lets the partitioner choose either behavior. +/// Executable buffers require closed sharding specs: a mesh axis or `.replicated`. pub const PartitionSpec = enum(u4) { mesh_axis_0 = 0, mesh_axis_1 = 1, @@ -414,18 +437,22 @@ pub const Partitioning = packed struct { const logical = if (@TypeOf(partitioning) == Logical) partitioning else - // capture all the comptime logic in the Logical struct. + // capture all the comptime logic in Logical. comptime Logical.parse(partitioning); return logical.applyUseAll(tags, mesh, default_value); } - pub const Logical = struct { - src: Shape.TagsArray, - tgt: Shape.TagsArray, + pub const Logical = union(enum) { + /// Replicates every dimension, including when used as a graph constraint. + replicated, + mapping: struct { + src: Shape.TagsArray, + tgt: Shape.TagsArray, + }, - const replicated: Shape.Tag = Shape.toTag(.replicated); - const open: Shape.Tag = Shape.toTag(.open); + const replicated_tag: Shape.Tag = Shape.toTag(.replicated); + const open_tag: Shape.Tag = Shape.toTag(.open); pub fn parse(v: anytype) Logical { const V = @TypeOf(v); @@ -434,7 +461,7 @@ pub const Partitioning = packed struct { var tgt_: Shape.TagsArray = .empty; if (comptime stdx.meta.isStruct(V)) { - const field_names = std.meta.fieldNames(V); + const field_names = comptime std.meta.fieldNames(V); stdx.debug.assertComptime(field_names.len <= @TypeOf(src_).max_len, "Too many tags in struct {} ({d})", .{ V, field_names.len }); inline for (field_names) |field_name| { const fv = @field(v, field_name); @@ -444,7 +471,7 @@ pub const Partitioning = packed struct { src_.appendAssumeCapacity(field_name); } } - return .{ .src = src_, .tgt = tgt_ }; + return .{ .mapping = .{ .src = src_, .tgt = tgt_ } }; } stdx.debug.compileError("Sharding.Partitioning.parse expects a struct of enum literals, eg {{ .b = .data, .d = .model }}, got: {any}", .{V}); @@ -453,23 +480,39 @@ pub const Partitioning = packed struct { test "Logical.parse" { const logical = comptime Logical.parse(.{ .a = .x, .b = .y }); - try std.testing.expectEqualSlices(Shape.Tag, &.{ "a".ptr, "b".ptr }, logical.src.constSlice()); - try std.testing.expectEqualSlices(Shape.Tag, &.{ "x".ptr, "y".ptr }, logical.tgt.constSlice()); + try std.testing.expectEqualSlices(Shape.Tag, &.{ "a".ptr, "b".ptr }, logical.mapping.src.constSlice()); + try std.testing.expectEqualSlices(Shape.Tag, &.{ "x".ptr, "y".ptr }, logical.mapping.tgt.constSlice()); + } + + test "Logical.replicated" { + const mesh: Mesh = .initTest(.logical, .{.model}); + const tags: []const Shape.Tag = &.{ "a", "b" }; + const logical: Logical = .replicated; + for ([_]PartitionSpec{ .open, .replicated }) |default| { + try std.testing.expectEqual(Partitioning.replicated(tags.len), logical.applyUseAll(tags, &mesh, default)); + } + const unspecified = Logical.parse(.{}); + try std.testing.expectEqual(Partitioning.open(tags.len), unspecified.applyUseAll(tags, &mesh, .open)); } pub fn apply(logical: *const Logical, tags: []const [*:0]const u8, mesh: *const Mesh, default_value: PartitionSpec) struct { Partitioning, ?u8 } { + const mapping = switch (logical.*) { + .replicated => return .{ .replicated(tags.len), null }, + .mapping => |*mapping_| mapping_, + }; + var partition_: Partitioning = .repeat(default_value, tags.len); var missing_src_axis: ?u8 = null; - for (0.., logical.src.slice(), logical.tgt.slice()) |i, shape_tag, mesh_tag| { + for (0.., mapping.src.slice(), mapping.tgt.slice()) |i, shape_tag, mesh_tag| { const shape_ax = Shape.axisFromTagMaybe(tags, shape_tag) orelse { missing_src_axis = @intCast(i); continue; }; - const spec: PartitionSpec = if (mesh_tag == Logical.replicated) + const spec: PartitionSpec = if (mesh_tag == Logical.replicated_tag) .replicated - else if (mesh_tag == Logical.open) + else if (mesh_tag == Logical.open_tag) .open else spec: { const mesh_ax = mesh.resolveLogicalAxis(mesh_tag) orelse { @@ -486,7 +529,7 @@ pub const Partitioning = packed struct { const p, const missing_shape_axis = logical.apply(tags, mesh, default_value); if (missing_shape_axis) |ax| { - std.debug.panic("{f} doesn't have an axis {s} to be partitioned on.", .{ stdx.fmt.stringsZ(tags), logical.src.get(ax) }); + std.debug.panic("{f} doesn't have an axis {s} to be partitioned on.", .{ stdx.fmt.stringsZ(tags), logical.mapping.src.get(ax) }); } return p; } @@ -1435,8 +1478,8 @@ pub const LogicalMesh = struct { /// * dimensions: dimension for each axis, -1 infers the size from underlying topology pub fn init(names: anytype, intents: []const LogicalAxisIntent, dimensions: []const i32) !LogicalMesh { if (@TypeOf(names) != []const Shape.Tag) { - const parsed_names = comptime Shape.parseTags(names); - return .init(@as([]const Shape.Tag, parsed_names.slice()), intents, dimensions); + const parsed_names: []const Shape.Tag = comptime Shape.parseTags(names); + return .init(parsed_names, intents, dimensions); } if (names.len == 0 or names.len > MAX_MESH_RANK or names.len != intents.len or names.len != dimensions.len) @panic("invalid logical mesh"); @@ -1472,6 +1515,10 @@ pub const LogicalMesh = struct { return .{ .axes = axes, .intents = intents, .dimensions = @splat(-1) }; } + pub fn hasAxis(m: *const LogicalMesh, logical_ax: Shape.Tag) ?u3 { + return @truncate(std.mem.findScalar(Shape.Tag, m.axes.slice(), logical_ax) orelse return null); + } + pub fn intent(self: LogicalMesh, tag: Shape.Tag) ?LogicalAxisIntent { const target = std.mem.span(tag); for (self.axes.slice(), self.intents.slice()) |t, i| { @@ -1598,7 +1645,7 @@ pub const Mesh = struct { return .{ .name = @tagName(name_), .logical = .{ - .axes = .init(axes_tags.slice()), + .axes = .init(axes_tags), .intents = .{ .buffer = @splat(.high_bandwidth), .len = axes_tags.len }, .dimensions = @splat(-1), }, @@ -1637,13 +1684,17 @@ pub const Mesh = struct { return @divExact(self.numDevices(), self.numPartitions()); } + pub fn axis(mesh: *const Mesh, ax: @EnumLiteral()) u3 { + return mesh.logical.hasAxis(Shape.toTag(ax)) orelse std.debug.panic("Mesh(name={s}) doesn't have axis {}: {f}", .{ mesh.name, ax, mesh.logical }); + } + pub fn sdyMeshAttr(self: *const Mesh, ctx: *mlir.Context) *const shardy.MeshAttr { var axes: stdx.BoundedArray(*const shardy.MeshAxisAttr, MAX_MESH_RANK + 1) = .empty; var tile_count: usize = 1; - for (self.logical.axes.slice(), 0..) |axis, i| { + for (self.logical.axes.slice(), 0..) |ax, i| { const size = self.grid.axis_sizes[i]; tile_count *= size; - axes.appendAssumeCapacity(.get(ctx, .{ .name = std.mem.span(axis), .size = @intCast(size) })); + axes.appendAssumeCapacity(.get(ctx, .{ .name = std.mem.span(ax), .size = @intCast(size) })); } const replicas = @divExact(self.grid.device_count, tile_count); if (replicas > 1) { @@ -1654,8 +1705,8 @@ pub const Mesh = struct { for (0..self.grid.device_count) |partition_index| { const coords: [MAX_MESH_RANK]u8 = self.grid.logical_coords[partition_index]; var tile_index: usize = 0; - for (0..self.logical.axes.len) |axis| { - tile_index = tile_index * self.grid.axis_sizes[axis] + coords[axis]; + for (0..self.logical.axes.len) |ax| { + tile_index = tile_index * self.grid.axis_sizes[ax] + coords[ax]; } // Device IDs here are executable partition indices, not PJRT device IDs. ids[tile_index * replicas + replica_indices[tile_index]] = @intCast(partition_index); @@ -1668,14 +1719,58 @@ pub const Mesh = struct { // There are at most MAX_MESH_RANK logical names, so one candidate is free. inline for (0..MAX_MESH_RANK + 1) |i| { const name = std.fmt.comptimePrint("__replica_{d}", .{i}); - const taken = for (self.logical.axes.slice()) |axis| { - if (std.mem.eql(u8, std.mem.span(axis), name)) break true; + const taken = for (self.logical.axes.slice()) |ax| { + if (std.mem.eql(u8, std.mem.span(ax), name)) break true; } else false; if (!taken) return name; } unreachable; } + /// Converts `shape`'s declared partitioning into a `#sdy.sharding` attribute + /// referencing this mesh. Entries in `[...]` follow tensor dimension order. + /// Mesh axis indices in the specs resolve to this mesh's logical axis names. + /// The tensor rank and logical mesh rank are independent. + /// + /// Each dimension spec is encoded as follows: + /// - `PartitionSpec.sharded(i)` -> `{"axis_name"}`: split the dimension over + /// logical mesh axis `i`. The dimension is closed to further partitioning. + /// - `.replicated` -> `{}`: keep the dimension whole on each shard. It is + /// closed, so the partitioner cannot assign a mesh axis to it. + /// - `.open` -> `{?}`: let the partitioner infer sharding from the graph. + /// It may acquire available mesh axes or remain replicated. + /// Closed therefore means no additional mesh axes, not necessarily replicated. + /// Every dimension within `shape.rank()` must have a spec other than `.out_of_bound`. + /// + /// Dimension replication (`{}`) applies to one tensor dimension. The optional + /// tensor-wide `replicated={...}` instead names mesh axes: each listed axis + /// must carry identical tensor data across its coordinates and cannot be used + /// to split any dimension, including an open one. + /// + /// For example, take a mesh `@dp_mp` with `data=2`, `model=2`, no extra replicas, + /// and a shape whose dimensions are ordered `batch`, `hidden`: + /// + /// ``` + /// Partitioning: .{ .batch = .data, .hidden = .open } + /// #sdy.sharding<@dp_mp, [{"data"}, {?}]> + /// ``` + /// `batch` is split over `data`. The partitioner may assign `model` to `hidden`. + /// The unused `model` axis stays available because an open dimension exists. + /// + /// ``` + /// Partitioning: .{ .batch = .data, .hidden = .replicated } + /// #sdy.sharding<@dp_mp, [{"data"}, {}], replicated={"model"}> + /// ``` + /// Both dimensions are closed. Unused logical axes are explicitly replicated, + /// so devices differing only in their `model` coordinate hold the same shard. + /// + /// ``` + /// Partitioning: .{ .batch = .open, .hidden = .replicated } + /// #sdy.sharding<@dp_mp, [{?}, {}]> + /// ``` + /// Only `batch` may acquire axes, and both `data` and `model` are available. + /// Adding `replicated={"data", "model"}` would prevent either axis from being + /// inferred for `batch`, despite its `{?}` entry. pub fn sdyShardingAttrForShape( data: *const Mesh, parent_allocator: std.mem.Allocator, @@ -1687,24 +1782,27 @@ pub const Mesh = struct { const allocator = arena.allocator(); const dimensions = try allocator.alloc(*const shardy.DimensionShardingAttr, shape.rank()); var used: [MAX_MESH_RANK]bool = @splat(false); - for (0.., dimensions) |ax, *d| { - d.* = switch (shape.partition(ax)) { + const has_open_dimension = shape.containsPartitionSpec(.open); + for (0.., dimensions) |i, *d| { + d.* = switch (shape.partition(i)) { .out_of_bound => @panic("undefined"), .replicated => .replicated(ctx), .open => .open(ctx, &.{}), else => d: { - const axis = shape.partition(ax).meshAxis().?; - used[axis] = true; - break :d .closed(ctx, &.{.named(ctx, std.mem.span(data.logical.axes.get(axis)))}); + const ax = shape.partition(i).meshAxis().?; + used[ax] = true; + break :d .closed(ctx, &.{.named(ctx, std.mem.span(data.logical.axes.get(ax)))}); }, }; } var replicated_axes: stdx.BoundedArray(*const shardy.AxisRefAttr, MAX_MESH_RANK + 1) = .empty; var tile_count: usize = 1; - for (data.logical.axes.slice(), 0..) |axis, i| { + for (data.logical.axes.slice(), 0..) |ax, i| { tile_count *= data.grid.axis_sizes[i]; - if (!used[i]) replicated_axes.appendAssumeCapacity(.named(ctx, std.mem.span(axis))); + // Open dimensions can acquire unused logical axes during inference. + // Only reserve those axes for replication when every dimension is closed. + if (!has_open_dimension and !used[i]) replicated_axes.appendAssumeCapacity(.named(ctx, std.mem.span(ax))); } if (tile_count < data.grid.device_count) { replicated_axes.appendAssumeCapacity(.named(ctx, data.sdyReplicaAxis())); @@ -1720,8 +1818,8 @@ pub const Mesh = struct { var tile_shape: stdx.BoundedArray(i64, Shape.MAX_RANK + 1) = .empty; var tile_count: usize = 1; - for (0..shape.rank()) |ax| { - const size = if (shape.partition(ax).meshAxis()) |axis| self.grid.axis_sizes[axis] else 1; + for (0..shape.rank()) |i| { + const size = if (shape.partition(i).meshAxis()) |ax| self.grid.axis_sizes[ax] else 1; tile_shape.appendAssumeCapacity(size); tile_count *= size; } @@ -1998,7 +2096,7 @@ pub const DeviceGrid = struct { for (strategy.foldSources(part.axis) orelse &.{part.axis}) |source| { if (used.contains(source)) continue; used.insert(source); - const info = mesh_.physical.axisInfo(source) orelse continue; + const info = mesh_.physical.axisInfo(source) orelse return error.InvalidPhysicalAxis; try grid.appendAxis(mesh_, logical_axis, source, 1, @intCast(info.size)); } } @@ -2741,7 +2839,7 @@ test "mesh: open and replicated dimension mix" { .expected_num_replicas = 1, .mesh = &mesh, .shape = Shape.init(.{ .batch = 8, .model = 8 }, .f32).withPartitioning(&mesh, .{ .batch = .open, .model = .replicated }), - .expected_sdy = "#sdy.sharding<@mix_mesh, [{?}, {}], replicated={\"batch\", \"model\"}>", + .expected_sdy = "#sdy.sharding<@mix_mesh, [{?}, {}]>", .expected_shards = &.{ .{ .device_id = 0, .slices = &.{ .{ 0, 8 }, .{ 0, 8 } } }, .{ .device_id = 1, .slices = &.{ .{ 0, 8 }, .{ 0, 8 } } }, diff --git a/zml/attention/Options.zig b/zml/attention/Options.zig new file mode 100644 index 00000000..539f753f --- /dev/null +++ b/zml/attention/Options.zig @@ -0,0 +1,3 @@ +/// Query j sees the keys up to token_index + j, as in a decoder. Otherwise every query sees every +/// key, as in an encoder or a diffusion transformer, and token_index is ignored. +causal: bool = true, diff --git a/zml/attention/attention.zig b/zml/attention/attention.zig index 49fe947b..edd9b4ca 100644 --- a/zml/attention/attention.zig +++ b/zml/attention/attention.zig @@ -93,6 +93,29 @@ pub const Metadata = union(BackendType) { } }; +pub const Options = @import("Options.zig"); + +pub const Partitioning = struct { + const Logical = zml.Sharding.Partitioning.Logical; + + in_specs: struct { + q: Logical, + k: Logical, + v: Logical, + token_index: Logical = .replicated, + /// Omit to replicate backend metadata tensors. + metadata: ?zml.ops.PartitionSpecs(Metadata) = null, + }, + out_specs: Logical, + + /// Shards the `.h` dimension of Q, K, V and the output over `mesh_axis`. + /// Other dimensions, token indices and backend metadata are replicated. + pub fn head(comptime mesh_axis: @EnumLiteral()) Partitioning { + const heads: Logical = .parse(.{ .h = mesh_axis }); + return .{ .in_specs = .{ .q = heads, .k = heads, .v = heads }, .out_specs = heads }; + } +}; + /// Causal attention as used in transformers. /// /// **Shapes**: @@ -104,19 +127,46 @@ pub const Metadata = union(BackendType) { /// - .q is the number of queries /// - .k is the number of keys /// - .hd is the head dimension -pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, metadata: Metadata, parameters: Parameters) zml.Tensor { - return switch (parameters) { - .stablehlo => b: { - // Generate the attention mask. - const seq_len = k.dim(.k); - var attn_mask = zml.nn.causalAttnMask(.{ .q = seq_len, .k = seq_len }, q.dtype(), null); - - // Note: in Pytorch it would be very inefficient to generate the full attn_mask, - // then slice into it, but XLA is able to optimize this correctly. - attn_mask = attn_mask.gatherSlices(zml.Shape.init(.{ .q = q.dim(.q) }, attn_mask.dtype()), token_index.appendAxes(.{.coord}), .{}); - const attn_output = zml.nn.sdpa(q, k, v, .{ .attn_mask = attn_mask }); - break :b attn_output; - }, - .flash_attention => flash_attention.attention(q, k, v, token_index, metadata.flash_attention, parameters.flash_attention), - }; +pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, metadata: Metadata, parameters: Parameters, mesh: *const zml.Mesh, partitioning: Partitioning) zml.Tensor { + return attentionOpts(q, k, v, token_index, metadata, parameters, .{}, mesh, partitioning); } + +/// `attention`, causal or not depending on `opts`. +pub fn attentionOpts(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, metadata: Metadata, parameters: Parameters, opts: Options, mesh: *const zml.Mesh, partitioning: Partitioning) zml.Tensor { + const args: Attention = .{ .q = q, .k = k, .v = v, .token_index = token_index, .metadata = metadata, .parameters = parameters, .opts = opts }; + return zml.ops.manualComputation(Attention.forward, args, q.shape(), mesh, .{ + .in_specs = .{ + .q = partitioning.in_specs.q, + .k = partitioning.in_specs.k, + .v = partitioning.in_specs.v, + .token_index = partitioning.in_specs.token_index, + .metadata = partitioning.in_specs.metadata orelse zml.ops.partitionSpecs(metadata, .replicated), + }, + .out_specs = partitioning.out_specs, + }); +} + +const Attention = struct { + q: zml.Tensor, + k: zml.Tensor, + v: zml.Tensor, + token_index: zml.Tensor, + metadata: Metadata, + parameters: Parameters, + opts: Options, + + fn forward(args: @This(), _: zml.Shape) zml.Tensor { + return switch (args.parameters) { + .stablehlo => b: { + const seq_len = args.k.dim(.k); + const mask: ?zml.Tensor = if (args.opts.causal) + zml.nn.causalAttnMask(.{ .q = seq_len, .k = seq_len }, args.q.dtype(), null) + .gatherSlices(zml.Shape.init(.{ .q = args.q.dim(.q) }, args.q.dtype()), args.token_index.appendAxes(.{.coord}), .{}) + else + null; + break :b zml.nn.sdpa(args.q, args.k, args.v, .{ .attn_mask = mask }); + }, + .flash_attention => flash_attention.attention(args.q, args.k, args.v, args.token_index, args.metadata.flash_attention, args.parameters.flash_attention, args.opts), + }; + } +}; diff --git a/zml/exe.zig b/zml/exe.zig index 85baa7e7..32c0ac48 100644 --- a/zml/exe.zig +++ b/zml/exe.zig @@ -387,6 +387,11 @@ pub const Exe = struct { }, } } + + pub fn memoryStats(self: *const Exe) pjrt.ApiError!pjrt.CompiledMemoryStats { + const exe = try self.exe.executable(self.platform.pjrt_api); + return try exe.getCompiledMemoryStats(self.platform.pjrt_api); + } }; /// A typed executable whose input and output buffer structures are derived from diff --git a/zml/io.zig b/zml/io.zig index 43bfdf23..ab8d8e6e 100644 --- a/zml/io.zig +++ b/zml/io.zig @@ -184,19 +184,22 @@ pub const TensorStore = struct { stdx.debug.panic("Checkpoint has no tensor named {s}{s}", .{ self.prefix() orelse "", subkey }); } - pub fn maybeCreateTensor(self: View, subkey: []const u8, comptime tags: anytype, mesh_: @EnumLiteral(), partitioning: anytype) ?Tensor { - const has_tags: bool = comptime @TypeOf(tags) != @TypeOf(null); - const parsed_tags: Shape.TagsArray = if (comptime has_tags) Shape.parseTags(tags) else undefined; + pub fn maybeCreateTensor(self: View, subkey: []const u8, tags: anytype, mesh_: @EnumLiteral(), partitioning: anytype) ?Tensor { + const parsed_tags: ?[]const Shape.Tag = comptime parseTags(tags); if (mesh_ == .replicated) { - return self.maybeCreateTensorInternal(subkey, if (has_tags) parsed_tags.constSlice() else null, .replicated); + return self.maybeCreateTensorInternal(subkey, parsed_tags, .replicated); } + const resolved_mesh = self.store.mesh(mesh_); + if (@TypeOf(partitioning) == Sharding.Partitioning) { + return self.maybeCreateTensorInternal(subkey, parsed_tags, .{ .mesh = resolved_mesh, .partition = partitioning }); + } // Unspecified tensor axes are treated as replicated - const partition: Sharding.Partitioning = if (has_tags) - .parse(parsed_tags.constSlice(), resolved_mesh, .replicated, partitioning) + const partition: Sharding.Partitioning = if (parsed_tags) |p_tags| + .parse(p_tags, resolved_mesh, .replicated, partitioning) else partitioning; - return self.maybeCreateTensorInternal(subkey, if (has_tags) parsed_tags.constSlice() else null, .{ .mesh = resolved_mesh, .partition = partition }); + return self.maybeCreateTensorInternal(subkey, parsed_tags, .{ .mesh = resolved_mesh, .partition = partition }); } pub fn createHostPinnedTensor(self: View, subkey: []const u8, tags: anytype, mesh_: @EnumLiteral(), partitioning: anytype) Tensor { @@ -204,8 +207,8 @@ pub const TensorStore = struct { stdx.debug.panic("Checkpoint has no tensor named {s}{s}", .{ self.prefix() orelse "", subkey }); } - pub fn maybeCreateHostPinnedTensor(self: View, subkey: []const u8, comptime tags: anytype, mesh_: @EnumLiteral(), comptime partitioning: anytype) ?Tensor { - const tensor = self.maybeCreateTensor(subkey, tags, mesh_, partitioning); + pub fn maybeCreateHostPinnedTensor(self: View, subkey: []const u8, tags: anytype, mesh_: @EnumLiteral(), comptime partitioning: anytype) ?Tensor { + const tensor = self.maybeCreateTensor(subkey, comptime parseTags(tags), mesh_, partitioning); if (tensor) |t| { const storage = self.store.id_to_sources.getPtr(t.id); storage.?.memory = .host_pinned; @@ -214,10 +217,7 @@ pub const TensorStore = struct { } pub fn createReplicatedTensor(self: View, subkey: []const u8, tags: anytype) Tensor { - const has_tags: bool = comptime @TypeOf(tags) != @TypeOf(null); - const parsed_tags: Shape.TagsArray = if (comptime has_tags) Shape.parseTags(tags) else undefined; - - return self.maybeCreateTensorInternal(subkey, if (has_tags) parsed_tags.slice() else null, .replicated) orelse + return self.maybeCreateTensorInternal(subkey, comptime parseTags(tags), .replicated) orelse stdx.debug.panic("Checkpoint has no tensor named {s}{s}", .{ self.prefix() orelse "", subkey }); } @@ -244,6 +244,11 @@ pub const TensorStore = struct { return tensor; } + /// Accepts "null" as tags + fn parseTags(tags: anytype) ?[]const Shape.Tag { + return if (@TypeOf(tags) == @TypeOf(null)) null else Shape.parseTags(tags); + } + pub fn maybeCreateBinding(self: View, sources: []const []const u8, shape: Shape) ?Tensor { const arena = self.store.arena.allocator(); @@ -1282,20 +1287,52 @@ pub const DirectMemoryWriter = struct { self.shard_progress = states; } - fn drain(w: *std.Io.Writer, data: []const []const u8, _: usize) std.Io.Writer.Error!usize { + fn drain(w: *std.Io.Writer, data: []const []const u8, splat: usize) std.Io.Writer.Error!usize { const self: *DirectMemoryWriter = @alignCast(@fieldParentPtr("interface", w)); - const chunk = data[0]; try self.commitWindow(); + if (self.interface.end == self.interface.buffer.len) return std.Io.Writer.Error.WriteFailed; - const writable = self.interface.buffer.len - self.interface.end; - if (writable == 0) return std.Io.Writer.Error.WriteFailed; + const consumed = consumed: { + var n: usize = 0; + for (data[0 .. data.len - 1]) |bytes| { + const copied = self.writeWindow(bytes); + n += copied; - const n = @min(writable, chunk.len); - @memcpy(self.interface.buffer[self.interface.end..][0..n], chunk[0..n]); + if (copied < bytes.len) break :consumed n; + } + + const pattern = data[data.len - 1]; + switch (pattern.len) { + 0 => {}, + 1 => { + const dest = self.interface.buffer[self.interface.end..]; + const count = @min(dest.len, splat); + @memset(dest[0..count], pattern[0]); + + self.interface.end += count; + n += count; + }, + else => for (0..splat) |_| { + const copied = self.writeWindow(pattern); + n += copied; + + if (copied < pattern.len) break; + }, + } + + break :consumed n; + }; + + try self.commitWindow(); + return consumed; + } + + fn writeWindow(self: *DirectMemoryWriter, bytes: []const u8) usize { + const dest = self.interface.buffer[self.interface.end..]; + const n = @min(dest.len, bytes.len); + @memcpy(dest[0..n], bytes[0..n]); self.interface.end += n; - try self.commitWindow(); - return n; } diff --git a/zml/kernel.zig b/zml/kernel.zig deleted file mode 100644 index 03a64d70..00000000 --- a/zml/kernel.zig +++ /dev/null @@ -1,1614 +0,0 @@ -const std = @import("std"); - -const cuda_tile_builder = @import("dsl/cuda_tile/builder"); -const cute_builder = @import("dsl/cute/builder"); -const fly_builder = @import("dsl/fly/builder"); -const func = @import("mlir/dialects/func"); -const metal_builder = @import("dsl/metal/builder"); -const mlir = @import("mlir"); -const mosaic_tpu_builder = @import("dsl/mosaic_tpu/builder"); -const mosaic_tpu_dialect = @import("mlir/dialects/mosaic_tpu"); -const stablehlo = @import("mlir/dialects/stablehlo"); -const stdx = @import("stdx"); -const tcl_builder = @import("dsl/tcl/builder"); -const triton_builder = @import("dsl/triton/builder"); - -const Compiler = @import("Compiler.zig"); -const DataType = @import("dtype.zig").DataType; -const mlirx = @import("mlirx.zig"); -const ops = @import("ops.zig"); -const platform_ = @import("platform.zig"); -const Shape = @import("shape.zig").Shape; -const Tensor = @import("tensor.zig").Tensor; - -/// Build a struct type with one field per name, all of type `FieldT`. -fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { - var name_slices: [names.len][]const u8 = undefined; - for (names, 0..) |n, i| name_slices[i] = n; - return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); -} - -fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { - mlir.registerPasses("Transforms"); - - const registry = try mlir.DialectRegistry.init(); - defer registry.deinit(); - - registerDialects(registry); - - func.registerExtensions(registry); - - const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); - ctx.loadAllAvailableDialects(); - - return ctx; -} - -fn resolveOutputOperandAliases( - aliases_opt: anytype, - operand_offset: i64, -) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { - var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; - - if (aliases_opt) |a| { - inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { - if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ - .output_index = @intCast(output_index), - .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, - }); - } - } - - return out; -} - -pub const triton = struct { - pub const Builder = triton_builder.Builder; - pub const Value = triton_builder.Value; - pub const DType = triton_builder.DType; - pub const FinishError = triton_builder.FinishError; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(triton_builder.registerDialects); - } - - pub fn from(dt: DataType) DType { - return switch (dt) { - .bool => .i1, - .i8, .u8 => .i8, - .i16, .u16 => .i16, - .i32, .u32 => .i32, - .i64, .u64 => .i64, - .f16 => .f16, - .bf16 => .bf16, - .f32 => .f32, - .f64 => .f64, - .f8e4m3fn => .f8e4m3fn, - .f8e4m3fnuz => .f8e4m3fnuz, - .f8e5m2 => .f8e5m2, - else => std.debug.panic("zml.kernel.triton.from: dtype {s} has no Triton equivalent", .{@tagName(dt)}), - }; - } - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - - pub const CallOpts = struct { - cfg: ConfigT, - grid: [3]i32, - num_stages: i32, - num_warps: i32, - output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, - debug: bool = false, - /// Bytes for Triton's implicit global scratch argument, across all CTAs. - global_scratch_memory_size: i32 = 0, - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try triton_builder.Builder.open(allocator, ctx, name); - defer b.deinit(); - - try spec.run(&b, cfg); - - return b.finish(); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - - const ttir = emit(cur.allocator, opts.cfg) catch |err| - std.debug.panic("zml.kernel.triton.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ttir); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); - - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); - - const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); - - const call_opts: ops.TritonOps = .{ - .debug = opts.debug, - .name = name, - .ir = ttir, - .grid = opts.grid, - .num_stages = opts.num_stages, - .num_warps = opts.num_warps, - .output_operand_aliases = aliases.constSlice(), - .is_tma_allowed = opts.global_scratch_memory_size > 0, - .global_scratch_memory_size = opts.global_scratch_memory_size, - }; - - const triton_results = if (opts.global_scratch_memory_size > 0) blk: { - // XLA maps the final custom-call result to Triton's implicit - // scratch pointer; it must not appear in the TTIR signature. - const extended = outputs_arr ++ [_]Shape{.init(.{opts.global_scratch_memory_size}, .u8)}; - const res = ops.triton(inputs_arr, extended, call_opts); - break :blk res[0..spec.outputs.len].*; - } else ops.triton(inputs_arr, outputs_arr, call_opts); - - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = triton_results[i]; - return results; - } - }; - } -}; - -pub const mosaic_tpu = struct { - const Builder = mosaic_tpu_builder.Builder; - const DType = mosaic_tpu_builder.DType; - const FinishError = mosaic_tpu_builder.FinishError; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(mosaic_tpu_builder.registerDialects); - } - - pub fn from(dt: DataType) DType { - return switch (dt) { - .bool => .i1, - .i8, .u8 => .i8, - .i16, .u16 => .i16, - .i32, .u32 => .i32, - .i64, .u64 => .i64, - .f16 => .f16, - .bf16 => .bf16, - .f32 => .f32, - .f64 => .f64, - .f8e4m3fn => .f8e4m3fn, - .f8e5m2 => .f8e5m2, - else => std.debug.panic("zml.kernel.mosaic_tpu.from: dtype {s} has no Mosaic-TPU equivalent", .{@tagName(dt)}), - }; - } - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - - pub const CallExtras = struct { - vmem_limit_bytes: ?i64 = null, - disable_bounds_checks: ?bool = null, - disable_semaphore_checks: ?bool = null, - has_communication: ?bool = null, - additional_attributes: []const mlir.NamedAttribute = &.{}, - output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, - dynamic_grid_bounds: []const Tensor = &.{}, - }; - - pub const CallOpts = struct { - cfg: ConfigT, - extras: CallExtras = .{}, - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try mosaic_tpu_builder.Builder.open(allocator, ctx, name); - defer b.deinit(); - - try spec.run(&b, cfg); - - return b.finishOpts(&.{}, .{ .canonicalize = true }); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - - const ir = emit(cur.allocator, opts.cfg) catch |err| - std.debug.panic("zml.kernel.mosaic_tpu.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ir); - - const backend_config = buildBackendConfig(cur.allocator, ir, .{ - .vmem_limit_bytes = opts.extras.vmem_limit_bytes, - .disable_bounds_checks = opts.extras.disable_bounds_checks, - .disable_semaphore_checks = opts.extras.disable_semaphore_checks, - .has_communication = opts.extras.has_communication, - }) catch |err| - std.debug.panic("zml.kernel.mosaic_tpu.Kernel({s}).call: backend_config build failed: {}", .{ name, err }); - defer cur.allocator.free(backend_config); - - var values: [spec.inputs.len]*const mlir.Value = undefined; - inline for (spec.inputs, 0..) |fname, i| values[i] = @field(inputs, fname).value(); - - var res_types: [spec.outputs.len]*const mlir.Type = undefined; - inline for (&res_types, spec.outputs) |*r, fname| { - r.* = mlirx.Type.rankedTensor(cur.mlir_ctx, @field(outputs, fname)); - } - - const aliases = resolveOutputOperandAliases(opts.extras.output_operand_aliases, @intCast(opts.extras.dynamic_grid_bounds.len)); - - const op = callTpuCustomCall(.{ - .inputs = &values, - .result_types = &res_types, - .backend_config = backend_config, - .aliases = aliases.constSlice(), - .dynamic_grid_bounds = opts.extras.dynamic_grid_bounds, - .additional_attributes = opts.extras.additional_attributes, - }); - - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| { - const out = @field(outputs, fname); - @field(results, fname) = Tensor._result(out, op.result(i)); - } - return results; - } - }; - } - - const CustomCallConfig = struct { - body: []const u8, - has_communication: ?bool = null, - serialization_format: i64 = 1, - needs_layout_passes: bool = true, - shape_invariant_numerics: bool = false, - disable_bounds_checks: ?bool = null, - disable_semaphore_checks: ?bool = null, - }; - - const ScopedMemoryConfig = struct { - memory_space: i64 = 1, - offset: i64 = 0, - size: i64, - }; - - const BackendConfig = struct { - custom_call_config: CustomCallConfig, - scoped_memory_configs: ?[]const ScopedMemoryConfig = null, - }; - - const BackendConfigOpts = struct { - vmem_limit_bytes: ?i64 = null, - disable_bounds_checks: ?bool = null, - disable_semaphore_checks: ?bool = null, - has_communication: ?bool = null, - }; - - fn buildBackendConfig( - allocator: std.mem.Allocator, - ir: [:0]const u8, - opts: BackendConfigOpts, - ) ![]u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - const module = try mlir.Module.parse(ctx, ir); - defer module.deinit(); - - mosaic_tpu_dialect.registerMosaicSerdePass(); - const prev_allow_unregistered = ctx.allowUnregisteredDialects(); - ctx.setAllowUnregisteredDialects(true); - defer ctx.setAllowUnregisteredDialects(prev_allow_unregistered); - - const pm = mlir.PassManager.init(ctx); - defer pm.deinit(); - try pm.asOpPassManager().addPipeline("mosaic-serde{serialize=true}"); - try pm.runOnOp(module.operation()); - - const Encoder = std.base64.standard.Encoder; - const b64 = b64: { - var bytecode: std.Io.Writer.Allocating = .init(allocator); - defer bytecode.deinit(); - try module.operation().writeBytecode(.{ .desired_emit_version = 0 }, &bytecode.writer); - const buf = try allocator.alloc(u8, Encoder.calcSize(bytecode.written().len)); - break :b64 Encoder.encode(buf, bytecode.written()); - }; - defer allocator.free(b64); - - const config = BackendConfig{ - .custom_call_config = .{ - .body = b64, - .has_communication = opts.has_communication, - .disable_bounds_checks = opts.disable_bounds_checks, - .disable_semaphore_checks = opts.disable_semaphore_checks, - }, - .scoped_memory_configs = if (opts.vmem_limit_bytes) |vmem| - &.{.{ .size = vmem }} - else - null, - }; - - var json: std.Io.Writer.Allocating = .init(allocator); - errdefer json.deinit(); - - try json.writer.print("{f}", .{std.json.fmt(config, .{ .emit_null_optional_fields = false })}); - return try json.toOwnedSlice(); - } - - const TpuCustomCallArgs = struct { - inputs: []const *const mlir.Value, - result_types: []const *const mlir.Type, - backend_config: []const u8, - aliases: []const stablehlo.CustomCallOpts.OutputOperandAlias, - dynamic_grid_bounds: []const Tensor, - additional_attributes: []const mlir.NamedAttribute, - }; - - fn callTpuCustomCall(args: TpuCustomCallArgs) *mlir.Operation { - const cur = Compiler.current(); - - var all_inputs: stdx.BoundedArray(*const mlir.Value, stablehlo.CustomCallOpts.MAX_OPERANDS) = .empty; - for (args.dynamic_grid_bounds) |t| all_inputs.appendAssumeCapacity(t.value()); - all_inputs.appendSliceAssumeCapacity(args.inputs); - - return stablehlo.custom_call( - cur.mlir_ctx, - all_inputs.constSlice(), - args.result_types, - .{ - .call_target_name = "tpu_custom_call", - .backend_config = .{ .original = args.backend_config }, - .has_side_effect = false, - .api_version = .original, - .additional_attributes = args.additional_attributes, - .output_operand_aliases = args.aliases, - }, - cur.location, - ).appendTo(cur.currentScope().block); - } -}; - -pub const cuda_tile = struct { - pub const Builder = cuda_tile_builder.Builder; - pub const Value = cuda_tile_builder.Value; - pub const DType = cuda_tile_builder.DType; - pub const FinishError = cuda_tile_builder.FinishError; - pub const EntryHint = cuda_tile_builder.EntryHint; - pub const Arch = cuda_tile_builder.Arch; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(cuda_tile_builder.registerDialects); - } - - pub fn from(dt: DataType) DType { - return switch (dt) { - .bool => .i1, - .i4, .u4 => .i4, - .i8, .u8 => .i8, - .i16, .u16 => .i16, - .i32, .u32 => .i32, - .i64, .u64 => .i64, - .f16 => .f16, - .bf16 => .bf16, - .f32 => .f32, - .f64 => .f64, - .f8e4m3fn => .f8e4m3fn, - .f8e5m2 => .f8e5m2, - .f8e8m0 => .f8e8m0fnu, - .f4e2m1 => .f4e2m1fn, - else => std.debug.panic("zml.kernel.cuda_tile.from: dtype {s} has no CUDA Tile IR equivalent", .{@tagName(dt)}), - }; - } - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - - /// No `num_warps`/`num_stages`: warp and CTA tuning is - /// `optimization_hints` inside the IR (`Opts.hints`). - pub const CallOpts = struct { - cfg: ConfigT, - grid: [3]i32, - /// Outputs XLA zeroes before the launch, by position in - /// `spec.outputs`, ascending. - zeroed_outputs: []const i32 = &.{}, - output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try cuda_tile_builder.Builder.open(allocator, ctx, name); - defer b.deinit(); - - try spec.run(&b, cfg); - - return b.finish(&.{}); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - - const ir = emit(cur.allocator, opts.cfg) catch |err| - std.debug.panic("zml.kernel.cuda_tile.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ir); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); - - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); - - const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); - - const tensor_results = ops.cudaTile(inputs_arr, outputs_arr, .{ - .name = name, - .ir = ir, - .grid = opts.grid, - .zeroed_outputs = opts.zeroed_outputs, - .output_operand_aliases = aliases.constSlice(), - }); - - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; - return results; - } - }; - } -}; - -pub const cute = struct { - pub const Builder = cute_builder.Builder; - pub const ArgSpec = cute_builder.ArgSpec; - pub const Value = cute_builder.Value; - pub const View = cute_builder.View; - pub const Atom = cute_builder.Atom; - pub const DType = cute_builder.DType; - pub const AlgebraToken = cute_builder.AlgebraToken; - pub const LayoutSpec = cute_builder.LayoutSpec; - pub const ConstrainedDynamic = cute_builder.ConstrainedDynamic; - pub const FinishError = cute_builder.FinishError; - pub const TmaConfig = cute_builder.TmaConfig; - pub const TmaFormat = cute_builder.TmaFormat; - pub const BlockScaledMmaConfig = cute_builder.BlockScaledMmaConfig; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(cute_builder.registerDialects); - } - - pub fn from(dt: DataType) DType { - return switch (dt) { - .bool => .i1, - .i8, .u8 => .i8, - .i16, .u16 => .i16, - .i32, .u32 => .i32, - .i64, .u64 => .i64, - .f16 => .f16, - .bf16 => .bf16, - .f32 => .f32, - .f64 => .f64, - .f4e2m1 => .f4e2m1fn, - .f8e4m3fn => .f8e4m3fn, - .f8e5m2 => .f8e5m2, - .f8e8m0 => .f8e8m0fnu, - else => std.debug.panic("zml.kernel.cute.from: dtype {s} has no CuTe equivalent", .{@tagName(dt)}), - }; - } - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - - /// `kernel.launch(grid=, block=)` in the Python DSL; dynamic shared - /// memory is what the kernel declares. - pub const CallOpts = struct { - cfg: ConfigT, - grid: [3]i32, - block: [3]i32, - zeroed_outputs: []const i32 = &.{}, - output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT, block: [3]i32) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try cute_builder.Builder.open(allocator, ctx, name); - defer b.deinit(); - - try spec.run(&b, cfg); - - // XLA passes one pointer per input then per output; it does - // not check the count against the kernel. - const expected = spec.inputs.len + spec.outputs.len; - if (b.args.len != expected) { - std.debug.panic("zml.kernel.cute.Kernel({s}): declareArgs declared {d} arguments, inputs + outputs are {d}", .{ name, b.args.len, expected }); - } - - return b.finish(block); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - - const ir = emit(cur.allocator, opts.cfg, opts.block) catch |err| - std.debug.panic("zml.kernel.cute.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ir); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); - - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); - - const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); - - const tensor_results = ops.cute(inputs_arr, outputs_arr, .{ - .name = name, - .ir = ir, - .grid = opts.grid, - .block = opts.block, - .zeroed_outputs = opts.zeroed_outputs, - .output_operand_aliases = aliases.constSlice(), - }); - - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; - return results; - } - }; - } - - fn ProgramSpec(comptime Config: type) type { - return struct { - /// Public host entry selected by XLA. The callback may give the - /// nested CUDA kernel a different (private) symbol. - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - /// A complete CuTe program: one or more `cuda.kernel` operations inside - /// a `gpu.module`, followed by a public `func.func` which constructs the - /// launch-time CuTe objects and calls `cuda.launch_ex`. - /// - /// This is the source-level equivalent of invoking a Python `@cute.jit` - /// object whose `__call__` builds TMA descriptors and launches an - /// `@cute.kernel`. XLA executes the generated host function on its CUDA - /// stream, so the custom call deliberately carries no external grid or - /// block dimensions. - pub fn Program( - comptime ConfigT: type, - comptime spec: ProgramSpec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - - pub const CallOpts = struct { - cfg: ConfigT, - zeroed_outputs: []const i32 = &.{}, - scalars: []const i64 = &.{}, - output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try cute_builder.Builder.openProgram(allocator, ctx, name); - defer b.deinit(); - try spec.run(&b, cfg); - return b.finishProgram(); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - const ir = emit(cur.allocator, opts.cfg) catch |err| - std.debug.panic("zml.kernel.cute.Program({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ir); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); - - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); - - const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); - const tensor_results = ops.cute(inputs_arr, outputs_arr, .{ - .name = name, - .ir = ir, - .zeroed_outputs = opts.zeroed_outputs, - .scalars = opts.scalars, - .output_operand_aliases = aliases.constSlice(), - }); - - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; - return results; - } - }; - } -}; - -pub const tcl = struct { - pub const Builder = tcl_builder.Builder; - pub const TensorOperation = tcl_builder.TensorOperation; - pub const Value = tcl_builder.Value; - pub const Axis = tcl_builder.Axis; - pub const Expr = tcl_builder.Expr; - pub const Dram = tcl_builder.Dram; - pub const Context = tcl_builder.Context; - pub const DType = tcl_builder.DType; - pub const VeOp = tcl_builder.VeOp; - pub const ReduceMode = tcl_builder.ReduceMode; - pub const Predicate = tcl_builder.Predicate; - pub const Tactic = tcl_builder.Tactic; - pub const FinishError = tcl_builder.FinishError; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(tcl_builder.registerDialects); - } - - pub fn from(dt: DataType) DType { - return switch (dt) { - .bool => .bool, - .i4 => .i4, - .i8 => .i8, - .i16 => .i16, - .i32 => .i32, - .i64 => .i64, - .u8 => .u8, - .u16 => .u16, - .u32 => .u32, - .u64 => .u64, - .f16 => .f16, - .bf16 => .bf16, - .f32 => .f32, - .f64 => .f64, - .f4e2m1 => .f4_e2, - .f8e4m3fn => .f8_e4, - .f8e5m2 => .f8_e5, - else => std.debug.panic("zml.kernel.tcl.from: dtype {s} has no TCL equivalent", .{@tagName(dt)}), - }; - } - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - /// A TCL function compiled by the Furiosa plugin as one native unit: - /// `declareArgs` declares the inputs in order, `ret` the outputs. - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - - pub const CallOpts = struct { - cfg: ConfigT, - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { - return build(allocator, cfg, null, null, null); - } - - /// `updated`, when given, receives the input each output updates - /// in place (`Builder.update`), or null. - fn build(allocator: std.mem.Allocator, cfg: ConfigT, inputs: ?[]const Shape, outputs: ?[]const Shape, updated: ?*[spec.outputs.len]?usize) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try Builder.open(allocator, ctx, name); - defer b.deinit(); - try spec.run(&b, cfg); - - const results = b.results orelse std.debug.panic("zml.kernel.tcl.Kernel({s}): run did not call ret", .{name}); - if (b.args.len != spec.inputs.len or results.len != spec.outputs.len) { - std.debug.panic("zml.kernel.tcl.Kernel({s}): {d} arguments and {d} results for {d} inputs and {d} outputs", .{ name, b.args.len, results.len, spec.inputs.len, spec.outputs.len }); - } - - if (inputs) |shapes| for (b.args, shapes, spec.inputs) |t, s, n| check(t, s, n); - if (outputs) |shapes| for (results, shapes, spec.outputs) |t, s, n| check(t, s, n); - if (updated) |u| for (u, results) |*dst, r| { - dst.* = b.updatedArgument(r); - }; - - return b.finish(); - } - - fn check(t: tcl_builder.Tensor, s: Shape, field: []const u8) void { - var same = from(s.dtype()) == t.dtype; - if (t.shards > 1) { - // A partition holds one chip's share of a chip-split DRAM tensor. - var n: i64 = 1; - for (t.axes) |a| n *= a.size; - same = same and n == @as(i64, @intCast(s.count())) * t.shards; - } else { - same = same and s.rank() == t.axes.len; - if (same) for (t.axes, 0..) |a, i| { - same = same and a.size == s.dim(i); - }; - } - - if (!same) std.debug.panic("zml.kernel.tcl.Kernel({s}): {s} is {f}, the TCL function has {s} over {d} axes", .{ name, field, s, @tagName(t.dtype), t.axes.len }); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - var input_shapes: [spec.inputs.len]Shape = undefined; - inline for (spec.inputs, 0..) |fname, i| { - inputs_arr[i] = @field(inputs, fname); - input_shapes[i] = inputs_arr[i].shape(); - } - - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); - - var updated: [spec.outputs.len]?usize = undefined; - const ir = build(cur.allocator, opts.cfg, &input_shapes, &outputs_arr, &updated) catch |err| - std.debug.panic("zml.kernel.tcl.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ir); - - var aliases: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, spec.outputs.len) = .empty; - for (updated, 0..) |input, output| if (input) |i| aliases.appendAssumeCapacity(.{ - .output_index = @intCast(output), - .operand_index = @intCast(i), - }); - const tensor_results = ops.tcl(inputs_arr, outputs_arr, .{ .ir = ir, .output_operand_aliases = aliases.constSlice() }); - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| { - @field(results, fname) = tensor_results[i]; - } - - return results; - } - }; - } -}; - -pub const metal = struct { - pub const Builder = metal_builder.Builder; - pub const Value = metal_builder.Value; - pub const Type = metal_builder.Type; - pub const Struct = metal_builder.Struct; - pub const ArgSpec = metal_builder.ArgSpec; - pub const LoopOpts = metal_builder.LoopOpts; - pub const FinishError = metal_builder.FinishError; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(metal_builder.registerDialects); - } - - pub fn from(dt: DataType) Type { - return switch (dt) { - .bool => .bool, - .i8 => .char, - .i16 => .short, - .i32 => .int, - .i64 => .long, - .u8 => .uchar, - .u16 => .ushort, - .u32 => .uint, - .u64 => .ulong, - .f16 => .half, - .bf16 => .bfloat, - .f32 => .float, - else => std.debug.panic("zml.kernel.metal.from: dtype {s} has no MSL equivalent", .{@tagName(dt)}), - }; - } - - /// The bytes of a `constant T &` argument: `.of(@as(i32, n))`, or a host - /// `extern struct` laid out like a `constant S &` argument. - pub const Constant = struct { - buffer: [256]u8 = undefined, - len: usize = 0, - - pub fn of(value: anytype) Constant { - const raw = std.mem.asBytes(&value); - var c: Constant = .{ .len = raw.len }; - @memcpy(c.buffer[0..raw.len], raw); - return c; - } - - pub fn bytes(self: *const Constant) []const u8 { - return self.buffer[0..self.len]; - } - }; - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - /// The arguments bound to input tensors. - inputs: []const [:0]const u8, - /// The arguments bound to output tensors; one named like an - /// input updates it in place. - outputs: []const [:0]const u8, - /// The `constant` arguments given by value at the call. - constants: []const [:0]const u8 = &.{}, - run: *const fn (*Builder, Config) FinishError!void, - fast_math: bool = true, - }; - } - - /// A Metal kernel: `run` declares its arguments with `declareArgs`, by - /// name, and builds its body. - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - pub const Constants = StructOf(spec.constants, Constant); - - pub const CallOpts = struct { - cfg: ConfigT, - /// Threadgroups per grid. - grid: [3]u32, - /// Threads per threadgroup. - threadgroup: [3]u32, - /// Bytes of `[[threadgroup(0)]]` memory. - threadgroup_memory: u64 = 0, - constants: Constants = std.mem.zeroes(Constants), - }; - - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { - var arguments: std.ArrayList(metal_builder.Argument) = .empty; - defer arguments.deinit(allocator); - return build(allocator, cfg, &arguments); - } - - fn build(allocator: std.mem.Allocator, cfg: ConfigT, arguments: *std.ArrayList(metal_builder.Argument)) ![:0]const u8 { - const ctx = try newContext(); - defer ctx.deinit(); - var b = Builder.open(allocator, ctx, name, .{ .fast_math = spec.fast_math }); - defer b.deinit(); - try spec.run(&b, cfg); - const ir = try b.finish(); - errdefer allocator.free(ir); - try arguments.appendSlice(allocator, b.arguments.items); - return ir; - } - - fn indexOf(names: []const [:0]const u8, arg: []const u8) ?usize { - for (names, 0..) |n, i| if (std.mem.eql(u8, n, arg)) return i; - return null; - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - const allocator = cur.allocator; - - var arguments: std.ArrayList(metal_builder.Argument) = .empty; - defer arguments.deinit(allocator); - const ir = build(allocator, opts.cfg, &arguments) catch |err| - std.debug.panic("zml.kernel.metal.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer allocator.free(ir); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - inline for (spec.inputs, 0..) |field, i| inputs_arr[i] = @field(inputs, field); - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |field, i| outputs_arr[i] = @field(outputs, field); - var constants: [spec.constants.len][]const u8 = undefined; - inline for (spec.constants, 0..) |field, i| constants[i] = @field(opts.constants, field).bytes(); - - var bound: [spec.outputs.len]bool = @splat(false); - var bound_inputs: [spec.inputs.len]bool = @splat(false); - var bound_constants: [spec.constants.len]bool = @splat(false); - var aliases: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, spec.outputs.len) = .empty; - var bindings: std.ArrayList(i32) = .empty; - defer bindings.deinit(allocator); - for (arguments.items) |arg| { - if (arg.kind == .threadgroup) { - // `threadgroup_memory` sizes `[[threadgroup(0)]]` alone. - if (arg.index != 0) std.debug.panic("zml.kernel.metal.Kernel({s}): threadgroup argument {s} is [[threadgroup({d})]], and only [[threadgroup(0)]] gets memory", .{ name, arg.name, arg.index }); - continue; - } - std.debug.assert(arg.index == bindings.items.len); - const input = indexOf(spec.inputs, arg.name); - const output = indexOf(spec.outputs, arg.name); - if (input) |i| bound_inputs[i] = true; - if (output) |o| bound[o] = true; - const binding: i32 = if (input) |i| blk: { - if (output) |o| aliases.appendAssumeCapacity(.{ .output_index = @intCast(o), .operand_index = @intCast(i) }); - break :blk @intCast(i); - } else if (output) |o| - @intCast(spec.inputs.len + o) - else if (indexOf(spec.constants, arg.name)) |k| blk: { - if (constants[k].len != arg.size) std.debug.panic("zml.kernel.metal.Kernel({s}): constant {s} has {d} bytes, the argument {d}", .{ name, arg.name, constants[k].len, arg.size }); - bound_constants[k] = true; - break :blk -1 - @as(i32, @intCast(k)); - } else std.debug.panic("zml.kernel.metal.Kernel({s}): argument {s} is no input, output or constant", .{ name, arg.name }); - bindings.append(allocator, binding) catch @panic("OOM"); - } - for (bound, spec.outputs) |b, output| if (!b) std.debug.panic("zml.kernel.metal.Kernel({s}): output {s} is no argument", .{ name, output }); - for (bound_inputs, spec.inputs) |b, input| if (!b) std.debug.panic("zml.kernel.metal.Kernel({s}): input {s} is no argument", .{ name, input }); - for (bound_constants, spec.constants) |b, constant| if (!b) std.debug.panic("zml.kernel.metal.Kernel({s}): constant {s} is no argument", .{ name, constant }); - - const tensor_results = ops.metal(inputs_arr, outputs_arr, .{ - .name = name, - .ir = ir, - .grid = opts.grid, - .threadgroup = opts.threadgroup, - .threadgroup_memory = opts.threadgroup_memory, - .arguments = bindings.items, - .constants = &constants, - .output_operand_aliases = aliases.constSlice(), - }); - var results: Results = undefined; - inline for (spec.outputs, 0..) |field, i| @field(results, field) = tensor_results[i]; - return results; - } - }; - } -}; - -pub const fly = struct { - /// Threads per wavefront: 64 on CDNA, 32 on RDNA. - pub fn waveSize(p: *const platform_.Platform) i32 { - const cc = platform_.rocm.computeCapability(p) orelse return 64; - return switch (cc.architecture()) { - .cdna1, .cdna2, .cdna3, .cdna4 => 64, - .rdna2, .rdna3, .rdna3_5, .rdna4 => 32, - }; - } - - /// The plugin sizes a block in wavefronts, so a thread count has to divide - /// evenly: a kernel's layouts are built for an exact number of threads and - /// spare ones would partition out of range. - fn warpsFor(p: *const platform_.Platform, threads: i32) i32 { - const wave = waveSize(p); - if (@rem(threads, wave) != 0) { - std.debug.panic("zml.kernel.fly: {d} threads is not a multiple of the {d}-wide wavefront", .{ threads, wave }); - } - return @divExact(threads, wave); - } - - pub const Builder = fly_builder.Builder; - pub const Value = fly_builder.Value; - pub const DType = fly_builder.DType; - pub const FinishError = fly_builder.FinishError; - pub const TiledCopy = fly_builder.TiledCopy; - pub const TiledMma = fly_builder.TiledMma; - pub const Arch = fly_builder.Arch; - pub const MmaFlavor = fly_builder.MmaFlavor; - pub const fly_rocdl = fly_builder.fly_rocdl; - - /// The matrix atom this device provides, or null when it has none: WMMA - /// arrived with RDNA3, so gfx1030 has no matrix instruction at all. - pub fn mmaFlavor(p: *const platform_.Platform) ?MmaFlavor { - const cc = platform_.rocm.computeCapability(p) orelse return .cdna3_mfma; - return switch (cc.architecture()) { - .cdna1, .cdna2, .cdna3, .cdna4 => .cdna3_mfma, - .rdna3, .rdna3_5 => .gfx11_wmma, - .rdna4 => .gfx120x_wmma, - .rdna2 => null, - }; - } - pub const layout = fly_builder.layout; - pub const Layout = fly_builder.Layout; - pub const Tile = fly_builder.Tile; - pub const IntTuple = fly_builder.IntTuple; - pub const L = fly_builder.L; - pub const rowMajor = fly_builder.rowMajor; - pub const colMajor = fly_builder.colMajor; - pub const ordered = fly_builder.ordered; - pub const tile = fly_builder.tile; - pub const it = fly_builder.it; - - pub fn newContext() std.mem.Allocator.Error!*mlir.Context { - return makeKernelContext(fly_builder.registerDialects); - } - - pub fn from(dt: DataType) DType { - return switch (dt) { - .bool => .i1, - .i8, .u8 => .i8, - .i16, .u16 => .i16, - .i32, .u32 => .i32, - .i64, .u64 => .i64, - .f16 => .f16, - .bf16 => .bf16, - .f32 => .f32, - .f64 => .f64, - .f8e4m3fn => .f8e4m3fn, - .f8e4m3fnuz => .f8e4m3fnuz, - .f8e5m2 => .f8e5m2, - .f8e5m2fnuz => .f8e5m2fnuz, - else => std.debug.panic("zml.kernel.fly.from: dtype {s} has no Fly equivalent", .{@tagName(dt)}), - }; - } - - fn Spec(comptime Config: type) type { - return struct { - name: [:0]const u8, - inputs: []const [:0]const u8, - outputs: []const [:0]const u8, - run: *const fn (*Builder, Config) FinishError!void, - }; - } - - pub fn Kernel( - comptime ConfigT: type, - comptime spec: Spec(ConfigT), - ) type { - return struct { - pub const name: [:0]const u8 = spec.name; - pub const Config = ConfigT; - pub const Inputs = StructOf(spec.inputs, Tensor); - pub const Outputs = StructOf(spec.outputs, Shape); - pub const Results = StructOf(spec.outputs, Tensor); - pub const arg_names = spec.inputs ++ spec.outputs; - /// By name, as `!fly.memref` views. - pub const Args = StructOf(arg_names, Value); - - pub const CallOpts = struct { - cfg: ConfigT, - grid: [3]i32, - /// Threads per block. Converted to the wavefronts the plugin - /// wants using the device's wave size, which is 64 on CDNA and - /// 32 on RDNA, so a kernel states the thread count its layouts - /// were built for and runs on both. - threads: i32, - /// > 0 sets `amdgpu-waves-per-eu = "N, N"`: a hard occupancy clamp. - waves_per_eu: i32 = 0, - /// Required when the body uses `fly.get_dyn_shared`. - shared_mem_bytes: i64 = 0, - /// Inputs first, then outputs; see `zeroedOutput`. A no-op - /// inside command buffers. - zeroed_args: []const i32 = &.{}, - output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, - }; - - /// The arguments `run` was declared with. - pub fn args(b: *Builder) Args { - var out: Args = undefined; - inline for (arg_names, 0..) |fname, i| @field(out, fname) = b.arg(i); - return out; - } - - /// `zeroed_args` index of an output. - pub fn zeroedOutput(comptime f: std.meta.FieldEnum(Outputs)) i32 { - return @intCast(spec.inputs.len + @backingInt(f)); - } - - /// `shapes` is inputs, then outputs. - pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT, shapes: []const Shape) ![:0]const u8 { - std.debug.assert(shapes.len == arg_names.len); - const ctx = try newContext(); - defer ctx.deinit(); - - var b = try fly_builder.Builder.open(allocator, ctx, name); - defer b.deinit(); - - var specs: [arg_names.len]Builder.ArgSpec = undefined; - inline for (arg_names, 0..) |fname, i| { - const shape = shapes[i]; - specs[i] = .{ - .name = fname, - .dtype = from(shape.dtype()), - .dims = if (shape.rank() > 0) shape.dims() else null, - }; - } - try b.declareArgsLow(&specs); - - try spec.run(&b, cfg); - - return b.finish(); - } - - pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { - const cur = Compiler.current(); - - var inputs_arr: [spec.inputs.len]Tensor = undefined; - inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); - - var outputs_arr: [spec.outputs.len]Shape = undefined; - inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); - - var shapes: [arg_names.len]Shape = undefined; - inline for (0..spec.inputs.len) |i| shapes[i] = inputs_arr[i].shape(); - inline for (0..spec.outputs.len) |i| shapes[spec.inputs.len + i] = outputs_arr[i]; - - const ir = emit(cur.allocator, opts.cfg, &shapes) catch |err| - std.debug.panic("zml.kernel.fly.Kernel({s}).call: emit failed: {}", .{ name, err }); - defer cur.allocator.free(ir); - - const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); - - const tensor_results = ops.fly(inputs_arr, outputs_arr, .{ - .name = name, - .ir = ir, - .grid = opts.grid, - .num_warps = warpsFor(cur.platform, opts.threads), - .waves_per_eu = opts.waves_per_eu, - .shared_mem_bytes = opts.shared_mem_bytes, - .zeroed_args = opts.zeroed_args, - .output_operand_aliases = aliases.constSlice(), - }); - - var results: Results = undefined; - inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; - return results; - } - }; - } -}; - -/// C = A + B over (M, N) f32, 128-bit vectorized and predicated on the -/// ragged edge. -const FlyVectorAdd = struct { - const Cfg = struct { - thr: fly.Layout = fly.ordered(.{ 8, 16 }, .{ 1, 0 }), - val: fly.Layout = fly.ordered(.{ 1, 4 }, .{ 0, 1 }), - }; - - const K = fly.Kernel(Cfg, .{ - .name = "vector_add", - .inputs = &.{ "a", "b" }, - .outputs = &.{"c"}, - .run = run, - }); - - fn run(b: *fly.Builder, cfg: Cfg) fly.FinishError!void { - const a = K.args(b); - fly_builder.emitVectorAdd(b, cfg.thr, cfg.val, a.a, a.b, a.c); - } -}; - -test "fly kernel emits a module XLA can parse" { - const shapes = [_]Shape{ .init(.{ 100, 1000 }, .f32), .init(.{ 100, 1000 }, .f32), .init(.{ 100, 1000 }, .f32) }; - const ir = try FlyVectorAdd.K.emit(std.testing.allocator, .{}, &shapes); - defer std.testing.allocator.free(ir); - try std.testing.expect(std.mem.indexOf(u8, ir, "gpu.func @vector_add(%arg0: !fly.ptr, %arg1: !fly.ptr, %arg2: !fly.ptr) kernel") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "fly.static : !fly.tile<[8|64]>") != null); -} - -/// C = A @ B^T with one MFMA-tiled block. -const FlyTiledMma = struct { - const Cfg = struct { flavor: fly.MmaFlavor }; - const K = fly.Kernel(Cfg, .{ - .name = "tiled_mma", - .inputs = &.{ "a", "b" }, - .outputs = &.{"c"}, - .run = run, - }); - - fn run(b: *fly.Builder, cfg: Cfg) fly.FinishError!void { - const a = K.args(b); - fly_builder.emitTiledMma(b, cfg.flavor, a.a, a.b, a.c); - } -}; - -test "fly kernels run on rocm" { - const zml = @import("zml.zig"); - const platform = zml.testing.env(); - if (platform.target != .rocm) return error.SkipZigTest; - - const allocator = std.testing.allocator; - - // Compile `forward`, upload both inputs, run, and bring the result back. - const call = struct { - fn f(comptime forward: anytype, p: *const zml.Platform, ta: Tensor, ha: anytype, tb: Tensor, hb: anytype) !zml.Slice { - var exe = try zml.module.compile(std.testing.allocator, std.testing.io, forward, .{ ta, tb }, p, .{}); - defer exe.deinit(); - var ba: zml.Buffer = try .fromBytes(std.testing.io, p, ta.shape(), std.mem.sliceAsBytes(ha)); - defer ba.deinit(); - var bb: zml.Buffer = try .fromBytes(std.testing.io, p, tb.shape(), std.mem.sliceAsBytes(hb)); - defer bb.deinit(); - var out = try exe.eval(std.testing.allocator, std.testing.io, .{ ba, bb }); - defer out.deinit(); - return out.toSliceAlloc(std.testing.allocator, std.testing.io); - } - }.f; - - // vector add: elementwise over a ragged (100, 1000), predicated at the edge. - { - const M, const N = .{ 100, 1000 }; - const Mod = struct { - pub fn forward(a: Tensor, b: Tensor) Tensor { - return FlyVectorAdd.K.call(.{ .a = a, .b = b }, .{ .c = a.shape() }, .{ - .cfg = .{}, - .grid = .{ (M + 8 - 1) / 8, (N + 64 - 1) / 64, 1 }, - .threads = 8 * 16, - }).c; - } - }; - - const ha = try allocator.alloc(f32, M * N); - defer allocator.free(ha); - const hb = try allocator.alloc(f32, M * N); - defer allocator.free(hb); - for (ha, hb, 0..) |*va, *vb, i| { - va.* = @floatFromInt(i % 977); - vb.* = @floatFromInt(1000 + (i % 313)); - } - - var host = try call(Mod.forward, platform, .init(.{ M, N }, .f32), ha, .init(.{ M, N }, .f32), hb); - defer host.free(allocator); - for (host.items(f32), 0..) |v, i| try std.testing.expectEqual(ha[i] + hb[i], v); - } - - // tiled mma: C = A @ B^T in one block of 256 threads, over whatever matrix - // atom this device has. CDNA multiplies f32 through MFMA; RDNA3+ has WMMA, - // whose verifier rejects f32 operands, so the operand type and K come from - // the flavour rather than being written into the test. - if (fly.mmaFlavor(platform)) |flavor| switch (flavor) { - inline else => |f| { - const Elem = switch (comptime f.operandDType()) { - .f32 => f32, - .f16 => f16, - else => @compileError("unhandled MMA operand type"), - }; - const M, const N = .{ 64, 64 }; - const Kd = comptime f.blockK(); - const operand_dtype: DataType = switch (comptime f.operandDType()) { - .f32 => .f32, - .f16 => .f16, - else => unreachable, - }; - - const Mod = struct { - pub fn forward(a: Tensor, b: Tensor) Tensor { - return FlyTiledMma.K.call(.{ .a = a, .b = b }, .{ .c = Shape.init(.{ M, N }, .f32) }, .{ - .cfg = .{ .flavor = f }, - .grid = .{ 1, 1, 1 }, - .threads = comptime f.blockThreads(), - }).c; - } - }; - - var ha: [M * Kd]Elem = undefined; - var hb: [N * Kd]Elem = undefined; - for (&ha, 0..) |*v, i| v.* = @as(Elem, @floatFromInt((i * 7) % 13)) - 6; - for (&hb, 0..) |*v, i| v.* = @as(Elem, @floatFromInt((i * 5) % 11)) - 5; - - var host = try call(Mod.forward, platform, .init(.{ M, Kd }, operand_dtype), &ha, .init(.{ N, Kd }, operand_dtype), &hb); - defer host.free(allocator); - - // f16 operands accumulate in f32, so the error is set by the input - // rounding rather than the sum. - const tol: f32 = if (Elem == f16) 1e-2 else 1e-3; - const out = host.items(f32); - for (0..M) |m| for (0..N) |n| { - var acc: f32 = 0; - for (0..Kd) |k| acc += @as(f32, ha[m * Kd + k]) * @as(f32, hb[n * Kd + k]); - try std.testing.expectApproxEqAbs(acc, out[m * N + n], tol); - }; - }, - }; -} - -/// `furiosa.tcl.examples.softmax` over f32. -const TclSoftmax = struct { - const Cfg = struct { b: i64, s: i64, d: i64 }; - const K = tcl.Kernel(Cfg, .{ - .name = "softmax_kernel", - .inputs = &.{"x"}, - .outputs = &.{"y"}, - .run = run, - }); - - fn run(b: *tcl.Builder, cfg: Cfg) tcl.FinishError!void { - const B = b.axis("B", cfg.b); - const S = b.axis("S", cfg.s); - const D = b.axis("D", cfg.d); - const t = try b.declareArgs(.{ .x = .{ .dtype = .f32, .axes = &.{ B, S, D } } }); - - const op = b.tensorOperation(.{}); - const x = op.fetch(t.x, .{}); - const e = x.subf(x.reduce(&.{D}, .maxf)).exp(); - b.ret(&.{try op.commit(e.divf(e.reduce(&.{D}, .addf)), .{})}); - } -}; - -test "tcl kernels run on furiosa" { - const zml = @import("zml.zig"); - const allocator = std.testing.allocator; - const cfg: TclSoftmax.Cfg = .{ .b = 2, .s = 16, .d = 256 }; - - const ir = try TclSoftmax.K.emit(allocator, cfg); - defer allocator.free(ir); - try std.testing.expect(std.mem.indexOf(u8, ir, "func.func @softmax_kernel(") != null); - - const platform = zml.testing.env(); - if (platform.target != .furiosa) return error.SkipZigTest; - - const Mod = struct { - pub fn forward(x: Tensor) Tensor { - return TclSoftmax.K.call(.{ .x = x }, .{ .y = x.shape() }, .{ .cfg = cfg }).y; - } - }; - const x: Tensor = .init(.{ cfg.b, cfg.s, cfg.d }, .f32); - var exe = try zml.module.compile(allocator, std.testing.io, Mod.forward, .{x}, platform, .{}); - defer exe.deinit(); - - var input: [cfg.b * cfg.s * cfg.d]f32 = undefined; - for (&input, 0..) |*v, i| v.* = @as(f32, @floatFromInt(i % 37)) * 0.25 - 4; - var x_buffer: zml.Buffer = try .fromBytes(std.testing.io, platform, x.shape(), std.mem.sliceAsBytes(&input)); - defer x_buffer.deinit(); - var result = try exe.eval(allocator, std.testing.io, .{x_buffer}); - defer result.deinit(); - var host = try result.toSliceAlloc(allocator, std.testing.io); - defer host.free(allocator); - - const out = host.items(f32); - var row: usize = 0; - while (row < cfg.b * cfg.s) : (row += 1) { - const in = input[row * cfg.d ..][0..cfg.d]; - var max: f32 = in[0]; - for (in) |v| max = @max(max, v); - var sum: f32 = 0; - for (in) |v| sum += @exp(v - max); - for (in, out[row * cfg.d ..][0..cfg.d]) |v, y| try std.testing.expectApproxEqAbs(@exp(v - max) / sum, y, 1e-4); - } -} - -test "cute kernel emits the module the CuTe compiler takes" { - const Cfg = struct { n: i64 }; - const AddOne = cute.Kernel(Cfg, .{ - .name = "add_one", - .inputs = &.{"x"}, - .outputs = &.{"out"}, - .run = struct { - fn run(b: *cute.Builder, cfg: Cfg) cute.FinishError!void { - const a = try b.declareArgs(.{ - .x = .{ .tensor = .{ .dtype = .f32, .shape = &.{cfg.n} } }, - .out = .{ .tensor = .{ .dtype = .f32, .shape = &.{cfg.n} } }, - }); - const i = b.blockIdx().x.mul(b.blockDim().x).add(b.threadIdx().x); - var guard = b.openIf(i.lt(cfg.n)); - { - a.out.set(.{i}, a.x.get(.{i}).add(1.0)); - guard.close(); - } - } - }.run, - }); - - const ir = try AddOne.emit(std.testing.allocator, .{ .n = 1000 }, .{ 128, 1, 1 }); - defer std.testing.allocator.free(ir); - try std.testing.expect(std.mem.indexOf(u8, ir, "\"func.func\"() <{function_type = (") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "sym_name = \"add_one\"") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "nvvm.reqntid = array") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "cute.memref.load") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "gpu.module") == null); -} - -test "cuda_tile kernel emits a module XLA can parse" { - const Cfg = struct { n: i64 }; - const AddOne = cuda_tile.Kernel(Cfg, .{ - .name = "add_one", - .inputs = &.{"x"}, - .outputs = &.{"out"}, - .run = struct { - fn run(b: *cuda_tile.Builder, cfg: Cfg) cuda_tile.FinishError!void { - const a = try b.declareArgs(.{ .x = .{ .ptr = .f32 }, .out = .{ .ptr = .f32 } }); - const vx = b.partitionView(b.tensorView(a.x, &.{cfg.n}, &.{1}), &.{128}, .{}); - const vo = b.partitionView(b.tensorView(a.out, &.{cfg.n}, &.{1}), &.{128}, .{}); - const bid = b.tileBlockId(); - _ = b.store(b.load(vx, &.{bid.x}).add(1.0), vo, &.{bid.x}); - } - }.run, - }); - - const ir = try AddOne.emit(std.testing.allocator, .{ .n = 1024 }); - defer std.testing.allocator.free(ir); - try std.testing.expect(std.mem.indexOf(u8, ir, "cuda_tile.module @add_one") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "entry @add_one(") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "load_view_tko") != null); - try std.testing.expect(std.mem.indexOf(u8, ir, "store_view_tko") != null); -} - -/// exp over a tensor split across two chips in one launch, returning the -/// split result and its all_gather. -const TclSplitExp = struct { - const Cfg = struct {}; - const K = tcl.Kernel(Cfg, .{ .name = "split_exp", .inputs = &.{"x"}, .outputs = &.{ "y", "all" }, .run = run }); - fn run(b: *tcl.Builder, _: Cfg) tcl.FinishError!void { - const C = b.axis("C", 2); - const A = b.axis("A", 1024); - const t = try b.declareArgs(.{ .x = .{ .dtype = .f32, .axes = &.{ C, A }, .dram = tcl.Dram{ .chip = &.{.{ .axis = C }} } } }); - const op = b.tensorOperation(.{ .context = .{ .layout = .{ .chip = .{ .axis = C } } } }); - const local = try op.commit(op.fetch(t.x, .{}).exp(), .{}); - const all = try b.allGather(local, C, .{}); - b.ret(&.{ try b.asDram(local, .{ .chip = &.{.{ .axis = C }} }), try b.asDram(all, .{}) }); - } - fn body(ctx: struct { x: Tensor }, shapes: [2]Shape) [2]Tensor { - const r = K.call(.{ .x = ctx.x }, .{ .y = shapes[0], .all = shapes[1] }, .{ .cfg = .{} }); - return .{ r.y, r.all }; - } -}; - -// Needs FURIOSA_VISIBLE_DEVICES=0,1. -test "tcl multi-chip kernels run on two furiosa chips" { - const zml = @import("zml.zig"); - const allocator = std.testing.allocator; - const platform = zml.testing.env(); - if (platform.target != .furiosa or platform.devices.len != 2) return error.SkipZigTest; - const model = platform.meshes.get("model").?; - const x_shape = Shape.init(.{ .c = 2, .a = 1024 }, .f32).withPartitioning(model, .{ .c = .model }); - const Mod = struct { - pub fn forward(x: Tensor) [2]Tensor { - const r = zml.ops.manualComputation(TclSplitExp.body, .{ .x = x }, .{ x.shape(), Shape.init(.{ .c = 2, .a = 1024 }, .f32) }); - return .{ r[0], r[1] }; - } - }; - const x: Tensor = .fromShape(x_shape); - var exe = try zml.module.compile(allocator, std.testing.io, Mod.forward, .{x}, platform, .{ .meshes = &.{model} }); - defer exe.deinit(); - var hx: [2 * 1024]f32 = undefined; - for (&hx, 0..) |*v, i| v.* = @as(f32, @floatFromInt(i % 13)) * 0.125 - 0.5; - var bx: zml.Buffer = try .fromBytes(std.testing.io, platform, x_shape, std.mem.sliceAsBytes(&hx)); - defer bx.deinit(); - var runner = try exe.runner(allocator); - defer runner.deinit(allocator); - var res: [2]zml.Buffer = undefined; - runner.run(std.testing.io, .{bx}, .{ &res[0], &res[1] }, .{ .wait = true }); - defer for (&res) |*r| r.deinit(); - for (res) |r| { - var host = try r.toSliceAlloc(allocator, std.testing.io); - defer host.free(allocator); - for (host.items(f32), 0..) |v, i| try std.testing.expectApproxEqRel(@exp(hx[i]), v, 1e-4); - } -} - -/// `out[i] = a[i] * scale + b[i]` over `n` floats. -const MetalAxpy = struct { - const Cfg = struct {}; - const K = metal.Kernel(Cfg, .{ - .name = "axpy", - .inputs = &.{ "a", "b" }, - .outputs = &.{"out"}, - .constants = &.{ "scale", "n" }, - .run = run, - }); - - fn run(b: *metal.Builder, _: Cfg) metal.FinishError!void { - const a = try b.declareArgs(.{ - .a = .{ .buffer = .{ .type = .float, .const_ = true } }, - .b = .{ .buffer = .{ .type = .float, .const_ = true } }, - .out = .{ .buffer = .{ .type = .float } }, - .scale = .{ .constant = .float }, - .n = .{ .constant = .uint }, - .tid = .{ .thread_position_in_grid = .uint }, - }); - b.returnIf(a.tid.ge(a.n)); - a.out.store(a.tid, b.fma(a.a.load(a.tid), a.scale, a.b.load(a.tid))); - } -}; - -test "metal kernels run on metal" { - const zml = @import("zml.zig"); - const allocator = std.testing.allocator; - - const ir = try MetalAxpy.K.emit(allocator, .{}); - defer allocator.free(ir); - try std.testing.expect(std.mem.indexOf(u8, ir, "air.fma") != null); - - const platform = zml.testing.env(); - if (platform.target != .metal) return error.SkipZigTest; - - const n = 1000; - const Mod = struct { - pub fn forward(x: Tensor, y: Tensor) Tensor { - return MetalAxpy.K.call(.{ .a = x, .b = y }, .{ .out = x.shape() }, .{ - .cfg = .{}, - .grid = .{ (n + 255) / 256, 1, 1 }, - .threadgroup = .{ 256, 1, 1 }, - .constants = .{ .scale = .of(@as(f32, 0.5)), .n = .of(@as(u32, n)) }, - }).out; - } - }; - const x: Tensor = .init(.{n}, .f32); - var exe = try zml.module.compile(allocator, std.testing.io, Mod.forward, .{ x, Tensor.init(.{n}, .f32) }, platform, .{}); - defer exe.deinit(); - - var hx: [n]f32 = undefined; - var hy: [n]f32 = undefined; - for (&hx, &hy, 0..) |*a, *b, i| { - a.* = @floatFromInt(i); - b.* = 1; - } - var bx: zml.Buffer = try .fromBytes(std.testing.io, platform, x.shape(), std.mem.sliceAsBytes(&hx)); - defer bx.deinit(); - var by: zml.Buffer = try .fromBytes(std.testing.io, platform, x.shape(), std.mem.sliceAsBytes(&hy)); - defer by.deinit(); - var result = try exe.eval(allocator, std.testing.io, .{ bx, by }); - defer result.deinit(); - var host = try result.toSliceAlloc(allocator, std.testing.io); - defer host.free(allocator); - for (host.items(f32), hx, hy) |out, a, b| try std.testing.expectEqual(a * 0.5 + b, out); -} diff --git a/zml/kernel/BUILD.bazel b/zml/kernel/BUILD.bazel new file mode 100644 index 00000000..61ab44f8 --- /dev/null +++ b/zml/kernel/BUILD.bazel @@ -0,0 +1,15 @@ +package(default_visibility = ["//zml:__pkg__"]) + +filegroup( + name = "kernel", + srcs = [ + "kernel.zig", + "//zml/kernel/cuda_tile", + "//zml/kernel/cute", + "//zml/kernel/fly", + "//zml/kernel/metal", + "//zml/kernel/mosaic_tpu", + "//zml/kernel/tcl", + "//zml/kernel/triton", + ], +) diff --git a/zml/kernel/cuda_tile/BUILD.bazel b/zml/kernel/cuda_tile/BUILD.bazel new file mode 100644 index 00000000..89b7977e --- /dev/null +++ b/zml/kernel/cuda_tile/BUILD.bazel @@ -0,0 +1,9 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "cuda_tile", + srcs = [ + "cuda_tile.zig", + "test.zig", + ], +) diff --git a/zml/kernel/cuda_tile/cuda_tile.zig b/zml/kernel/cuda_tile/cuda_tile.zig new file mode 100644 index 00000000..3f6d8734 --- /dev/null +++ b/zml/kernel/cuda_tile/cuda_tile.zig @@ -0,0 +1,163 @@ +const std = @import("std"); + +const cuda_tile_builder = @import("dsl/cuda_tile/builder"); +pub const Builder = cuda_tile_builder.Builder; +pub const Value = cuda_tile_builder.Value; +pub const DType = cuda_tile_builder.DType; +pub const FinishError = cuda_tile_builder.FinishError; +pub const EntryHint = cuda_tile_builder.EntryHint; +pub const Arch = cuda_tile_builder.Arch; +const func = @import("mlir/dialects/func"); +const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const mlirx = @import("../../mlirx.zig"); +const ops = @import("../../ops.zig"); +const platform_ = @import("../../platform.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +test { + _ = @import("test.zig"); +} + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} + +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(cuda_tile_builder.registerDialects); +} + +pub fn from(dt: DataType) DType { + return switch (dt) { + .bool => .i1, + .i4, .u4 => .i4, + .i8, .u8 => .i8, + .i16, .u16 => .i16, + .i32, .u32 => .i32, + .i64, .u64 => .i64, + .f16 => .f16, + .bf16 => .bf16, + .f32 => .f32, + .f64 => .f64, + .f8e4m3fn => .f8e4m3fn, + .f8e5m2 => .f8e5m2, + .f8e8m0 => .f8e8m0fnu, + .f4e2m1 => .f4e2m1fn, + else => std.debug.panic("zml.kernel.cuda_tile.from: dtype {s} has no CUDA Tile IR equivalent", .{@tagName(dt)}), + }; +} + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + + /// No `num_warps`/`num_stages`: warp and CTA tuning is + /// `optimization_hints` inside the IR (`Opts.hints`). + pub const CallOpts = struct { + cfg: ConfigT, + grid: [3]i32, + /// Outputs XLA zeroes before the launch, by position in + /// `spec.outputs`, ascending. + zeroed_outputs: []const i32 = &.{}, + output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try cuda_tile_builder.Builder.open(allocator, ctx, name); + defer b.deinit(); + + try spec.run(&b, cfg); + + return b.finish(&.{}); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + + const ir = emit(cur.allocator, opts.cfg) catch |err| + std.debug.panic("zml.kernel.cuda_tile.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ir); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); + + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); + + const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); + + const tensor_results = ops.cudaTile(inputs_arr, outputs_arr, .{ + .name = name, + .ir = ir, + .grid = opts.grid, + .zeroed_outputs = opts.zeroed_outputs, + .output_operand_aliases = aliases.constSlice(), + }); + + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; + return results; + } + }; +} diff --git a/zml/kernel/cuda_tile/test.zig b/zml/kernel/cuda_tile/test.zig new file mode 100644 index 00000000..c929b0b9 --- /dev/null +++ b/zml/kernel/cuda_tile/test.zig @@ -0,0 +1,28 @@ +const std = @import("std"); + +const cuda_tile = @import("cuda_tile.zig"); + +test "cuda_tile kernel emits a module XLA can parse" { + const Cfg = struct { n: i64 }; + const AddOne = cuda_tile.Kernel(Cfg, .{ + .name = "add_one", + .inputs = &.{"x"}, + .outputs = &.{"out"}, + .run = struct { + fn run(b: *cuda_tile.Builder, cfg: Cfg) cuda_tile.FinishError!void { + const a = try b.declareArgs(.{ .x = .{ .ptr = .f32 }, .out = .{ .ptr = .f32 } }); + const vx = b.partitionView(b.tensorView(a.x, &.{cfg.n}, &.{1}), &.{128}, .{}); + const vo = b.partitionView(b.tensorView(a.out, &.{cfg.n}, &.{1}), &.{128}, .{}); + const bid = b.tileBlockId(); + _ = b.store(b.load(vx, &.{bid.x}).add(1.0), vo, &.{bid.x}); + } + }.run, + }); + + const ir = try AddOne.emit(std.testing.allocator, .{ .n = 1024 }); + defer std.testing.allocator.free(ir); + try std.testing.expect(std.mem.indexOf(u8, ir, "cuda_tile.module @add_one") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "entry @add_one(") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "load_view_tko") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "store_view_tko") != null); +} diff --git a/zml/kernel/cute/BUILD.bazel b/zml/kernel/cute/BUILD.bazel new file mode 100644 index 00000000..7a613072 --- /dev/null +++ b/zml/kernel/cute/BUILD.bazel @@ -0,0 +1,9 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "cute", + srcs = [ + "cute.zig", + "test.zig", + ], +) diff --git a/zml/kernel/cute/cute.zig b/zml/kernel/cute/cute.zig new file mode 100644 index 00000000..2e8c919f --- /dev/null +++ b/zml/kernel/cute/cute.zig @@ -0,0 +1,264 @@ +const std = @import("std"); + +const cute_builder = @import("dsl/cute/builder"); +pub const AlgebraExpr = cute_builder.AlgebraExpr; +pub const Builder = cute_builder.Builder; +pub const ArgSpec = cute_builder.ArgSpec; +pub const Value = cute_builder.Value; +pub const View = cute_builder.View; +pub const Atom = cute_builder.Atom; +pub const DType = cute_builder.DType; +pub const AlgebraToken = cute_builder.AlgebraToken; +pub const LayoutSpec = cute_builder.LayoutSpec; +pub const ConstrainedDynamic = cute_builder.ConstrainedDynamic; +pub const FinishError = cute_builder.FinishError; +pub const TmaConfig = cute_builder.TmaConfig; +pub const TmaFormat = cute_builder.TmaFormat; +pub const BlockScaledMmaConfig = cute_builder.BlockScaledMmaConfig; +pub const CtaGroup = cute_builder.CtaGroup; +pub const ctaCount = cute_builder.ctaCount; +pub const MBarrierScope = cute_builder.MBarrierScope; +pub const OperandMajorMode = cute_builder.OperandMajorMode; +pub const Tcgen05SmemSwizzle = cute_builder.Tcgen05SmemSwizzle; +pub const Tcgen05SmemDesc = cute_builder.Tcgen05SmemDesc; +pub const Tcgen05InstrDesc = cute_builder.Tcgen05InstrDesc; +pub const CopyBulkTensorTileG2SOp = cute_builder.CopyBulkTensorTileG2SOp; +pub const CopyBulkTensor2DGather4G2SOp = cute_builder.CopyBulkTensor2DGather4G2SOp; +pub const CopyBulkTensorTileS2GOp = cute_builder.CopyBulkTensorTileS2GOp; +const func = @import("mlir/dialects/func"); +const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const mlirx = @import("../../mlirx.zig"); +const ops = @import("../../ops.zig"); +const platform_ = @import("../../platform.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +test { + _ = @import("test.zig"); +} + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(cute_builder.registerDialects); +} + +pub fn from(dt: DataType) DType { + return switch (dt) { + .bool => .i1, + .i8, .u8 => .i8, + .i16, .u16 => .i16, + .i32, .u32 => .i32, + .i64, .u64 => .i64, + .f16 => .f16, + .bf16 => .bf16, + .f32 => .f32, + .f64 => .f64, + .f4e2m1 => .f4e2m1fn, + .f8e4m3fn => .f8e4m3fn, + .f8e5m2 => .f8e5m2, + .f8e8m0 => .f8e8m0fnu, + else => std.debug.panic("zml.kernel.cute.from: dtype {s} has no CuTe equivalent", .{@tagName(dt)}), + }; +} + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + + /// `kernel.launch(grid=, block=)` in the Python DSL; dynamic shared + /// memory is what the kernel declares. + pub const CallOpts = struct { + cfg: ConfigT, + grid: [3]i32, + block: [3]i32, + zeroed_outputs: []const i32 = &.{}, + output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT, block: [3]i32) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try cute_builder.Builder.open(allocator, ctx, name); + defer b.deinit(); + + try spec.run(&b, cfg); + + // XLA passes one pointer per input then per output; it does + // not check the count against the kernel. + const expected = spec.inputs.len + spec.outputs.len; + if (b.args.len != expected) { + std.debug.panic("zml.kernel.cute.Kernel({s}): declareArgs declared {d} arguments, inputs + outputs are {d}", .{ name, b.args.len, expected }); + } + + return b.finish(block); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + + const ir = emit(cur.allocator, opts.cfg, opts.block) catch |err| + std.debug.panic("zml.kernel.cute.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ir); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); + + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); + + const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); + + const tensor_results = ops.cute(inputs_arr, outputs_arr, .{ + .name = name, + .ir = ir, + .grid = opts.grid, + .block = opts.block, + .zeroed_outputs = opts.zeroed_outputs, + .output_operand_aliases = aliases.constSlice(), + }); + + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; + return results; + } + }; +} + +fn ProgramSpec(comptime Config: type) type { + return struct { + /// Public host entry selected by XLA. The callback may give the + /// nested CUDA kernel a different (private) symbol. + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +/// A complete CuTe program: one or more `cuda.kernel` operations inside +/// a `gpu.module`, followed by a public `func.func` which constructs the +/// launch-time CuTe objects and calls `cuda.launch_ex`. +/// +/// This is the source-level equivalent of invoking a Python `@cute.jit` +/// object whose `__call__` builds TMA descriptors and launches an +/// `@cute.kernel`. XLA executes the generated host function on its CUDA +/// stream, so the custom call deliberately carries no external grid or +/// block dimensions. +pub fn Program( + comptime ConfigT: type, + comptime spec: ProgramSpec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + + pub const CallOpts = struct { + cfg: ConfigT, + zeroed_outputs: []const i32 = &.{}, + scalars: []const i64 = &.{}, + symmetric_workspace: ?ops.CuteOps.SymmetricWorkspace = null, + output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try cute_builder.Builder.openProgram(allocator, ctx, name); + defer b.deinit(); + try spec.run(&b, cfg); + return b.finishProgram(); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + const ir = emit(cur.allocator, opts.cfg) catch |err| + std.debug.panic("zml.kernel.cute.Program({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ir); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); + + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); + + const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); + const tensor_results = ops.cute(inputs_arr, outputs_arr, .{ + .name = name, + .ir = ir, + .zeroed_outputs = opts.zeroed_outputs, + .scalars = opts.scalars, + .symmetric_workspace = opts.symmetric_workspace, + .output_operand_aliases = aliases.constSlice(), + }); + + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; + return results; + } + }; +} diff --git a/zml/kernel/cute/test.zig b/zml/kernel/cute/test.zig new file mode 100644 index 00000000..f4d457a3 --- /dev/null +++ b/zml/kernel/cute/test.zig @@ -0,0 +1,34 @@ +const std = @import("std"); + +const cute = @import("cute.zig"); + +test "cute kernel emits the module the CuTe compiler takes" { + const Cfg = struct { n: i64 }; + const AddOne = cute.Kernel(Cfg, .{ + .name = "add_one", + .inputs = &.{"x"}, + .outputs = &.{"out"}, + .run = struct { + fn run(b: *cute.Builder, cfg: Cfg) cute.FinishError!void { + const a = try b.declareArgs(.{ + .x = .{ .tensor = .{ .dtype = .f32, .shape = &.{cfg.n} } }, + .out = .{ .tensor = .{ .dtype = .f32, .shape = &.{cfg.n} } }, + }); + const i = b.blockIdx().x.mul(b.blockDim().x).add(b.threadIdx().x); + var guard = b.openIf(i.lt(cfg.n)); + { + a.out.set(.{i}, a.x.get(.{i}).add(1.0)); + guard.close(); + } + } + }.run, + }); + + const ir = try AddOne.emit(std.testing.allocator, .{ .n = 1000 }, .{ 128, 1, 1 }); + defer std.testing.allocator.free(ir); + try std.testing.expect(std.mem.indexOf(u8, ir, "\"func.func\"() <{function_type = (") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "sym_name = \"add_one\"") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "nvvm.reqntid = array") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "cute.memref.load") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "gpu.module") == null); +} diff --git a/zml/kernel/fly/BUILD.bazel b/zml/kernel/fly/BUILD.bazel new file mode 100644 index 00000000..57203521 --- /dev/null +++ b/zml/kernel/fly/BUILD.bazel @@ -0,0 +1,9 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "fly", + srcs = [ + "fly.zig", + "test.zig", + ], +) diff --git a/zml/kernel/fly/fly.zig b/zml/kernel/fly/fly.zig new file mode 100644 index 00000000..b7d6c943 --- /dev/null +++ b/zml/kernel/fly/fly.zig @@ -0,0 +1,248 @@ +const std = @import("std"); + +const fly_builder = @import("dsl/fly/builder"); +pub const Builder = fly_builder.Builder; +pub const Value = fly_builder.Value; +pub const DType = fly_builder.DType; +pub const FinishError = fly_builder.FinishError; +pub const TiledCopy = fly_builder.TiledCopy; +pub const TiledMma = fly_builder.TiledMma; +pub const Arch = fly_builder.Arch; +pub const MmaFlavor = fly_builder.MmaFlavor; +pub const fly_rocdl = fly_builder.fly_rocdl; +pub const layout = fly_builder.layout; +pub const Layout = fly_builder.Layout; +pub const Tile = fly_builder.Tile; +pub const IntTuple = fly_builder.IntTuple; +pub const L = fly_builder.L; +pub const rowMajor = fly_builder.rowMajor; +pub const colMajor = fly_builder.colMajor; +pub const ordered = fly_builder.ordered; +pub const tile = fly_builder.tile; +pub const it = fly_builder.it; +const func = @import("mlir/dialects/func"); +const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const mlirx = @import("../../mlirx.zig"); +const ops = @import("../../ops.zig"); +const platform_ = @import("../../platform.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +test { + _ = @import("test.zig"); +} + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} + +/// Threads per wavefront: 64 on CDNA, 32 on RDNA. +pub fn waveSize(p: *const platform_.Platform) i32 { + const cc = platform_.rocm.computeCapability(p) orelse return 64; + return switch (cc.architecture()) { + .cdna1, .cdna2, .cdna3, .cdna4 => 64, + .rdna2, .rdna3, .rdna3_5, .rdna4 => 32, + }; +} + +/// The plugin sizes a block in wavefronts, so a thread count has to divide +/// evenly: a kernel's layouts are built for an exact number of threads and +/// spare ones would partition out of range. +fn warpsFor(p: *const platform_.Platform, threads: i32) i32 { + const wave = waveSize(p); + if (@rem(threads, wave) != 0) { + std.debug.panic("zml.kernel.fly: {d} threads is not a multiple of the {d}-wide wavefront", .{ threads, wave }); + } + return @divExact(threads, wave); +} + +/// The matrix atom this device provides, or null when it has none: WMMA +/// arrived with RDNA3, so gfx1030 has no matrix instruction at all. +pub fn mmaFlavor(p: *const platform_.Platform) ?MmaFlavor { + const cc = platform_.rocm.computeCapability(p) orelse return .cdna3_mfma; + return switch (cc.architecture()) { + .cdna1, .cdna2, .cdna3, .cdna4 => .cdna3_mfma, + .rdna3, .rdna3_5 => .gfx11_wmma, + .rdna4 => .gfx120x_wmma, + .rdna2 => null, + }; +} +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(fly_builder.registerDialects); +} + +pub fn from(dt: DataType) DType { + return switch (dt) { + .bool => .i1, + .i8, .u8 => .i8, + .i16, .u16 => .i16, + .i32, .u32 => .i32, + .i64, .u64 => .i64, + .f16 => .f16, + .bf16 => .bf16, + .f32 => .f32, + .f64 => .f64, + .f8e4m3fn => .f8e4m3fn, + .f8e4m3fnuz => .f8e4m3fnuz, + .f8e5m2 => .f8e5m2, + .f8e5m2fnuz => .f8e5m2fnuz, + else => std.debug.panic("zml.kernel.fly.from: dtype {s} has no Fly equivalent", .{@tagName(dt)}), + }; +} + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + pub const arg_names = spec.inputs ++ spec.outputs; + /// By name, as `!fly.memref` views. + pub const Args = StructOf(arg_names, Value); + + pub const CallOpts = struct { + cfg: ConfigT, + grid: [3]i32, + /// Threads per block. Converted to the wavefronts the plugin + /// wants using the device's wave size, which is 64 on CDNA and + /// 32 on RDNA, so a kernel states the thread count its layouts + /// were built for and runs on both. + threads: i32, + /// > 0 sets `amdgpu-waves-per-eu = "N, N"`: a hard occupancy clamp. + waves_per_eu: i32 = 0, + /// Required when the body uses `fly.get_dyn_shared`. + shared_mem_bytes: i64 = 0, + /// Inputs first, then outputs; see `zeroedOutput`. A no-op + /// inside command buffers. + zeroed_args: []const i32 = &.{}, + output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, + }; + + /// The arguments `run` was declared with. + pub fn args(b: *Builder) Args { + var out: Args = undefined; + inline for (arg_names, 0..) |fname, i| @field(out, fname) = b.arg(i); + return out; + } + + /// `zeroed_args` index of an output. + pub fn zeroedOutput(comptime f: std.meta.FieldEnum(Outputs)) i32 { + return @intCast(spec.inputs.len + @backingInt(f)); + } + + /// `shapes` is inputs, then outputs. + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT, shapes: []const Shape) ![:0]const u8 { + std.debug.assert(shapes.len == arg_names.len); + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try fly_builder.Builder.open(allocator, ctx, name); + defer b.deinit(); + + var specs: [arg_names.len]Builder.ArgSpec = undefined; + inline for (arg_names, 0..) |fname, i| { + const shape = shapes[i]; + specs[i] = .{ + .name = fname, + .dtype = from(shape.dtype()), + .dims = if (shape.rank() > 0) shape.dims() else null, + }; + } + try b.declareArgsLow(&specs); + + try spec.run(&b, cfg); + + return b.finish(); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); + + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); + + var shapes: [arg_names.len]Shape = undefined; + inline for (0..spec.inputs.len) |i| shapes[i] = inputs_arr[i].shape(); + inline for (0..spec.outputs.len) |i| shapes[spec.inputs.len + i] = outputs_arr[i]; + + const ir = emit(cur.allocator, opts.cfg, &shapes) catch |err| + std.debug.panic("zml.kernel.fly.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ir); + + const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); + + const tensor_results = ops.fly(inputs_arr, outputs_arr, .{ + .name = name, + .ir = ir, + .grid = opts.grid, + .num_warps = warpsFor(cur.platform, opts.threads), + .waves_per_eu = opts.waves_per_eu, + .shared_mem_bytes = opts.shared_mem_bytes, + .zeroed_args = opts.zeroed_args, + .output_operand_aliases = aliases.constSlice(), + }); + + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = tensor_results[i]; + return results; + } + }; +} diff --git a/zml/kernel/fly/test.zig b/zml/kernel/fly/test.zig new file mode 100644 index 00000000..ae788d96 --- /dev/null +++ b/zml/kernel/fly/test.zig @@ -0,0 +1,149 @@ +const std = @import("std"); + +const fly_builder = @import("dsl/fly/builder"); + +const fly = @import("fly.zig"); +const zml = @import("../../zml.zig"); + +/// C = A + B over (M, N) f32, 128-bit vectorized and predicated on the +/// ragged edge. +const FlyVectorAdd = struct { + const Cfg = struct { + thr: fly.Layout = fly.ordered(.{ 8, 16 }, .{ 1, 0 }), + val: fly.Layout = fly.ordered(.{ 1, 4 }, .{ 0, 1 }), + }; + + const K = fly.Kernel(Cfg, .{ + .name = "vector_add", + .inputs = &.{ "a", "b" }, + .outputs = &.{"c"}, + .run = run, + }); + + fn run(b: *fly.Builder, cfg: Cfg) fly.FinishError!void { + const a = K.args(b); + fly_builder.emitVectorAdd(b, cfg.thr, cfg.val, a.a, a.b, a.c); + } +}; + +/// C = A @ B^T with one MFMA-tiled block. +const FlyTiledMma = struct { + const Cfg = struct { flavor: fly.MmaFlavor }; + const K = fly.Kernel(Cfg, .{ + .name = "tiled_mma", + .inputs = &.{ "a", "b" }, + .outputs = &.{"c"}, + .run = run, + }); + + fn run(b: *fly.Builder, cfg: Cfg) fly.FinishError!void { + const a = K.args(b); + fly_builder.emitTiledMma(b, cfg.flavor, a.a, a.b, a.c); + } +}; + +test "fly kernel emits a module XLA can parse" { + const shapes = [_]zml.Shape{ .init(.{ 100, 1000 }, .f32), .init(.{ 100, 1000 }, .f32), .init(.{ 100, 1000 }, .f32) }; + const ir = try FlyVectorAdd.K.emit(std.testing.allocator, .{}, &shapes); + defer std.testing.allocator.free(ir); + try std.testing.expect(std.mem.indexOf(u8, ir, "gpu.func @vector_add(%arg0: !fly.ptr, %arg1: !fly.ptr, %arg2: !fly.ptr) kernel") != null); + try std.testing.expect(std.mem.indexOf(u8, ir, "fly.static : !fly.tile<[8|64]>") != null); +} + +test "fly kernels run on rocm" { + const platform = zml.testing.env(); + if (platform.target != .rocm) return error.SkipZigTest; + + const allocator = std.testing.allocator; + + // Compile `forward`, upload both inputs, run, and bring the result back. + const call = struct { + fn f(comptime forward: anytype, p: *const zml.Platform, ta: zml.Tensor, ha: anytype, tb: zml.Tensor, hb: anytype) !zml.Slice { + var exe = try zml.module.compile(std.testing.allocator, std.testing.io, forward, .{ ta, tb }, p, .{}); + defer exe.deinit(); + var ba: zml.Buffer = try .fromBytes(std.testing.io, p, ta.shape(), std.mem.sliceAsBytes(ha)); + defer ba.deinit(); + var bb: zml.Buffer = try .fromBytes(std.testing.io, p, tb.shape(), std.mem.sliceAsBytes(hb)); + defer bb.deinit(); + var out = try exe.eval(std.testing.allocator, std.testing.io, .{ ba, bb }); + defer out.deinit(); + return out.toSliceAlloc(std.testing.allocator, std.testing.io); + } + }.f; + + // vector add: elementwise over a ragged (100, 1000), predicated at the edge. + { + const M, const N = .{ 100, 1000 }; + const Mod = struct { + pub fn forward(a: zml.Tensor, b: zml.Tensor) zml.Tensor { + return FlyVectorAdd.K.call(.{ .a = a, .b = b }, .{ .c = a.shape() }, .{ + .cfg = .{}, + .grid = .{ (M + 8 - 1) / 8, (N + 64 - 1) / 64, 1 }, + .threads = 8 * 16, + }).c; + } + }; + + const ha = try allocator.alloc(f32, M * N); + defer allocator.free(ha); + const hb = try allocator.alloc(f32, M * N); + defer allocator.free(hb); + for (ha, hb, 0..) |*va, *vb, i| { + va.* = @floatFromInt(i % 977); + vb.* = @floatFromInt(1000 + (i % 313)); + } + + var host = try call(Mod.forward, platform, .init(.{ M, N }, .f32), ha, .init(.{ M, N }, .f32), hb); + defer host.free(allocator); + for (host.items(f32), 0..) |v, i| try std.testing.expectEqual(ha[i] + hb[i], v); + } + + // tiled mma: C = A @ B^T in one block of 256 threads, over whatever matrix + // atom this device has. CDNA multiplies f32 through MFMA; RDNA3+ has WMMA, + // whose verifier rejects f32 operands, so the operand type and K come from + // the flavour rather than being written into the test. + if (fly.mmaFlavor(platform)) |flavor| switch (flavor) { + inline else => |f| { + const Elem = switch (comptime f.operandDType()) { + .f32 => f32, + .f16 => f16, + else => @compileError("unhandled MMA operand type"), + }; + const M, const N = .{ 64, 64 }; + const Kd = comptime f.blockK(); + const operand_dtype: zml.DataType = switch (comptime f.operandDType()) { + .f32 => .f32, + .f16 => .f16, + else => unreachable, + }; + + const Mod = struct { + pub fn forward(a: zml.Tensor, b: zml.Tensor) zml.Tensor { + return FlyTiledMma.K.call(.{ .a = a, .b = b }, .{ .c = zml.Shape.init(.{ M, N }, .f32) }, .{ + .cfg = .{ .flavor = f }, + .grid = .{ 1, 1, 1 }, + .threads = comptime f.blockThreads(), + }).c; + } + }; + + var ha: [M * Kd]Elem = undefined; + var hb: [N * Kd]Elem = undefined; + for (&ha, 0..) |*v, i| v.* = @as(Elem, @floatFromInt((i * 7) % 13)) - 6; + for (&hb, 0..) |*v, i| v.* = @as(Elem, @floatFromInt((i * 5) % 11)) - 5; + + var host = try call(Mod.forward, platform, .init(.{ M, Kd }, operand_dtype), &ha, .init(.{ N, Kd }, operand_dtype), &hb); + defer host.free(allocator); + + // f16 operands accumulate in f32, so the error is set by the input + // rounding rather than the sum. + const tol: f32 = if (Elem == f16) 1e-2 else 1e-3; + const out = host.items(f32); + for (0..M) |m| for (0..N) |n| { + var acc: f32 = 0; + for (0..Kd) |k| acc += @as(f32, ha[m * Kd + k]) * @as(f32, hb[n * Kd + k]); + try std.testing.expectApproxEqAbs(acc, out[m * N + n], tol); + }; + }, + }; +} diff --git a/zml/kernel/kernel.zig b/zml/kernel/kernel.zig new file mode 100644 index 00000000..204c6204 --- /dev/null +++ b/zml/kernel/kernel.zig @@ -0,0 +1,12 @@ +pub const cuda_tile = @import("cuda_tile/cuda_tile.zig"); +pub const cute = @import("cute/cute.zig"); +pub const fly = @import("fly/fly.zig"); +pub const metal = @import("metal/metal.zig"); +pub const mosaic_tpu = @import("mosaic_tpu/mosaic_tpu.zig"); +pub const tcl = @import("tcl/tcl.zig"); +pub const triton = @import("triton/triton.zig"); + +test { + const std = @import("std"); + std.testing.refAllDecls(@This()); +} diff --git a/zml/kernel/metal/BUILD.bazel b/zml/kernel/metal/BUILD.bazel new file mode 100644 index 00000000..4007beb4 --- /dev/null +++ b/zml/kernel/metal/BUILD.bazel @@ -0,0 +1,9 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "metal", + srcs = [ + "metal.zig", + "test.zig", + ], +) diff --git a/zml/kernel/metal/metal.zig b/zml/kernel/metal/metal.zig new file mode 100644 index 00000000..9626a1de --- /dev/null +++ b/zml/kernel/metal/metal.zig @@ -0,0 +1,237 @@ +const std = @import("std"); + +const func = @import("mlir/dialects/func"); +const metal_builder = @import("dsl/metal/builder"); +pub const Builder = metal_builder.Builder; +pub const Value = metal_builder.Value; +pub const Type = metal_builder.Type; +pub const Struct = metal_builder.Struct; +pub const ArgSpec = metal_builder.ArgSpec; +pub const LoopOpts = metal_builder.LoopOpts; +pub const FinishError = metal_builder.FinishError; +const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const mlirx = @import("../../mlirx.zig"); +const ops = @import("../../ops.zig"); +const platform_ = @import("../../platform.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +test { + _ = @import("test.zig"); +} + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} + +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(metal_builder.registerDialects); +} + +pub fn from(dt: DataType) Type { + return switch (dt) { + .bool => .bool, + .i8 => .char, + .i16 => .short, + .i32 => .int, + .i64 => .long, + .u8 => .uchar, + .u16 => .ushort, + .u32 => .uint, + .u64 => .ulong, + .f16 => .half, + .bf16 => .bfloat, + .f32 => .float, + else => std.debug.panic("zml.kernel.metal.from: dtype {s} has no MSL equivalent", .{@tagName(dt)}), + }; +} + +/// The bytes of a `constant T &` argument: `.of(@as(i32, n))`, or a host +/// `extern struct` laid out like a `constant S &` argument. +pub const Constant = struct { + buffer: [256]u8 = undefined, + len: usize = 0, + + pub fn of(value: anytype) Constant { + const raw = std.mem.asBytes(&value); + var c: Constant = .{ .len = raw.len }; + @memcpy(c.buffer[0..raw.len], raw); + return c; + } + + pub fn bytes(self: *const Constant) []const u8 { + return self.buffer[0..self.len]; + } +}; + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + /// The arguments bound to input tensors. + inputs: []const [:0]const u8, + /// The arguments bound to output tensors; one named like an + /// input updates it in place. + outputs: []const [:0]const u8, + /// The `constant` arguments given by value at the call. + constants: []const [:0]const u8 = &.{}, + run: *const fn (*Builder, Config) FinishError!void, + fast_math: bool = true, + }; +} + +/// A Metal kernel: `run` declares its arguments with `declareArgs`, by +/// name, and builds its body. +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + pub const Constants = StructOf(spec.constants, Constant); + + pub const CallOpts = struct { + cfg: ConfigT, + /// Threadgroups per grid. + grid: [3]u32, + /// Threads per threadgroup. + threadgroup: [3]u32, + /// Bytes of `[[threadgroup(0)]]` memory. + threadgroup_memory: u64 = 0, + constants: Constants = std.mem.zeroes(Constants), + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { + var arguments: std.ArrayList(metal_builder.Argument) = .empty; + defer arguments.deinit(allocator); + return build(allocator, cfg, &arguments); + } + + fn build(allocator: std.mem.Allocator, cfg: ConfigT, arguments: *std.ArrayList(metal_builder.Argument)) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + var b = Builder.open(allocator, ctx, name, .{ .fast_math = spec.fast_math }); + defer b.deinit(); + try spec.run(&b, cfg); + const ir = try b.finish(); + errdefer allocator.free(ir); + try arguments.appendSlice(allocator, b.arguments.items); + return ir; + } + + fn indexOf(names: []const [:0]const u8, arg: []const u8) ?usize { + for (names, 0..) |n, i| if (std.mem.eql(u8, n, arg)) return i; + return null; + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + const allocator = cur.allocator; + + var arguments: std.ArrayList(metal_builder.Argument) = .empty; + defer arguments.deinit(allocator); + const ir = build(allocator, opts.cfg, &arguments) catch |err| + std.debug.panic("zml.kernel.metal.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer allocator.free(ir); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + inline for (spec.inputs, 0..) |field, i| inputs_arr[i] = @field(inputs, field); + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |field, i| outputs_arr[i] = @field(outputs, field); + var constants: [spec.constants.len][]const u8 = undefined; + inline for (spec.constants, 0..) |field, i| constants[i] = @field(opts.constants, field).bytes(); + + var bound: [spec.outputs.len]bool = @splat(false); + var bound_inputs: [spec.inputs.len]bool = @splat(false); + var bound_constants: [spec.constants.len]bool = @splat(false); + var aliases: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, spec.outputs.len) = .empty; + var bindings: std.ArrayList(i32) = .empty; + defer bindings.deinit(allocator); + for (arguments.items) |arg| { + if (arg.kind == .threadgroup) { + // `threadgroup_memory` sizes `[[threadgroup(0)]]` alone. + if (arg.index != 0) std.debug.panic("zml.kernel.metal.Kernel({s}): threadgroup argument {s} is [[threadgroup({d})]], and only [[threadgroup(0)]] gets memory", .{ name, arg.name, arg.index }); + continue; + } + std.debug.assert(arg.index == bindings.items.len); + const input = indexOf(spec.inputs, arg.name); + const output = indexOf(spec.outputs, arg.name); + if (input) |i| bound_inputs[i] = true; + if (output) |o| bound[o] = true; + const binding: i32 = if (input) |i| blk: { + if (output) |o| aliases.appendAssumeCapacity(.{ .output_index = @intCast(o), .operand_index = @intCast(i) }); + break :blk @intCast(i); + } else if (output) |o| + @intCast(spec.inputs.len + o) + else if (indexOf(spec.constants, arg.name)) |k| blk: { + if (constants[k].len != arg.size) std.debug.panic("zml.kernel.metal.Kernel({s}): constant {s} has {d} bytes, the argument {d}", .{ name, arg.name, constants[k].len, arg.size }); + bound_constants[k] = true; + break :blk -1 - @as(i32, @intCast(k)); + } else std.debug.panic("zml.kernel.metal.Kernel({s}): argument {s} is no input, output or constant", .{ name, arg.name }); + bindings.append(allocator, binding) catch @panic("OOM"); + } + for (bound, spec.outputs) |b, output| if (!b) std.debug.panic("zml.kernel.metal.Kernel({s}): output {s} is no argument", .{ name, output }); + for (bound_inputs, spec.inputs) |b, input| if (!b) std.debug.panic("zml.kernel.metal.Kernel({s}): input {s} is no argument", .{ name, input }); + for (bound_constants, spec.constants) |b, constant| if (!b) std.debug.panic("zml.kernel.metal.Kernel({s}): constant {s} is no argument", .{ name, constant }); + + const tensor_results = ops.metal(inputs_arr, outputs_arr, .{ + .name = name, + .ir = ir, + .grid = opts.grid, + .threadgroup = opts.threadgroup, + .threadgroup_memory = opts.threadgroup_memory, + .arguments = bindings.items, + .constants = &constants, + .output_operand_aliases = aliases.constSlice(), + }); + var results: Results = undefined; + inline for (spec.outputs, 0..) |field, i| @field(results, field) = tensor_results[i]; + return results; + } + }; +} diff --git a/zml/kernel/metal/test.zig b/zml/kernel/metal/test.zig new file mode 100644 index 00000000..0a342af0 --- /dev/null +++ b/zml/kernel/metal/test.zig @@ -0,0 +1,71 @@ +const std = @import("std"); + +const metal = @import("metal.zig"); +const zml = @import("../../zml.zig"); + +/// `out[i] = a[i] * scale + b[i]` over `n` floats. +const MetalAxpy = struct { + const Cfg = struct {}; + const K = metal.Kernel(Cfg, .{ + .name = "axpy", + .inputs = &.{ "a", "b" }, + .outputs = &.{"out"}, + .constants = &.{ "scale", "n" }, + .run = run, + }); + + fn run(b: *metal.Builder, _: Cfg) metal.FinishError!void { + const a = try b.declareArgs(.{ + .a = .{ .buffer = .{ .type = .float, .const_ = true } }, + .b = .{ .buffer = .{ .type = .float, .const_ = true } }, + .out = .{ .buffer = .{ .type = .float } }, + .scale = .{ .constant = .float }, + .n = .{ .constant = .uint }, + .tid = .{ .thread_position_in_grid = .uint }, + }); + b.returnIf(a.tid.ge(a.n)); + a.out.store(a.tid, b.fma(a.a.load(a.tid), a.scale, a.b.load(a.tid))); + } +}; + +test "metal kernels run on metal" { + const allocator = std.testing.allocator; + + const ir = try MetalAxpy.K.emit(allocator, .{}); + defer allocator.free(ir); + try std.testing.expect(std.mem.indexOf(u8, ir, "air.fma") != null); + + const platform = zml.testing.env(); + if (platform.target != .metal) return error.SkipZigTest; + + const n = 1000; + const Mod = struct { + pub fn forward(x: zml.Tensor, y: zml.Tensor) zml.Tensor { + return MetalAxpy.K.call(.{ .a = x, .b = y }, .{ .out = x.shape() }, .{ + .cfg = .{}, + .grid = .{ (n + 255) / 256, 1, 1 }, + .threadgroup = .{ 256, 1, 1 }, + .constants = .{ .scale = .of(@as(f32, 0.5)), .n = .of(@as(u32, n)) }, + }).out; + } + }; + const x: zml.Tensor = .init(.{n}, .f32); + var exe = try zml.module.compile(allocator, std.testing.io, Mod.forward, .{ x, zml.Tensor.init(.{n}, .f32) }, platform, .{}); + defer exe.deinit(); + + var hx: [n]f32 = undefined; + var hy: [n]f32 = undefined; + for (&hx, &hy, 0..) |*a, *b, i| { + a.* = @floatFromInt(i); + b.* = 1; + } + var bx: zml.Buffer = try .fromBytes(std.testing.io, platform, x.shape(), std.mem.sliceAsBytes(&hx)); + defer bx.deinit(); + var by: zml.Buffer = try .fromBytes(std.testing.io, platform, x.shape(), std.mem.sliceAsBytes(&hy)); + defer by.deinit(); + var result = try exe.eval(allocator, std.testing.io, .{ bx, by }); + defer result.deinit(); + var host = try result.toSliceAlloc(allocator, std.testing.io); + defer host.free(allocator); + for (host.items(f32), hx, hy) |out, a, b| try std.testing.expectEqual(a * 0.5 + b, out); +} diff --git a/zml/kernel/mosaic_tpu/BUILD.bazel b/zml/kernel/mosaic_tpu/BUILD.bazel new file mode 100644 index 00000000..ba3b6b66 --- /dev/null +++ b/zml/kernel/mosaic_tpu/BUILD.bazel @@ -0,0 +1,6 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "mosaic_tpu", + srcs = ["mosaic_tpu.zig"], +) diff --git a/zml/kernel/mosaic_tpu/mosaic_tpu.zig b/zml/kernel/mosaic_tpu/mosaic_tpu.zig new file mode 100644 index 00000000..a09e8368 --- /dev/null +++ b/zml/kernel/mosaic_tpu/mosaic_tpu.zig @@ -0,0 +1,284 @@ +const std = @import("std"); + +const func = @import("mlir/dialects/func"); +const mlir = @import("mlir"); +const mosaic_tpu_builder = @import("dsl/mosaic_tpu/builder"); +const Builder = mosaic_tpu_builder.Builder; +const DType = mosaic_tpu_builder.DType; +const FinishError = mosaic_tpu_builder.FinishError; +const mosaic_tpu_dialect = @import("mlir/dialects/mosaic_tpu"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const mlirx = @import("../../mlirx.zig"); +const ops = @import("../../ops.zig"); +const platform_ = @import("../../platform.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} + +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(mosaic_tpu_builder.registerDialects); +} + +pub fn from(dt: DataType) DType { + return switch (dt) { + .bool => .i1, + .i8, .u8 => .i8, + .i16, .u16 => .i16, + .i32, .u32 => .i32, + .i64, .u64 => .i64, + .f16 => .f16, + .bf16 => .bf16, + .f32 => .f32, + .f64 => .f64, + .f8e4m3fn => .f8e4m3fn, + .f8e5m2 => .f8e5m2, + else => std.debug.panic("zml.kernel.mosaic_tpu.from: dtype {s} has no Mosaic-TPU equivalent", .{@tagName(dt)}), + }; +} + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + + pub const CallExtras = struct { + vmem_limit_bytes: ?i64 = null, + disable_bounds_checks: ?bool = null, + disable_semaphore_checks: ?bool = null, + has_communication: ?bool = null, + additional_attributes: []const mlir.NamedAttribute = &.{}, + output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, + dynamic_grid_bounds: []const Tensor = &.{}, + }; + + pub const CallOpts = struct { + cfg: ConfigT, + extras: CallExtras = .{}, + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try mosaic_tpu_builder.Builder.open(allocator, ctx, name); + defer b.deinit(); + + try spec.run(&b, cfg); + + return b.finishOpts(&.{}, .{ .canonicalize = true }); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + + const ir = emit(cur.allocator, opts.cfg) catch |err| + std.debug.panic("zml.kernel.mosaic_tpu.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ir); + + const backend_config = buildBackendConfig(cur.allocator, ir, .{ + .vmem_limit_bytes = opts.extras.vmem_limit_bytes, + .disable_bounds_checks = opts.extras.disable_bounds_checks, + .disable_semaphore_checks = opts.extras.disable_semaphore_checks, + .has_communication = opts.extras.has_communication, + }) catch |err| + std.debug.panic("zml.kernel.mosaic_tpu.Kernel({s}).call: backend_config build failed: {}", .{ name, err }); + defer cur.allocator.free(backend_config); + + var values: [spec.inputs.len]*const mlir.Value = undefined; + inline for (spec.inputs, 0..) |fname, i| values[i] = @field(inputs, fname).value(); + + var res_types: [spec.outputs.len]*const mlir.Type = undefined; + inline for (&res_types, spec.outputs) |*r, fname| { + r.* = mlirx.Type.rankedTensor(cur.mlir_ctx, @field(outputs, fname)); + } + + const aliases = resolveOutputOperandAliases(opts.extras.output_operand_aliases, @intCast(opts.extras.dynamic_grid_bounds.len)); + + const op = callTpuCustomCall(.{ + .inputs = &values, + .result_types = &res_types, + .backend_config = backend_config, + .aliases = aliases.constSlice(), + .dynamic_grid_bounds = opts.extras.dynamic_grid_bounds, + .additional_attributes = opts.extras.additional_attributes, + }); + + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| { + const out = @field(outputs, fname); + @field(results, fname) = Tensor._result(out, op.result(i)); + } + return results; + } + }; +} + +const CustomCallConfig = struct { + body: []const u8, + has_communication: ?bool = null, + serialization_format: i64 = 1, + needs_layout_passes: bool = true, + shape_invariant_numerics: bool = false, + disable_bounds_checks: ?bool = null, + disable_semaphore_checks: ?bool = null, +}; + +const ScopedMemoryConfig = struct { + memory_space: i64 = 1, + offset: i64 = 0, + size: i64, +}; + +const BackendConfig = struct { + custom_call_config: CustomCallConfig, + scoped_memory_configs: ?[]const ScopedMemoryConfig = null, +}; + +const BackendConfigOpts = struct { + vmem_limit_bytes: ?i64 = null, + disable_bounds_checks: ?bool = null, + disable_semaphore_checks: ?bool = null, + has_communication: ?bool = null, +}; + +fn buildBackendConfig( + allocator: std.mem.Allocator, + ir: [:0]const u8, + opts: BackendConfigOpts, +) ![]u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + const module = try mlir.Module.parse(ctx, ir); + defer module.deinit(); + + mosaic_tpu_dialect.registerMosaicSerdePass(); + const prev_allow_unregistered = ctx.allowUnregisteredDialects(); + ctx.setAllowUnregisteredDialects(true); + defer ctx.setAllowUnregisteredDialects(prev_allow_unregistered); + + const pm = mlir.PassManager.init(ctx); + defer pm.deinit(); + try pm.asOpPassManager().addPipeline("mosaic-serde{serialize=true}"); + try pm.runOnOp(module.operation()); + + const Encoder = std.base64.standard.Encoder; + const b64 = b64: { + var bytecode: std.Io.Writer.Allocating = .init(allocator); + defer bytecode.deinit(); + try module.operation().writeBytecode(.{ .desired_emit_version = 0 }, &bytecode.writer); + const buf = try allocator.alloc(u8, Encoder.calcSize(bytecode.written().len)); + break :b64 Encoder.encode(buf, bytecode.written()); + }; + defer allocator.free(b64); + + const config = BackendConfig{ + .custom_call_config = .{ + .body = b64, + .has_communication = opts.has_communication, + .disable_bounds_checks = opts.disable_bounds_checks, + .disable_semaphore_checks = opts.disable_semaphore_checks, + }, + .scoped_memory_configs = if (opts.vmem_limit_bytes) |vmem| + &.{.{ .size = vmem }} + else + null, + }; + + var json: std.Io.Writer.Allocating = .init(allocator); + errdefer json.deinit(); + + try json.writer.print("{f}", .{std.json.fmt(config, .{ .emit_null_optional_fields = false })}); + return try json.toOwnedSlice(); +} + +const TpuCustomCallArgs = struct { + inputs: []const *const mlir.Value, + result_types: []const *const mlir.Type, + backend_config: []const u8, + aliases: []const stablehlo.CustomCallOpts.OutputOperandAlias, + dynamic_grid_bounds: []const Tensor, + additional_attributes: []const mlir.NamedAttribute, +}; + +fn callTpuCustomCall(args: TpuCustomCallArgs) *mlir.Operation { + const cur = Compiler.current(); + + var all_inputs: stdx.BoundedArray(*const mlir.Value, stablehlo.CustomCallOpts.MAX_OPERANDS) = .empty; + for (args.dynamic_grid_bounds) |t| all_inputs.appendAssumeCapacity(t.value()); + all_inputs.appendSliceAssumeCapacity(args.inputs); + + return stablehlo.custom_call( + cur.mlir_ctx, + all_inputs.constSlice(), + args.result_types, + .{ + .call_target_name = "tpu_custom_call", + .backend_config = .{ .original = args.backend_config }, + .has_side_effect = false, + .api_version = .original, + .additional_attributes = args.additional_attributes, + .output_operand_aliases = args.aliases, + }, + cur.location, + ).appendTo(cur.currentScope().block); +} diff --git a/zml/kernel/tcl/BUILD.bazel b/zml/kernel/tcl/BUILD.bazel new file mode 100644 index 00000000..bc7a0aa2 --- /dev/null +++ b/zml/kernel/tcl/BUILD.bazel @@ -0,0 +1,9 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "tcl", + srcs = [ + "tcl.zig", + "test.zig", + ], +) diff --git a/zml/kernel/tcl/tcl.zig b/zml/kernel/tcl/tcl.zig new file mode 100644 index 00000000..023e862f --- /dev/null +++ b/zml/kernel/tcl/tcl.zig @@ -0,0 +1,206 @@ +const std = @import("std"); + +const func = @import("mlir/dialects/func"); +const metal_builder = @import("dsl/metal/builder"); +const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); +const tcl_builder = @import("dsl/tcl/builder"); +pub const Builder = tcl_builder.Builder; +pub const TensorOperation = tcl_builder.TensorOperation; +pub const Value = tcl_builder.Value; +pub const Axis = tcl_builder.Axis; +pub const Expr = tcl_builder.Expr; +pub const Dram = tcl_builder.Dram; +pub const Context = tcl_builder.Context; +pub const DType = tcl_builder.DType; +pub const VeOp = tcl_builder.VeOp; +pub const ReduceMode = tcl_builder.ReduceMode; +pub const Predicate = tcl_builder.Predicate; +pub const Tactic = tcl_builder.Tactic; +pub const FinishError = tcl_builder.FinishError; + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const mlirx = @import("../../mlirx.zig"); +const ops = @import("../../ops.zig"); +const platform_ = @import("../../platform.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +test { + _ = @import("test.zig"); +} + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} + +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(tcl_builder.registerDialects); +} + +pub fn from(dt: DataType) DType { + return switch (dt) { + .bool => .bool, + .i4 => .i4, + .i8 => .i8, + .i16 => .i16, + .i32 => .i32, + .i64 => .i64, + .u8 => .u8, + .u16 => .u16, + .u32 => .u32, + .u64 => .u64, + .f16 => .f16, + .bf16 => .bf16, + .f32 => .f32, + .f64 => .f64, + .f4e2m1 => .f4_e2, + .f8e4m3fn => .f8_e4, + .f8e5m2 => .f8_e5, + else => std.debug.panic("zml.kernel.tcl.from: dtype {s} has no TCL equivalent", .{@tagName(dt)}), + }; +} + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +/// A TCL function compiled by the Furiosa plugin as one native unit: +/// `declareArgs` declares the inputs in order, `ret` the outputs. +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + + pub const CallOpts = struct { + cfg: ConfigT, + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { + return build(allocator, cfg, null, null, null); + } + + /// `updated`, when given, receives the input each output updates + /// in place (`Builder.update`), or null. + fn build(allocator: std.mem.Allocator, cfg: ConfigT, inputs: ?[]const Shape, outputs: ?[]const Shape, updated: ?*[spec.outputs.len]?usize) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try Builder.open(allocator, ctx, name); + defer b.deinit(); + try spec.run(&b, cfg); + + const results = b.results orelse std.debug.panic("zml.kernel.tcl.Kernel({s}): run did not call ret", .{name}); + if (b.args.len != spec.inputs.len or results.len != spec.outputs.len) { + std.debug.panic("zml.kernel.tcl.Kernel({s}): {d} arguments and {d} results for {d} inputs and {d} outputs", .{ name, b.args.len, results.len, spec.inputs.len, spec.outputs.len }); + } + + if (inputs) |shapes| for (b.args, shapes, spec.inputs) |t, s, n| check(t, s, n); + if (outputs) |shapes| for (results, shapes, spec.outputs) |t, s, n| check(t, s, n); + if (updated) |u| for (u, results) |*dst, r| { + dst.* = b.updatedArgument(r); + }; + + return b.finish(); + } + + fn check(t: tcl_builder.Tensor, s: Shape, field: []const u8) void { + var same = from(s.dtype()) == t.dtype; + if (t.shards > 1) { + // A partition holds one chip's share of a chip-split DRAM tensor. + var n: i64 = 1; + for (t.axes) |a| n *= a.size; + same = same and n == @as(i64, @intCast(s.count())) * t.shards; + } else { + same = same and s.rank() == t.axes.len; + if (same) for (t.axes, 0..) |a, i| { + same = same and a.size == s.dim(i); + }; + } + + if (!same) std.debug.panic("zml.kernel.tcl.Kernel({s}): {s} is {f}, the TCL function has {s} over {d} axes", .{ name, field, s, @tagName(t.dtype), t.axes.len }); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + var input_shapes: [spec.inputs.len]Shape = undefined; + inline for (spec.inputs, 0..) |fname, i| { + inputs_arr[i] = @field(inputs, fname); + input_shapes[i] = inputs_arr[i].shape(); + } + + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); + + var updated: [spec.outputs.len]?usize = undefined; + const ir = build(cur.allocator, opts.cfg, &input_shapes, &outputs_arr, &updated) catch |err| + std.debug.panic("zml.kernel.tcl.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ir); + + var aliases: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, spec.outputs.len) = .empty; + for (updated, 0..) |input, output| if (input) |i| aliases.appendAssumeCapacity(.{ + .output_index = @intCast(output), + .operand_index = @intCast(i), + }); + const tensor_results = ops.tcl(inputs_arr, outputs_arr, .{ .ir = ir, .output_operand_aliases = aliases.constSlice() }); + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| { + @field(results, fname) = tensor_results[i]; + } + + return results; + } + }; +} diff --git a/zml/kernel/tcl/test.zig b/zml/kernel/tcl/test.zig new file mode 100644 index 00000000..b9ae1875 --- /dev/null +++ b/zml/kernel/tcl/test.zig @@ -0,0 +1,123 @@ +const std = @import("std"); + +const zml = @import("../../zml.zig"); +const tcl = @import("tcl.zig"); + +/// `furiosa.tcl.examples.softmax` over f32. +const TclSoftmax = struct { + const Cfg = struct { b: i64, s: i64, d: i64 }; + const K = tcl.Kernel(Cfg, .{ + .name = "softmax_kernel", + .inputs = &.{"x"}, + .outputs = &.{"y"}, + .run = run, + }); + + fn run(b: *tcl.Builder, cfg: Cfg) tcl.FinishError!void { + const B = b.axis("B", cfg.b); + const S = b.axis("S", cfg.s); + const D = b.axis("D", cfg.d); + const t = try b.declareArgs(.{ .x = .{ .dtype = .f32, .axes = &.{ B, S, D } } }); + + const op = b.tensorOperation(.{}); + const x = op.fetch(t.x, .{}); + const e = x.subf(x.reduce(&.{D}, .maxf)).exp(); + b.ret(&.{try op.commit(e.divf(e.reduce(&.{D}, .addf)), .{})}); + } +}; + +test "tcl kernels run on furiosa" { + const allocator = std.testing.allocator; + const cfg: TclSoftmax.Cfg = .{ .b = 2, .s = 16, .d = 256 }; + + const ir = try TclSoftmax.K.emit(allocator, cfg); + defer allocator.free(ir); + try std.testing.expect(std.mem.indexOf(u8, ir, "func.func @softmax_kernel(") != null); + + const platform = zml.testing.env(); + if (platform.target != .furiosa) return error.SkipZigTest; + + const Mod = struct { + pub fn forward(x: zml.Tensor) zml.Tensor { + return TclSoftmax.K.call(.{ .x = x }, .{ .y = x.shape() }, .{ .cfg = cfg }).y; + } + }; + const x: zml.Tensor = .init(.{ cfg.b, cfg.s, cfg.d }, .f32); + var exe = try zml.module.compile(allocator, std.testing.io, Mod.forward, .{x}, platform, .{}); + defer exe.deinit(); + + var input: [cfg.b * cfg.s * cfg.d]f32 = undefined; + for (&input, 0..) |*v, i| v.* = @as(f32, @floatFromInt(i % 37)) * 0.25 - 4; + var x_buffer: zml.Buffer = try .fromBytes(std.testing.io, platform, x.shape(), std.mem.sliceAsBytes(&input)); + defer x_buffer.deinit(); + var result = try exe.eval(allocator, std.testing.io, .{x_buffer}); + defer result.deinit(); + var host = try result.toSliceAlloc(allocator, std.testing.io); + defer host.free(allocator); + + const out = host.items(f32); + var row: usize = 0; + while (row < cfg.b * cfg.s) : (row += 1) { + const in = input[row * cfg.d ..][0..cfg.d]; + var max: f32 = in[0]; + for (in) |v| max = @max(max, v); + var sum: f32 = 0; + for (in) |v| sum += @exp(v - max); + for (in, out[row * cfg.d ..][0..cfg.d]) |v, y| try std.testing.expectApproxEqAbs(@exp(v - max) / sum, y, 1e-4); + } +} + +/// exp over a tensor split across two chips in one launch, returning the +/// split result and its all_gather. +const TclSplitExp = struct { + const Cfg = struct {}; + const K = tcl.Kernel(Cfg, .{ .name = "split_exp", .inputs = &.{"x"}, .outputs = &.{ "y", "all" }, .run = run }); + fn run(b: *tcl.Builder, _: Cfg) tcl.FinishError!void { + const C = b.axis("C", 2); + const A = b.axis("A", 1024); + const t = try b.declareArgs(.{ .x = .{ .dtype = .f32, .axes = &.{ C, A }, .dram = tcl.Dram{ .chip = &.{.{ .axis = C }} } } }); + const op = b.tensorOperation(.{ .context = .{ .layout = .{ .chip = .{ .axis = C } } } }); + const local = try op.commit(op.fetch(t.x, .{}).exp(), .{}); + const all = try b.allGather(local, C, .{}); + b.ret(&.{ try b.asDram(local, .{ .chip = &.{.{ .axis = C }} }), try b.asDram(all, .{}) }); + } + fn body(ctx: struct { x: zml.Tensor }, shapes: [2]zml.Shape) [2]zml.Tensor { + const r = K.call(.{ .x = ctx.x }, .{ .y = shapes[0], .all = shapes[1] }, .{ .cfg = .{} }); + return .{ r.y, r.all }; + } +}; + +// Needs FURIOSA_VISIBLE_DEVICES=0,1. +test "tcl multi-chip kernels run on two furiosa chips" { + const allocator = std.testing.allocator; + const platform = zml.testing.env(); + if (platform.target != .furiosa or platform.devices.len != 2) return error.SkipZigTest; + const tp = zml.testing.tpMesh(); + const x_shape = zml.Shape.init(.{ .c = 2, .a = 1024 }, .f32).withPartitioning(tp, .{ .c = .model }); + const Mod = struct { + pub fn forward(x: zml.Tensor, mesh: *const zml.Sharding.Mesh) [2]zml.Tensor { + const r = zml.ops.manualComputation(TclSplitExp.body, .{ .x = x }, .{ x.shape(), zml.Shape.init(.{ .c = 2, .a = 1024 }, .f32) }, mesh, .{ + .in_specs = .{ .x = .parse(.{ .c = .model }) }, + .out_specs = .{ .parse(.{ .c = .model }), .replicated }, + }); + return .{ r[0], r[1] }; + } + }; + const x: zml.Tensor = .fromShape(x_shape); + var exe = try zml.module.compile(allocator, std.testing.io, Mod.forward, .{ x, tp }, platform, .{ .meshes = &.{tp} }); + defer exe.deinit(); + var hx: [2 * 1024]f32 = undefined; + for (&hx, 0..) |*v, i| v.* = @as(f32, @floatFromInt(i % 13)) * 0.125 - 0.5; + var bx: zml.Buffer = try .fromBytes(std.testing.io, platform, x_shape, std.mem.sliceAsBytes(&hx)); + defer bx.deinit(); + var runner = try exe.runner(allocator); + defer runner.deinit(allocator); + var res: [2]zml.Buffer = undefined; + runner.run(std.testing.io, .{bx}, .{ &res[0], &res[1] }, .{ .wait = true }); + defer for (&res) |*r| r.deinit(); + for (res) |r| { + var host = try r.toSliceAlloc(allocator, std.testing.io); + defer host.free(allocator); + for (host.items(f32), 0..) |v, i| try std.testing.expectApproxEqRel(@exp(hx[i]), v, 1e-4); + } +} diff --git a/zml/kernel/triton/BUILD.bazel b/zml/kernel/triton/BUILD.bazel new file mode 100644 index 00000000..1e992352 --- /dev/null +++ b/zml/kernel/triton/BUILD.bazel @@ -0,0 +1,6 @@ +package(default_visibility = ["//zml/kernel:__pkg__"]) + +filegroup( + name = "triton", + srcs = ["triton.zig"], +) diff --git a/zml/kernel/triton/triton.zig b/zml/kernel/triton/triton.zig new file mode 100644 index 00000000..6a91a03f --- /dev/null +++ b/zml/kernel/triton/triton.zig @@ -0,0 +1,165 @@ +const std = @import("std"); + +const func = @import("mlir/dialects/func"); +const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); +const stdx = @import("stdx"); +const triton_builder = @import("dsl/triton/builder"); +pub const Builder = triton_builder.Builder; +pub const Value = triton_builder.Value; +pub const DType = triton_builder.DType; +pub const FinishError = triton_builder.FinishError; + +const Compiler = @import("../../Compiler.zig"); +const DataType = @import("../../dtype.zig").DataType; +const ops = @import("../../ops.zig"); +const Shape = @import("../../shape.zig").Shape; +const Tensor = @import("../../tensor.zig").Tensor; + +/// Build a struct type with one field per name, all of type `FieldT`. +fn StructOf(comptime names: []const [:0]const u8, comptime FieldT: type) type { + var name_slices: [names.len][]const u8 = undefined; + for (names, 0..) |n, i| name_slices[i] = n; + return @Struct(.auto, null, &name_slices, &@splat(FieldT), &@splat(.{})); +} + +fn makeKernelContext(comptime registerDialects: fn (*mlir.DialectRegistry) void) std.mem.Allocator.Error!*mlir.Context { + mlir.registerPasses("Transforms"); + + const registry = try mlir.DialectRegistry.init(); + defer registry.deinit(); + + registerDialects(registry); + + func.registerExtensions(registry); + + const ctx = try mlir.Context.init(.{ .registry = registry, .threading = false }); + ctx.loadAllAvailableDialects(); + + return ctx; +} + +fn resolveOutputOperandAliases( + aliases_opt: anytype, + operand_offset: i64, +) stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) { + var out: stdx.BoundedArray(stablehlo.CustomCallOpts.OutputOperandAlias, stablehlo.CustomCallOpts.MAX_RESULTS) = .empty; + + if (aliases_opt) |a| { + inline for (@typeInfo(@TypeOf(a)).@"struct".field_names, 0..) |field_name, output_index| { + if (@field(a, field_name)) |operand| out.appendAssumeCapacity(.{ + .output_index = @intCast(output_index), + .operand_index = @as(i64, @intCast(@backingInt(operand))) + operand_offset, + }); + } + } + + return out; +} + +pub fn newContext() std.mem.Allocator.Error!*mlir.Context { + return makeKernelContext(triton_builder.registerDialects); +} + +pub fn from(dt: DataType) DType { + return switch (dt) { + .bool => .i1, + .i8, .u8 => .i8, + .i16, .u16 => .i16, + .i32, .u32 => .i32, + .i64, .u64 => .i64, + .f16 => .f16, + .bf16 => .bf16, + .f32 => .f32, + .f64 => .f64, + .f8e4m3fn => .f8e4m3fn, + .f8e4m3fnuz => .f8e4m3fnuz, + .f8e5m2 => .f8e5m2, + else => std.debug.panic("zml.kernel.triton.from: dtype {s} has no Triton equivalent", .{@tagName(dt)}), + }; +} + +fn Spec(comptime Config: type) type { + return struct { + name: [:0]const u8, + inputs: []const [:0]const u8, + outputs: []const [:0]const u8, + run: *const fn (*Builder, Config) FinishError!void, + }; +} + +pub fn Kernel( + comptime ConfigT: type, + comptime spec: Spec(ConfigT), +) type { + return struct { + pub const name: [:0]const u8 = spec.name; + pub const Config = ConfigT; + pub const Inputs = StructOf(spec.inputs, Tensor); + pub const Outputs = StructOf(spec.outputs, Shape); + pub const Results = StructOf(spec.outputs, Tensor); + + pub const CallOpts = struct { + cfg: ConfigT, + grid: [3]i32, + num_stages: i32, + num_warps: i32, + output_operand_aliases: ?ops.CustomCallOutputOperandAliases(Inputs, Outputs) = null, + debug: bool = false, + /// Bytes for Triton's implicit global scratch argument, across all CTAs. + global_scratch_memory_size: i32 = 0, + }; + + pub fn emit(allocator: std.mem.Allocator, cfg: ConfigT) ![:0]const u8 { + const ctx = try newContext(); + defer ctx.deinit(); + + var b = try triton_builder.Builder.open(allocator, ctx, name); + defer b.deinit(); + + try spec.run(&b, cfg); + + return b.finish(); + } + + pub fn call(inputs: Inputs, outputs: Outputs, opts: CallOpts) Results { + const cur = Compiler.current(); + + const ttir = emit(cur.allocator, opts.cfg) catch |err| + std.debug.panic("zml.kernel.triton.Kernel({s}).call: emit failed: {}", .{ name, err }); + defer cur.allocator.free(ttir); + + var inputs_arr: [spec.inputs.len]Tensor = undefined; + inline for (spec.inputs, 0..) |fname, i| inputs_arr[i] = @field(inputs, fname); + + var outputs_arr: [spec.outputs.len]Shape = undefined; + inline for (spec.outputs, 0..) |fname, i| outputs_arr[i] = @field(outputs, fname); + + const aliases = resolveOutputOperandAliases(opts.output_operand_aliases, 0); + + const call_opts: ops.TritonOps = .{ + .debug = opts.debug, + .name = name, + .ir = ttir, + .grid = opts.grid, + .num_stages = opts.num_stages, + .num_warps = opts.num_warps, + .output_operand_aliases = aliases.constSlice(), + .is_tma_allowed = opts.global_scratch_memory_size > 0, + .global_scratch_memory_size = opts.global_scratch_memory_size, + }; + + const triton_results = if (opts.global_scratch_memory_size > 0) blk: { + // XLA maps the final custom-call result to Triton's implicit + // scratch pointer; it must not appear in the TTIR signature. + const extended = outputs_arr ++ [_]Shape{.init(.{opts.global_scratch_memory_size}, .u8)}; + const res = ops.triton(inputs_arr, extended, call_opts); + break :blk res[0..spec.outputs.len].*; + } else ops.triton(inputs_arr, outputs_arr, call_opts); + + var results: Results = undefined; + inline for (spec.outputs, 0..) |fname, i| @field(results, fname) = triton_results[i]; + return results; + } + }; +} diff --git a/zml/kernels/flash_attention/README.md b/zml/kernels/flash_attention/README.md index b5fe7682..81a35cf4 100644 --- a/zml/kernels/flash_attention/README.md +++ b/zml/kernels/flash_attention/README.md @@ -265,7 +265,7 @@ are not uniformly supported by the current kernels; an existing prefill path requires equal query/key lengths and assumes positions begin at zero. Treat these as current conformance gaps, not alternative definitions of causality. -The tests in [flash_attention.zig](flash_attention.zig) compare results +The tests in [tests.zig](tests.zig) compare results against `sdpa` with the gathered causal mask. The original shared cases cover `D = 64`, MHA and GQA, single-token decode at full and partial context lengths, batched decode with diff --git a/zml/kernels/flash_attention/cute/attention.zig b/zml/kernels/flash_attention/cute/attention.zig index 0d01c357..ec0fce16 100644 --- a/zml/kernels/flash_attention/cute/attention.zig +++ b/zml/kernels/flash_attention/cute/attention.zig @@ -1,4 +1,7 @@ const std = @import("std"); +const mlir = @import("mlir"); +const llvm = @import("mlir/dialects/llvm"); +const vector = @import("mlir/dialects/vector"); const zml = @import("../../../zml.zig"); const cute = zml.kernel.cute; @@ -14,6 +17,8 @@ pub const Config = struct { head_dim: i64, dtype: cute.DType, splits: i64 = 1, + /// Query j sees the keys up to positions + j. Otherwise every query sees every key and positions is ignored. + causal: bool = true, }; /// Planning uses local (post-sharding) dimensions. XLA owns the compiled CUDA @@ -144,13 +149,16 @@ fn runSized(b: *B, cfg: Config, comptime fragments: usize) cute.FinishError!void const query = row.div(group).minimum(cfg.query_length - 1); const head = kv_head.mul(group).add(row.rem(group)); const start = a.positions.get(.{batch}); - const position = start.add(query); - const causal_end = start.add(first_row.add(3).div(group).minimum(cfg.query_length - 1)).add(1); + const position = if (cfg.causal) start.add(query) else b.constant(.i32, cfg.key_length - 1); + const live_end = if (cfg.causal) + start.add(first_row.add(3).div(group).minimum(cfg.query_length - 1)).add(1) + else + b.constant(.i32, cfg.key_length); // Partition the live cache, rather than its capacity, so decode uses all // blocks even when only a prefix of the cache is populated. - const chunk = causal_end.add(cfg.splits - 1).div(cfg.splits); + const chunk = live_end.add(cfg.splits - 1).div(cfg.splits); const begin = if (split) b.blockIdx().z.mul(chunk) else b.constant(.i32, 0); - const end = if (split) begin.add(chunk).minimum(causal_end) else causal_end; + const end = if (split) begin.add(chunk).minimum(live_end) else live_end; const zero = b.constant(.f32, 0); const neg_inf = b.constant(.f32, -std.math.inf(f32)); @@ -305,6 +313,494 @@ fn merge(b: *B, cfg: Config) cute.FinishError!void { } } +// ==================== Prefill ==================== +// +// FlashAttention-2 forward for `query_length > 1`, with the tiling FA2's sm80 kernel uses on sm_120 +// (`Flash_fwd_kernel_traits<128, 128, 64, 4>`): a 128-thread block owns `prefillRows` queries of one +// head and every warp owns whole rows (all the keys of a tile, the whole head dimension). Scores and +// output stay in `mma.sync` accumulators, so P is reused from registers as the A operand of P @ V, +// without cross-warp reductions. Q, K and V tiles go through swizzled shared memory: V_j is copied +// while S_j = Q K_j^T is computed, K_{j+1} while O += P_j V_j is. +// +// It is a program, not a bare kernel: its host function launches the kernel with the dynamic shared +// memory the kernel allocates (up to 96 KiB). A bare kernel's shared memory is static, which ptxas +// caps at 48 KiB before sm_120. + +pub const Prefill = cute.Program(Config, .{ + .name = "flash_attention_cute_prefill", + .inputs = &.{ "q", "k", "v", "positions" }, + .outputs = &.{"out"}, + .run = prefillProgram, +}); + +const prefill_kernel_name = "flash_attention_cute_prefill_kernel"; + +/// Keys per tile (FA2's kBlockN). +const prefill_keys = 64; + +/// Whether `Prefill` handles `cfg`; the other configurations use `Kernel`. +/// Same as FA2 we only support 16-bit inputs +pub fn prefillSupported(cfg: Config) bool { + const dtype_ok = cfg.dtype == .bf16 or cfg.dtype == .f16; + const head_dim_ok = cfg.head_dim == 32 or cfg.head_dim == 64 or cfg.head_dim == 128 or cfg.head_dim == 256; + return cfg.query_length > 1 and cfg.splits == 1 and dtype_ok and head_dim_ok; +} + +/// m16 row tiles per warp: two up to head dimension 128, one above so the output accumulators fit in registers. +fn prefillRowTiles(head_dim: i64) i64 { + return if (head_dim <= 128) 2 else 1; +} + +/// Queries per block (FA2's kBlockM). +fn prefillRows(cfg: Config) i64 { + return 4 * 16 * prefillRowTiles(cfg.head_dim); +} + +/// Query tiles on x, so the blocks resident at the same time share the K/V of one head in L2. +fn prefillGrid(cfg: Config) [3]i32 { + return .{ @intCast(@divTrunc(cfg.query_length + prefillRows(cfg) - 1, prefillRows(cfg))), @intCast(cfg.query_heads), @intCast(cfg.batch) }; +} + +fn prefillProgram(b: *B, cfg: Config) cute.FinishError!void { + std.debug.assert(prefillSupported(cfg)); + switch (cfg.head_dim) { + inline 32, 64, 128, 256 => |d| switch (cfg.dtype) { + inline .bf16, .f16 => |dt| { + b.beginFunction(prefill_kernel_name, .cuda_kernel); + try prefillSized(b, cfg, d, dt); + b.endFunction(block); + + b.beginFunction(Prefill.name, .host); + const host = try b.declareArgs(.{ + .q = .{ .ptr = dt }, + .k = .{ .ptr = dt }, + .v = .{ .ptr = dt }, + .positions = .{ .ptr = cute.DType.i32 }, + .out = .{ .ptr = dt }, + }); + const grid = prefillGrid(cfg); + const config = b.makeLaunchConfig(.{ + .grid = .{ b.constant(.i32, grid[0]), b.constant(.i32, grid[1]), b.constant(.i32, grid[2]) }, + .block = .{ b.constant(.i32, block[0]), b.constant(.i32, block[1]), b.constant(.i32, block[2]) }, + .dynamic_smem = b.kernelSmemSize(prefill_kernel_name), + .stream = b.cudaStream(), + }); + const launch = b.launchEx(prefill_kernel_name, config, .{ host.q, host.k, host.v, host.positions, host.out }); + b.returnHostStatus(b.cudaResultStatus(launch)); + b.endFunction(null); + }, + else => unreachable, + }, + else => unreachable, + } +} + +/// PTX `max.f32` (FMNMX). `Value.maximum` and `Value.maxnum` both lower to NaN-handling compare and +/// select sequences (three instructions); prefill scores are never NaN, masked ones are -inf. +fn fmax(b: *B, x: Value, y: Value) Value { + return b.callIntrinsic("llvm.nvvm.fmax.f", &.{ x, y }, .float(b.ctx, .f32)).?; +} + +fn unpackStruct(b: *B, comptime n: usize, value: Value, dtype: cute.DType) [n]Value { + var result: [n]Value = undefined; + for (&result, 0..) |*r, i| { + r.* = b.emit(llvm.extractvalue(b.ctx, value.inner, dtype.toMlir(b.ctx), &.{@intCast(i)}, b.loc())); + } + return result; +} + +/// `ldmatrix.x4`: lanes 8i..8i+7 address the rows of matrix i, the result holds the four 8x8 fragments. +fn matrixLoad(b: *B, comptime transpose: bool, address: Value) [4]Value { + const assembly = "ldmatrix.sync.aligned.m8n8.x4" ++ (if (transpose) ".trans" else "") ++ ".shared.b16 {$0, $1, $2, $3}, [$4];"; + const result = b.emit(llvm.inline_asm(b.ctx, assembly, &.{address.inner}, .{ + .result_type = llvm.structType(b.ctx, &@as([4]*const mlir.Type, @splat(.int(b.ctx, .i32)))), + .constraints = "=r,=r,=r,=r,r", + .has_side_effects = true, + }, b.loc())); + return unpackStruct(b, 4, result, .i32); +} + +/// `mma.sync.m16n8k16` with f32 accumulators: acc += a (16x16, row) @ b (16x8, col). +fn mma(b: *B, comptime dtype: cute.DType, a: [4]Value, rhs: [2]Value, acc: [4]Value) [4]Value { + const assembly = "mma.sync.aligned.m16n8k16.row.col.f32." ++ @tagName(dtype) ++ "." ++ @tagName(dtype) ++ ".f32 " ++ + "{$0, $1, $2, $3}, {$4, $5, $6, $7}, {$8, $9}, {$10, $11, $12, $13};"; + const result = b.emit(llvm.inline_asm(b.ctx, assembly, &.{ a[0].inner, a[1].inner, a[2].inner, a[3].inner, rhs[0].inner, rhs[1].inner, acc[0].inner, acc[1].inner, acc[2].inner, acc[3].inner }, .{ + .result_type = llvm.structType(b.ctx, &@as([4]*const mlir.Type, @splat(.float(b.ctx, .f32)))), + .constraints = "=f,=f,=f,=f,r,r,r,r,r,r,f,f,f,f", + }, b.loc())); + return unpackStruct(b, 4, result, .f32); +} + +/// The four f32 accumulators of one mma.sync, carried across loop iterations as one value (the loop +/// carries at most 64 values). +fn packAcc(b: *B, values: [4]Value) Value { + return b.emit(vector.from_elements(b.ctx, &.{ values[0].inner, values[1].inner, values[2].inner, values[3].inner }, .vector(&.{4}, .float(b.ctx, .f32)), b.loc())); +} + +fn unpackAcc(b: *B, value: Value) [4]Value { + var result: [4]Value = undefined; + for (&result, 0..) |*r, i| { + r.* = b.vectorExtract(value, @intCast(i), .f32); + } + return result; +} + +/// Comptime layout of one prefill block, shared by the setup and the per-tile steps. +/// +/// Tiles are row major with the 16-byte vector index of every row XORed with row % swizzle_rows, so +/// 16-byte cp.async stores and 8-row ldmatrix reads are free of bank conflicts. Shared memory addresses +/// are folded the way CuTe's static layouts let nvcc fold them: the XORed row bits of every address a +/// thread uses are a per-thread constant, and the k-step part of a column offset is bit-disjoint from +/// the per-thread part, so each ldmatrix / cp.async address is a per-thread base plus a constant +/// (XORed into the 16-byte vector index). +fn PrefillTile(comptime d: i64, comptime dt: cute.DType) type { + return struct { + const row_tiles = prefillRowTiles(d); + const rows = 4 * 16 * row_tiles; + const n = prefill_keys; + const score_tiles = n / 8; + const out_tiles = d / 8; + const row_bytes = d * 2; + /// The XORed vector index must stay in its row: rows of d=32 only have 4 vectors. + const swizzle_rows = @min(8, d / 8); + const q_bytes = rows * row_bytes; + const kv_bytes = n * row_bytes; + /// Rows covered by one 16-byte cp.async per thread of the block. + const copy_rows = @divExact(128, d / 8); + /// Running max and per-thread partial sum of the two rows of every row tile. + const stats = 2 * row_tiles; + /// Carried by the key loop: maxima, sums, then the packed output accumulators of every + /// (row tile, 8 output features). + const carried = 2 * stats + row_tiles * out_tiles; + const scale: f32 = std.math.log2e / std.math.sqrt(@as(f32, @floatFromInt(d))); + + comptime { + // copy_rows rows starting at a multiple of copy_rows never carry into the XORed bits. + std.debug.assert(std.math.isPowerOfTwo(copy_rows)); + std.debug.assert(std.math.isPowerOfTwo(swizzle_rows)); + } + + b: *B, + cfg: Config, + lane: Value, + warp: Value, + first_query: Value, + start: Value, + q_smem: Value, + k_smem: Value, + v_smem: Value, + k_ptr: Value, + v_ptr: Value, + kv_base: Value, + tiles: Value, + /// cp.async: this thread copies the 16-byte vector `copy_col` of rows `copy_row + i * copy_rows`. + copy_row: Value, + copy_col: Value, + /// Byte offset in a tile of this thread's first cp.async destination, without the XORed part. + copy_row_offset: Value, + copy_vector_offset: Value, + /// ldmatrix rows of this thread, as byte offsets, and their XORed 16-byte vector offsets. + q_frag_row: Value, + k_frag_row: Value, + v_frag_row: Value, + qv_frag_vector: Value, + k_frag_vector: Value, + + const Self = @This(); + + fn init(b: *B, cfg: Config, tid: Value, first_query: Value, start: Value, shared: Value, k_ptr: Value, v_ptr: Value, kv_base: Value, tiles: Value) Self { + const lane = tid.rem(32); + const warp = tid.div(32); + const copy_row = tid.div(d / 8); + const copy_col = tid.rem(d / 8); + const lane8 = lane.rem(8); + // The fragment rows a thread reads are lane % 8 plus multiples of 8. + const lane_xor = lane.rem(swizzle_rows); + return .{ + .b = b, + .cfg = cfg, + .lane = lane, + .warp = warp, + .first_query = first_query, + .start = start, + .q_smem = shared, + .k_smem = shared.add(q_bytes), + .v_smem = shared.add(q_bytes + kv_bytes), + .k_ptr = k_ptr, + .v_ptr = v_ptr, + .kv_base = kv_base, + .tiles = tiles, + .copy_row = copy_row, + .copy_col = copy_col, + .copy_row_offset = copy_row.mul(row_bytes), + .copy_vector_offset = copy_col.bitXor(copy_row.rem(swizzle_rows)).mul(16), + // Q (A operand): rows lane % 16, vectors lane / 16 (+ 2 per k-step). + .q_frag_row = shared.add(warp.mul(16 * row_tiles).add(lane.rem(16)).mul(row_bytes)), + // K (B operand): keys lane % 8 + 8 * (lane / 16), vectors (lane / 8) % 2 (+ 2 per k-step). + .k_frag_row = shared.add(q_bytes).add(lane8.add(lane.div(16).mul(8)).mul(row_bytes)), + // V (B operand, transposed): keys lane % 8 + 8 * ((lane / 8) % 2), vectors lane / 16 (+ 2 per 16 features). + .v_frag_row = shared.add(q_bytes + kv_bytes).add(lane8.add(lane.div(8).rem(2).mul(8)).mul(row_bytes)), + .qv_frag_vector = lane.div(16).bitXor(lane_xor).mul(16), + .k_frag_vector = lane.div(8).rem(2).bitXor(lane_xor).mul(16), + }; + } + + /// Destination of the i-th cp.async of this thread in a tile starting at `base`. + fn copyDst(t: Self, base: Value, comptime i: usize) Value { + const xor = ((copy_rows * i) % swizzle_rows) * 16; + const vector_offset = if (xor == 0) t.copy_vector_offset else t.copy_vector_offset.bitXor(xor); + return base.add(t.copy_row_offset).add(vector_offset).add(copy_rows * i * row_bytes); + } + + /// Copy keys `key_base..key_base + n` of `src` into `dst`; keys past the end are zero-filled. + fn loadTile(t: Self, src: Value, dst: Value, key_base: Value) void { + const b = t.b; + const stride = t.cfg.kv_heads * d; + const first_key = key_base.add(t.copy_row); + const safe = src.add(t.kv_base.mul(2)); + const first = safe.add(first_key.to(.i64).mul(stride).add(t.copy_col.mul(8).to(.i64)).mul(2)); + inline for (0..@divExact(n, copy_rows)) |i| { + const valid = first_key.add(copy_rows * i).lt(t.cfg.key_length); + const source = b.select(valid, first.add(copy_rows * i * stride * 2), safe); + b.cpAsync16(t.copyDst(dst, i), source, valid); + } + } + + fn queryOf(t: Self, mi: usize, h: usize) Value { + return t.first_query.add(t.warp.mul(16 * row_tiles)).add(mi * 16).add(t.lane.div(4)).add(h * 8); + } + + fn accIndex(comptime mi: usize, comptime di: usize) usize { + return @intCast(2 * stats + @as(i64, mi) * out_tiles + @as(i64, di)); + } + + /// One key tile. `masked` tiles apply the causal and key length masks; the other ones are + /// fully visible to every query of the block. + fn step(t: Self, comptime masked: bool, tile: Value, carried_in: [carried]Value) [carried]Value { + const b = t.b; + const lane = t.lane; + const zero = b.constant(.f32, 0); + const neg_inf = b.constant(.f32, -std.math.inf(f32)); + const key_base = tile.mul(n); + b.cpAsyncWaitAll(); + b.syncThreads(); + t.loadTile(t.v_ptr, t.v_smem, key_base); + + // S = Q K^T, k-steps of 16 features. + var s: [row_tiles][score_tiles][4]Value = @splat(@splat(@splat(zero))); + inline for (0..d / 16) |ks| { + var qa: [row_tiles][4]Value = undefined; + inline for (0..row_tiles) |mi| { + qa[mi] = matrixLoad(b, false, t.q_frag_row.add(mi * 16 * row_bytes).add(t.qv_frag_vector.bitXor(ks * 32))); + } + const k_vector = t.k_frag_vector.bitXor(ks * 32); + inline for (0..score_tiles / 2) |np| { + const kb = matrixLoad(b, false, t.k_frag_row.add(np * 16 * row_bytes).add(k_vector)); + inline for (0..row_tiles) |mi| { + s[mi][2 * np] = mma(b, dt, qa[mi], .{ kb[0], kb[1] }, s[mi][2 * np]); + s[mi][2 * np + 1] = mma(b, dt, qa[mi], .{ kb[2], kb[3] }, s[mi][2 * np + 1]); + } + } + } + + // Online softmax in base 2 on the raw scores, the scale is folded into the exponent: + // p = exp2(s * scale - m * scale). Thread rows: lane / 4 (accumulators 0, 1) and lane / 4 + 8 (2, 3). + var next: [carried]Value = undefined; + var alpha: [row_tiles][2]Value = undefined; + inline for (0..row_tiles) |mi| { + if (masked) { + inline for (0..score_tiles) |ni| { + inline for (0..4) |j| { + const key = key_base.add(ni * 8).add(lane.rem(4).mul(2)).add(j % 2); + const in_range = key.lt(t.cfg.key_length); + const visible = if (t.cfg.causal) in_range.bitAnd(key.le(t.start.add(t.queryOf(mi, j / 2)))) else in_range; + s[mi][ni][j] = b.select(visible, s[mi][ni][j], neg_inf); + } + } + } + inline for (0..2) |h| { + var row_max = s[mi][0][h * 2]; + inline for (0..score_tiles) |ni| { + inline for (0..2) |j| { + row_max = fmax(b, row_max, s[mi][ni][h * 2 + j]); + } + } + inline for (.{ 2, 1 }) |offset| { + row_max = fmax(b, row_max, b.shuffleXor(row_max, offset)); + } + const old_max = carried_in[mi * 2 + h]; + const new_max = fmax(b, old_max, row_max); + // A row without any visible key so far keeps m = -inf: exponentiate against a finite value. + const neg_scaled_max = fmax(b, new_max, b.constant(.f32, -std.math.floatMax(f32))).mul(-scale); + alpha[mi][h] = b.exp2Approx(old_max.mul(scale).add(neg_scaled_max)); + next[mi * 2 + h] = new_max; + var sum = carried_in[stats + mi * 2 + h].mul(alpha[mi][h]); + inline for (0..score_tiles) |ni| { + inline for (0..2) |j| { + s[mi][ni][h * 2 + j] = b.exp2Approx(s[mi][ni][h * 2 + j].mul(scale).add(neg_scaled_max)); + sum = sum.add(s[mi][ni][h * 2 + j]); + } + } + next[stats + mi * 2 + h] = sum; + } + } + + // All warps are done with K_j once V_j has landed: start copying K_{j+1}. + b.cpAsyncWaitAll(); + b.syncThreads(); + var prefetch = b.openIf(tile.add(1).lt(t.tiles)); + { + t.loadTile(t.k_ptr, t.k_smem, key_base.add(n)); + prefetch.close(); + } + + // O = O * alpha + P V. The accumulators of two adjacent score tiles are the A operand of one k-step. + var o: [row_tiles][out_tiles][4]Value = undefined; + inline for (0..row_tiles) |mi| { + inline for (0..out_tiles) |di| { + const prev = unpackAcc(b, carried_in[accIndex(mi, di)]); + inline for (0..4) |j| { + o[mi][di][j] = prev[j].mul(alpha[mi][j / 2]); + } + } + } + inline for (0..n / 16) |ks| { + var pa: [row_tiles][4]Value = undefined; + inline for (0..row_tiles) |mi| { + pa[mi] = .{ + b.packHalf2(s[mi][2 * ks][0], s[mi][2 * ks][1], dt), + b.packHalf2(s[mi][2 * ks][2], s[mi][2 * ks][3], dt), + b.packHalf2(s[mi][2 * ks + 1][0], s[mi][2 * ks + 1][1], dt), + b.packHalf2(s[mi][2 * ks + 1][2], s[mi][2 * ks + 1][3], dt), + }; + } + const v_row = t.v_frag_row.add(ks * 16 * row_bytes); + inline for (0..out_tiles / 2) |dp| { + const vb = matrixLoad(b, true, v_row.add(t.qv_frag_vector.bitXor(dp * 32))); + inline for (0..row_tiles) |mi| { + o[mi][2 * dp] = mma(b, dt, pa[mi], .{ vb[0], vb[1] }, o[mi][2 * dp]); + o[mi][2 * dp + 1] = mma(b, dt, pa[mi], .{ vb[2], vb[3] }, o[mi][2 * dp + 1]); + } + } + } + inline for (0..row_tiles) |mi| { + inline for (0..out_tiles) |di| { + next[accIndex(mi, di)] = packAcc(b, o[mi][di]); + } + } + return next; + } + }; +} + +fn prefillSized(b: *B, cfg: Config, comptime d: i64, comptime dt: cute.DType) cute.FinishError!void { + const T = PrefillTile(d, dt); + const a = try b.declareArgs(.{ + .q = .{ .ptr = dt }, + .k = .{ .ptr = dt }, + .v = .{ .ptr = dt }, + .positions = .{ .ptr = cute.DType.i32 }, + .out = .{ .ptr = dt }, + }); + b.setFunctionAttribute(.named(b.ctx, "cu_attrs", .dict(b.ctx, &.{.named(b.ctx, "max_dynamic_shared_size_bytes", b.devMaxSharedMemoryOptin())}))); + + const tid = b.threadIdx().x; + const first_query = b.blockIdx().x.mul(T.rows); + const head = b.blockIdx().y; + const batch = b.blockIdx().z; + const kv_head = head.div(@divExact(cfg.query_heads, cfg.kv_heads)); + // One shared storage object (like a Python CuTe `SharedStorage`): the host compiler sizes the + // launch's dynamic shared memory from it, which `kernelSmemSize` returns. `allocSmem` and + // `allocSmemView` both become static shared memory instead. + b.setFunctionAttribute(.named(b.ctx, "smem.partition_num", .int(b.ctx, .i32, 1))); + const fields = [_][]const u8{ + std.fmt.comptimePrint("sQ:{d}:0", .{T.q_bytes}), + std.fmt.comptimePrint("sK:{d}:{d}", .{ T.kv_bytes, T.q_bytes }), + std.fmt.comptimePrint("sV:{d}:{d}", .{ T.kv_bytes, T.q_bytes + T.kv_bytes }), + }; + const smem = b.allocSmemStorage(.i8, 128, b.layoutType(b.layoutSpec(.{T.q_bytes + 2 * T.kv_bytes}, .{1})), 0, &fields); + const shared = b.ptrToInt(b.getIter(smem, b.ptrTy(.i8, .smem, 128) catch @panic("bad shared memory pointer")).value(), .i32); + const q_ptr = b.ptrToInt(a.q, .i64); + const out_ptr = b.ptrToInt(a.out, .i64); + const start = b.loadWords(1, b.ptrToInt(a.positions, .i64).add(batch.to(.i64).mul(4)), .global, .{})[0]; + + // Element offsets of this (batch, head) in q/out and (batch, kv head) in k/v, and the q/out row stride. + const q_stride = cfg.query_heads * d; + const q_base = batch.to(.i64).mul(cfg.query_length * q_stride).add(head.to(.i64).mul(d)); + const kv_base = batch.to(.i64).mul(cfg.key_length * cfg.kv_heads * d).add(kv_head.to(.i64).mul(d)); + + // Causal: the last query of the block sees keys up to start + last_query, and every query of the + // block sees the keys up to start + first_query. Otherwise only the last tile, past the end of the + // keys, is masked. + const last_query = first_query.add(T.rows).minimum(cfg.query_length).sub(1); + const tiles = if (cfg.causal) + start.add(last_query).add(1).minimum(cfg.key_length).add(T.n - 1).div(T.n) + else + b.constant(.i32, @divTrunc(cfg.key_length + T.n - 1, T.n)); + const full_tiles = if (cfg.causal) + start.add(first_query).add(1).minimum(cfg.key_length).div(T.n).minimum(tiles) + else + b.constant(.i32, @divTrunc(cfg.key_length, T.n)); + + const t: T = .init(b, cfg, tid, first_query, start, shared, b.ptrToInt(a.k, .i64), b.ptrToInt(a.v, .i64), kv_base, tiles); + + // Out of range rows are zero-filled and never read from global memory. + { + const first_row = first_query.add(t.copy_row); + const safe = q_ptr.add(q_base.mul(2)); + const first = safe.add(first_row.to(.i64).mul(q_stride).add(t.copy_col.mul(8).to(.i64)).mul(2)); + inline for (0..@divExact(T.rows, T.copy_rows)) |i| { + const valid = first_row.add(T.copy_rows * i).lt(cfg.query_length); + b.cpAsync16(t.copyDst(t.q_smem, i), b.select(valid, first.add(T.copy_rows * i * q_stride * 2), safe), valid); + } + } + t.loadTile(t.k_ptr, t.k_smem, b.constant(.i32, 0)); + + var initial: [T.carried]Value = undefined; + const zero = b.constant(.f32, 0); + for (initial[0..T.stats]) |*v| { + v.* = b.constant(.f32, -std.math.inf(f32)); + } + for (initial[T.stats .. 2 * T.stats]) |*v| { + v.* = zero; + } + for (initial[2 * T.stats ..]) |*v| { + v.* = packAcc(b, .{ zero, zero, zero, zero }); + } + + var full = b.openFor(0, full_tiles, 1, initial); + { + full.yield(t.step(false, full.iv, full.carried[0..T.carried].*)); + } + var diagonal = b.openFor(full_tiles, tiles, 1, full.results[0..T.carried].*); + { + diagonal.yield(t.step(true, diagonal.iv, diagonal.carried[0..T.carried].*)); + } + const results = diagonal.results[0..T.carried].*; + + // Normalize by the row sums (reduced over the quad) and store pairs of features as one 32-bit word. + inline for (0..T.row_tiles) |mi| { + inline for (0..2) |h| { + var sum = results[T.stats + mi * 2 + h]; + inline for (.{ 2, 1 }) |offset| { + sum = sum.add(b.shuffleXor(sum, offset)); + } + const norm = b.select(sum.gt(0), b.rcpApprox(sum), zero); + const query = t.queryOf(mi, h); + var valid = b.openIf(query.lt(cfg.query_length)); + { + const row_offset = q_base.add(query.to(.i64).mul(q_stride)).add(t.lane.rem(4).mul(2).to(.i64)); + inline for (0..T.out_tiles) |di| { + const acc = unpackAcc(b, results[T.accIndex(mi, di)]); + const word = b.packHalf2(acc[h * 2].mul(norm), acc[h * 2 + 1].mul(norm), dt); + b.storeWords(1, out_ptr.add(row_offset.add(di * 8).mul(2)), .global, .{word}); + } + valid.close(); + } + } + } +} + test "CuTe attention emits tiled online softmax through head dimension 512" { for ([_]i64{ 64, 80, 128, 256, 512 }) |head_dim| { const ir = try Kernel.emit(std.testing.allocator, .{ diff --git a/zml/kernels/flash_attention/cute/cute.zig b/zml/kernels/flash_attention/cute/cute.zig index 2958888d..36e0953b 100644 --- a/zml/kernels/flash_attention/cute/cute.zig +++ b/zml/kernels/flash_attention/cute/cute.zig @@ -15,7 +15,8 @@ pub fn isAvailable(platform: *const zml.Platform) bool { return cc.major >= 8; } -pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor) zml.Tensor { +/// `causal`: query j sees the keys up to token_index + j. Otherwise every key, and token_index is ignored. +pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, causal: bool) zml.Tensor { stdx.debug.assert(q.shape().hasTags(.{ .q, .h, .hd }) and k.shape().hasTags(.{ .k, .h, .hd }) and v.shape().hasTags(.{ .k, .h, .hd }), "CuTe attention requires q/h/hd and k/h/hd tags", .{}); const batched = q.shape().hasTag(.b) != null; const rank: usize = if (batched) 4 else 3; @@ -24,7 +25,7 @@ pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.T stdx.debug.assert(q.dtype() == .bf16 or q.dtype() == .f16 or q.dtype() == .f32, "CuTe attention supports bf16, f16, and f32", .{}); stdx.debug.assert(q.dim(.hd) > 0 and q.dim(.hd) <= 512 and q.dim(.hd) == k.dim(.hd) and k.dim(.hd) == v.dim(.hd), "CuTe attention requires matching head dimensions in [1, 512]", .{}); stdx.debug.assert(k.dim(.h) > 0 and q.dim(.h) > 0 and @rem(q.dim(.h), k.dim(.h)) == 0 and k.dim(.h) == v.dim(.h), "CuTe attention requires Hq divisible by Hkv and matching K/V heads", .{}); - stdx.debug.assert(q.dim(.q) > 0 and q.dim(.q) <= k.dim(.k) and k.dim(.k) == v.dim(.k), "CuTe attention requires 0 < Lq <= Lk and matching K/V lengths", .{}); + stdx.debug.assert(q.dim(.q) > 0 and (!causal or q.dim(.q) <= k.dim(.k)) and k.dim(.k) == v.dim(.k), "CuTe attention requires 0 < Lq, Lq <= Lk when causal, and matching K/V lengths", .{}); stdx.debug.assert(token_index.dtype().isInteger(), "CuTe attention requires integer token positions", .{}); if (batched) { stdx.debug.assert(k.shape().hasTag(.b) != null and v.shape().hasTag(.b) != null and q.dim(.b) == k.dim(.b) and q.dim(.b) == v.dim(.b), "CuTe attention requires matching batch dimensions", .{}); @@ -33,64 +34,51 @@ pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.T stdx.debug.assert(token_index.rank() == 0, "Unbatched CuTe attention requires a scalar token position", .{}); } - const mesh = zml.Compiler.current().resolveMesh(.{.model}); - const q_sharded = q.withPartitioning(mesh, .{ .h = .model }); - const k_sharded = k.withPartitioning(mesh, .{ .h = .model }); - const v_sharded = v.withPartitioning(mesh, .{ .h = .model }); const positions = if (batched) token_index.convert(.i32).broad(zml.Shape.init(.{ .b = q.dim(.b) }, .i32)) else token_index.convert(.i32); - return zml.ops.manualComputation(Context.body, Context{ .q = q_sharded, .k = k_sharded, .v = v_sharded, .positions = positions }, q_sharded.shape()); -} + const q_ = if (batched) q.transpose(.{ .b, .q, .h, .hd }) else q.transpose(.{ .q, .h, .hd }); + const k_ = if (batched) k.transpose(.{ .b, .k, .h, .hd }) else k.transpose(.{ .k, .h, .hd }); + const v_ = if (batched) v.transpose(.{ .b, .k, .h, .hd }) else v.transpose(.{ .k, .h, .hd }); -const Context = struct { - q: zml.Tensor, - k: zml.Tensor, - v: zml.Tensor, - positions: zml.Tensor, + const splits = kernel.decodeSplits(q_.shape(), k_.shape(), decodeResources(zml.Compiler.current().platform)); - fn body(self: Context, output_shape: zml.Shape) zml.Tensor { - const batched = self.q.shape().hasTag(.b) != null; - const q = if (batched) self.q.transpose(.{ .b, .q, .h, .hd }) else self.q.transpose(.{ .q, .h, .hd }); - const k = if (batched) self.k.transpose(.{ .b, .k, .h, .hd }) else self.k.transpose(.{ .k, .h, .hd }); - const v = if (batched) self.v.transpose(.{ .b, .k, .h, .hd }) else self.v.transpose(.{ .k, .h, .hd }); + const cfg: kernel.Config = .{ + .batch = if (batched) q_.dim(.b) else 1, + .query_length = q_.dim(.q), + .key_length = k_.dim(.k), + .query_heads = q_.dim(.h), + .kv_heads = k_.dim(.h), + .head_dim = q_.dim(.hd), + .dtype = zml.kernel.cute.from(q_.dtype()), + .splits = splits, + .causal = causal, + }; - const splits = kernel.decodeSplits(q.shape(), k.shape(), decodeResources(zml.Compiler.current().platform)); - - const cfg: kernel.Config = .{ - .batch = if (batched) q.dim(.b) else 1, - .query_length = q.dim(.q), - .key_length = k.dim(.k), - .query_heads = q.dim(.h), - .kv_heads = k.dim(.h), - .head_dim = q.dim(.hd), - .dtype = zml.kernel.cute.from(q.dtype()), - .splits = splits, - }; - - const out = if (cfg.splits > 1) b: { - const partials = kernel.Kernel.call(.{ .q = q, .k = k, .v = v, .positions = self.positions }, .{ - .out = zml.Shape.init(.{ cfg.batch, cfg.splits, cfg.query_length, cfg.query_heads, cfg.head_dim + 2 }, .f32), - }, .{ - .cfg = cfg, - .grid = .{ @intCast(@divTrunc(@divExact(cfg.query_heads, cfg.kv_heads) + 3, 4)), @intCast(cfg.kv_heads * cfg.batch), @intCast(cfg.splits) }, - .block = kernel.block, - }).out; - break :b kernel.Merge.call(.{ .partials = partials }, .{ .out = q.shape() }, .{ - .cfg = cfg, - .grid = .{ @intCast(cfg.query_heads), @intCast(cfg.batch), 1 }, - .block = kernel.block, - }).out; - } else kernel.Kernel.call(.{ .q = q, .k = k, .v = v, .positions = self.positions }, .{ .out = q.shape() }, .{ + const out = if (kernel.prefillSupported(cfg)) + kernel.Prefill.call(.{ .q = q_, .k = k_, .v = v_, .positions = positions }, .{ .out = q_.shape() }, .{ .cfg = cfg }).out + else if (cfg.splits > 1) b: { + const partials = kernel.Kernel.call(.{ .q = q_, .k = k_, .v = v_, .positions = positions }, .{ + .out = zml.Shape.init(.{ cfg.batch, cfg.splits, cfg.query_length, cfg.query_heads, cfg.head_dim + 2 }, .f32), + }, .{ .cfg = cfg, - .grid = .{ @intCast(@divTrunc(cfg.query_length * @divExact(cfg.query_heads, cfg.kv_heads) + 3, 4)), @intCast(cfg.kv_heads * cfg.batch), 1 }, + .grid = .{ @intCast(@divTrunc(@divExact(cfg.query_heads, cfg.kv_heads) + 3, 4)), @intCast(cfg.kv_heads * cfg.batch), @intCast(cfg.splits) }, .block = kernel.block, }).out; + break :b kernel.Merge.call(.{ .partials = partials }, .{ .out = q_.shape() }, .{ + .cfg = cfg, + .grid = .{ @intCast(cfg.query_heads), @intCast(cfg.batch), 1 }, + .block = kernel.block, + }).out; + } else kernel.Kernel.call(.{ .q = q_, .k = k_, .v = v_, .positions = positions }, .{ .out = q_.shape() }, .{ + .cfg = cfg, + .grid = .{ @intCast(@divTrunc(cfg.query_length * @divExact(cfg.query_heads, cfg.kv_heads) + 3, 4)), @intCast(cfg.kv_heads * cfg.batch), 1 }, + .block = kernel.block, + }).out; - return out.transpose(output_shape); - } -}; + return out.transpose(q.shape()); +} /// Returns the minimum number of multiprocessors and shared memory per block across all devices in the platform. /// If there are no devices, returns 0 for both values. diff --git a/zml/kernels/flash_attention/flash_attention.zig b/zml/kernels/flash_attention/flash_attention.zig index c1a27a36..3fb898db 100644 --- a/zml/kernels/flash_attention/flash_attention.zig +++ b/zml/kernels/flash_attention/flash_attention.zig @@ -6,9 +6,11 @@ pub const cute = @import("cute/cute.zig"); pub const metal = @import("metal/metal.zig"); pub const nki = @import("nki/nki.zig"); const zml = @import("../../zml.zig"); +const Options = @import("../../attention/Options.zig"); test { std.testing.refAllDecls(@This()); + _ = @import("tests.zig"); } pub fn isAvailable(platform: *const zml.Platform) bool { @@ -126,239 +128,12 @@ pub const Metadata = union(Backend) { /// - .q is the number of queries /// - .k is the number of keys /// - .hd is the head dimension -pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, metadata: Metadata, parameters: Parameters) zml.Tensor { +pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, metadata: Metadata, parameters: Parameters, opts: Options) zml.Tensor { + zml.stdx.debug.assert(zml.Compiler.current().manual_computation_depth > 0, "Attention kernels require an explicit manualComputation boundary", .{}); return switch (parameters) { - .cute => cute.attention(q, k, v, token_index), - .nki => nki.attention(q, k, v, token_index, parameters.nki), - .metal => metal.attention(q, k, v, token_index, metadata.metal), + .cute => cute.attention(q, k, v, token_index, opts.causal), + // The NKI and Metal kernels are causal only. + .nki => if (opts.causal) nki.attention(q, k, v, token_index, parameters.nki) else zml.nn.sdpa(q, k, v, .{}), + .metal => if (opts.causal) metal.attention(q, k, v, token_index, metadata.metal) else zml.nn.sdpa(q, k, v, .{}), }; } - -test "attention: q=1,qh=64,kh=8" { - try testAttention( - .init(.{ .q = 1, .h = 64, .hd = 64 }, .bf16), - .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{63}, - ); - - try testAttention( - .init(.{ .q = 1, .h = 64, .hd = 64 }, .bf16), - .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{36}, - ); -} - -test "attention: b=4,q=1,qh=64,kh=8" { - // Full attention - try testAttention( - .init(.{ .b = 4, .q = 1, .h = 64, .hd = 64 }, .bf16), - .init(.{ .b = 4, .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{ 63, 63, 63, 63 }, - ); - - // Partial attention - try testAttention( - .init(.{ .b = 4, .q = 1, .h = 64, .hd = 64 }, .bf16), - .init(.{ .b = 4, .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{ 61, 57, 23, 63 }, - ); -} - -test "attention: q=1,qh=8,kh=8" { - try testAttention( - .init(.{ .q = 1, .h = 8, .hd = 64 }, .bf16), - .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{62}, - ); - try testAttention( - .init(.{ .q = 1, .h = 8, .hd = 64 }, .bf16), - .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{63}, - ); -} - -test "attention: q=8,qh=64,kh=8" { - try testAttention( - .init(.{ .q = 8, .h = 64, .hd = 64 }, .bf16), - .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), - &.{56}, - ); -} - -pub fn testAttention(q_shape: zml.Shape, k_shape: zml.Shape, token_index_h: []const u32) !void { - return testAttentionBackends(q_shape, k_shape, token_index_h, std.enums.values(Backend)); -} - -fn testAttentionBackends(q_shape: zml.Shape, k_shape: zml.Shape, token_index_h: []const u32, backends: []const Backend) !void { - const io = std.testing.io; - const allocator = std.testing.allocator; - - const platform = zml.testing.env(); - - var arena_state: std.heap.ArenaAllocator = .init(allocator); - defer arena_state.deinit(); - const arena = arena_state.allocator(); - - const token_index_shape: zml.Shape = if (token_index_h.len > 1) b: { - std.debug.assert(token_index_h.len == q_shape.dim(.b)); - break :b .init(.{ .b = q_shape.dim(.b) }, .u32); - } else .init(.{}, .u32); - - const max_k_idx: i64 = k_shape.dim(.k) - 1; - const max_q_idx: i64 = q_shape.dim(.q) - 1; - for (token_index_h) |index| { - // Check for out of bound reads - std.debug.assert(index + max_q_idx <= max_k_idx); - } - - const tensors: struct { q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor } = .{ - .q = .fromShape(q_shape), - .k = .fromShape(k_shape), - .v = .fromShape(k_shape), - .token_index = .fromShape(token_index_shape), - }; - - var q = try attentionTestInput(q_shape, 17); - defer q.deinit(); - var k = try attentionTestInput(k_shape, 29); - defer k.deinit(); - var v = try attentionTestInput(k_shape, 43); - defer v.deinit(); - var token_index = try zml.Buffer.fromBytes(io, platform, token_index_shape, @ptrCast(token_index_h)); - defer token_index.deinit(); - - const meshes = platform.meshes.values(); - const reference_metadata: zml.attention.Metadata = .{ .stablehlo = {} }; - const reference_parameters: zml.attention.Parameters = .{ .stablehlo = {} }; - const reference_exe = try platform.compileFn(allocator, io, zml.attention.attention, .{ tensors.q, tensors.k, tensors.v, tensors.token_index, reference_metadata, reference_parameters }, .{ - .program_name = "attention_stablehlo", - .meshes = meshes, - }); - defer reference_exe.deinit(); - - var reference_d = try reference_exe.eval(allocator, io, .{ q, k, v, token_index, reference_metadata }); - defer reference_d.deinit(); - try reference_d.await(io); - const reference_h: zml.Slice = try reference_d.toSliceAlloc(allocator, io); - defer reference_h.free(allocator); - - for (backends) |backend| { - if (!backend.isAvailable(platform)) continue; - - const metadata: Metadata = .init(.fromBackend(backend)); - const parameters: Parameters = .init(.fromBackend(backend)); - const exe = try platform.compileFn( - allocator, - io, - attention, - .{ tensors.q, tensors.k, tensors.v, tensors.token_index, metadata, parameters }, - .{ - .program_name = try std.fmt.allocPrint(arena, "attention_{t}", .{backend}), - .meshes = meshes, - }, - ); - defer exe.deinit(); - - var metadata_d = try metadata.initBuffer(io, platform); - defer Metadata.deinitBuffer(&metadata_d); - - var output_d = try exe.eval(allocator, io, .{ q, k, v, token_index, metadata_d }); - defer output_d.deinit(); - try output_d.await(io); - const output_h = try output_d.toSliceAlloc(allocator, io); - defer output_h.free(allocator); - - errdefer std.log.err( - \\ Attention test failed, {0t} output doesn't match reference. - \\ - reference: {1d} - \\ - {0t}: {2d} - , .{ backend, reference_h, output_h }); - try zml.testing.expectClose(io, reference_h, output_h, .{ - .absolute_tolerance = 5e-3, - .relative_tolerance = 1e-2, - .epsilon_relative = 1e-3, - .minimum_close_fraction = 0.99, - }); - } -} - -fn attentionTestInput(shape: zml.Shape, seed: u64) !zml.Buffer { - const host = try zml.Slice.alloc(std.testing.allocator, shape); - defer host.free(std.testing.allocator); - var prng: std.Random.DefaultPrng = .init(seed); - const random = prng.random(); - switch (shape.dtype()) { - inline .bf16, .f16, .f32 => |dtype| { - for (host.items(dtype.toZigType())) |*value| { - const x = random.float(f32) * 4 - 2; - value.* = if (dtype == .bf16) .fromF32(x) else @floatCast(x); - } - }, - else => unreachable, - } - return zml.Buffer.fromSlice(std.testing.io, zml.testing.env(), host); -} - -/// Independent f64 oracle: the SDPA reference uses fast dot precision, so its -/// f32 result alone cannot establish the accuracy of the kernel's f32 arithmetic. -fn checkAttentionF32(q: zml.Slice, k: zml.Slice, v: zml.Slice, positions: []const u32, output: zml.Slice) !void { - const Layout = struct { - b: usize, - s: usize, - h: usize, - d: usize, - - fn stride(shape: zml.Shape, tag: anytype) usize { - const axis = shape.hasTag(tag) orelse return 0; - var result: usize = 1; - for (shape.dims()[axis + 1 ..]) |dim| result *= @intCast(dim); - return result; - } - - fn init(shape: zml.Shape) @This() { - return .{ .b = stride(shape, .b), .s = if (shape.hasTag(.q) != null) stride(shape, .q) else stride(shape, .k), .h = stride(shape, .h), .d = stride(shape, .hd) }; - } - - fn offset(self: @This(), b: usize, s: usize, h: usize, d: usize) usize { - return b * self.b + s * self.s + h * self.h + d * self.d; - } - }; - const ql = Layout.init(q.shape); - const kl = Layout.init(k.shape); - const vl = Layout.init(v.shape); - const batch: usize = if (q.shape.hasTag(.b) != null) @intCast(q.shape.dim(.b)) else 1; - const queries: usize = @intCast(q.shape.dim(.q)); - const heads: usize = @intCast(q.shape.dim(.h)); - const dim: usize = @intCast(q.shape.dim(.hd)); - const group: usize = @intCast(@divExact(q.shape.dim(.h), k.shape.dim(.h))); - const scores = try std.testing.allocator.alloc(f64, @intCast(k.shape.dim(.k))); - defer std.testing.allocator.free(scores); - const scale = 1.0 / @sqrt(@as(f64, @floatFromInt(dim))); - for (0..batch) |b| { - const start: usize = positions[if (positions.len == 1) 0 else b]; - for (0..queries) |i| { - for (0..heads) |h| { - var maximum: f64 = -std.math.inf(f64); - const active = scores[0 .. start + i + 1]; - for (active, 0..) |*score, j| { - var dot: f64 = 0; - for (0..dim) |d| dot += @as(f64, q.items(f32)[ql.offset(b, i, h, d)]) * @as(f64, k.items(f32)[kl.offset(b, j, h / group, d)]); - score.* = dot * scale; - maximum = @max(maximum, score.*); - } - var sum: f64 = 0; - for (active) |*score| { - score.* = @exp(score.* - maximum); - sum += score.*; - } - for (0..dim) |d| { - var expected: f64 = 0; - for (active, 0..) |weight, j| expected += weight * @as(f64, v.items(f32)[vl.offset(b, j, h / group, d)]); - expected /= sum; - const actual = output.items(f32)[ql.offset(b, i, h, d)]; - try std.testing.expectApproxEqAbs(expected, @as(f64, actual), 2e-5 + 2e-5 * @abs(expected)); - } - } - } - } -} diff --git a/zml/kernels/flash_attention/metal/BUILD.bazel b/zml/kernels/flash_attention/metal/BUILD.bazel index 855bdcca..f82ade45 100644 --- a/zml/kernels/flash_attention/metal/BUILD.bazel +++ b/zml/kernels/flash_attention/metal/BUILD.bazel @@ -2,5 +2,8 @@ package(default_visibility = ["//zml:__pkg__"]) filegroup( name = "metal", - srcs = ["metal.zig"], + srcs = [ + "flash_attn.zig", + "metal.zig", + ], ) diff --git a/zml/kernels/flash_attention/metal/flash_attn.zig b/zml/kernels/flash_attention/metal/flash_attn.zig new file mode 100644 index 00000000..de4b68da --- /dev/null +++ b/zml/kernels/flash_attention/metal/flash_attn.zig @@ -0,0 +1,732 @@ +//! Metal flash attention over contiguous `[batch, keys, kv_heads, head_dim]` +//! K/V, causal from `token_index`: fa_vec and fa_vec_reduce for decode, +//! fa_tiled otherwise. + +const std = @import("std"); + +const metal = @import("../../../kernel/kernel.zig").metal; + +const ArgSpec = metal.ArgSpec; +const Builder = metal.Builder; +const Type = metal.Type; +const Value = metal.Value; + +const flt_max = std.math.floatMax(f32); +/// A masked score: below any real one yet finite, since fast math assumes +/// no infinities. Halved so subtracting a running maximum cannot overflow. +const masked_score = -flt_max / 2.0; +const max_half = 65504.0; +const simd_width = 32; + +/// fa_vec: bf16 decode, one query per batch row. Simdgroup s takes query head +/// s % heads of the threadgroup's KV head group and position split s / heads. +pub const FaVec = struct { + pub const Cfg = struct { + head_dim: i32, + heads: i32, + splits: i32, + }; + + pub const Args: metal.Struct = .{ .name = "fa_vec_args", .fields = &.{ + .{ .name = "num_heads", .type = .int }, + .{ .name = "n_kv", .type = .int }, + .{ .name = "k_len", .type = .int }, + .{ .name = "scale", .type = .float }, + } }; + + pub const HostArgs = extern struct { + num_heads: i32, + n_kv: i32, + k_len: i32, + scale: f32, + }; + + pub const Kernel = metal.Kernel(Cfg, .{ + .name = "fa_vec", + .inputs = &.{ "q", "k", "v", "token_index" }, + .outputs = &.{"dst"}, + .constants = &.{"args"}, + .run = run, + }); + + fn run(b: *Builder, cfg: Cfg) metal.FinishError!void { + switch (cfg.head_dim) { + 32, 64, 128, 256, 512 => {}, + else => std.debug.panic("fa_vec: head_dim {d} unsupported (32, 64, 128, 256 or 512)", .{cfg.head_dim}), + } + const char_in: ArgSpec = .{ .buffer = .{ .type = .char, .const_ = true } }; + const a = try b.declareArgs(.{ + .args = .{ .constant_struct = Args }, + .q = char_in, + .k = char_in, + .v = char_in, + .token_index = .{ .buffer = .{ .type = .int, .const_ = true } }, + .dst = .{ .buffer = .{ .type = .char } }, + .tgpig = .{ .threadgroup_position_in_grid = .uint3 }, + .tgpg = .{ .threadgroups_per_grid = .uint3 }, + .tiisg = .{ .thread_index_in_simdgroup = .ushort }, + .sgitg = .{ .simdgroup_index_in_threadgroup = .ushort }, + .nsg = .{ .simdgroups_per_threadgroup = .ushort }, + }); + impl(b, cfg, a); + } + + fn impl(b: *Builder, cfg: Cfg, a: anytype) void { + const D = cfg.head_dim; + const D4 = @divExact(D, 4); + // A K/V row spans NL lanes of NI float4s; a simdgroup reads R rows per step. + const NL = @min(D4, simd_width); + const R = @divExact(simd_width, NL); + const NI = @divExact(D4, NL); + const C = simd_width; + const HG = cfg.heads; + const NWG = cfg.splits; + const full: metal.LoopOpts = .{ .unroll = .full }; + + const num_heads = a.args.field("num_heads"); + const n_kv = a.args.field("n_kv"); + const group = num_heads.div(n_kv); + const groups = group.div(HG); + const r = a.tgpig.at(2).to(.int); + const tg_head = a.tgpig.at(1).to(.int); + const ikv = tg_head.div(groups); + const iq2 = ikv.mul(group).add(tg_head.rem(groups).mul(HG)).add(a.sgitg.rem(HG).to(.int)); + const iwg = a.tgpig.at(0).to(.int).mul(a.nsg.div(HG).to(.int)).add(a.sgitg.div(HG).to(.int)); + const rid = r.to(.long).mul(num_heads.to(.uint)).add(iq2); + const nrows = a.tgpg.at(2).to(.long).mul(num_heads.to(.uint)); + + const kv_len = a.token_index.load(r).add(1); + const live = if (NWG > 1) liveSplits(b, kv_len, NWG) else b.lift(@as(i32, 1)); + + if (NWG > 1) { + var idle = b.openIf(iwg.ge(live)); + { + b.return_(); + idle.close(); + } + } + + const position_bytes = n_kv.to(.ulong).mul(D * 2); + const first = r.to(.ulong).mul(a.args.field("k_len").to(.ulong)).mul(position_bytes).add(ikv.to(.ulong).mul(D * 2)); + const k0 = a.k.add(first); + const v0 = a.v.add(first); + + const lane = if (R == 1) a.tiisg else a.tiisg.rem(NL); + const sub = a.tiisg.div(NL).to(.int); + const q4 = a.q.add(rid.mul(D).mul(2)).to(.bfloat4).add(lane); + const qr = b.array(.float4, @intCast(NI)); + const o = b.array(.float4, @intCast(NI)); + var init = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + const ii = init.iv; + qr.store(ii, q4.load(ii.mul(NL)).to(.float4)); + o.store(ii, b.constant(.float4, 0)); + init.close(); + } + + const S = b.local(.float, 0.0); + const M = b.local(.float, -flt_max / 2.0); + + const ic0 = b.local(.int, iwg); + var blocks = b.openLoop(.{}); + { + const ic = ic0.mul(C); + var past = b.openIf(ic.ge(kv_len)); + { + blocks.breakLoop(); + past.close(); + } + + const s = b.local(.float, 0.0); + var qk = b.openForOpts(b.constant(.short, 0), @divExact(C, R), 1, full); + { + const cc = qk.iv; + const p = b.min(if (R == 1) ic.add(cc) else ic.add(cc.to(.int).mul(R)).add(sub), kv_len.sub(1)); + const pk4 = k0.add(p.to(.ulong).mul(position_bytes)).to(.bfloat4).add(lane); + const d = b.local(.float, 0.0); + var dots = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + const ii = dots.iv; + d.assign(d.add(b.dot(pk4.load(ii.mul(NL)).to(.float4), qr.load(ii)))); + dots.close(); + } + if (R == 1) { + s.assign(b.ternary(a.tiisg.to(.short).eq(cc), b.simdSum(d.rvalue()), s.rvalue())); + } else { + var x = d.rvalue(); + var mask: i32 = 1; + while (mask < NL) : (mask *= 2) x = x.add(b.simdShuffleXor(x, b.constant(.ushort, mask))); + const t = b.simdShuffle(x, a.tiisg.rem(R).mul(NL)); + s.assign(b.ternary(a.tiisg.div(R).to(.short).eq(cc), t, s.rvalue())); + } + qk.close(); + } + + const pos = ic.add(a.tiisg.to(.int)); + const score = b.ternary(pos.lt(kv_len), s.mul(a.args.field("scale")), b.constant(.float, -max_half)); + + const m = M.rvalue(); + M.assign(b.simdMax(b.max(M.rvalue(), score))); + const ms = b.exp(m.sub(M)); + const vs = b.exp(score.sub(M)); + S.assign(S.mul(ms).add(b.simdSum(vs))); + + var rescale = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + o.store(rescale.iv, o.load(rescale.iv).mul(ms)); + rescale.close(); + } + + var pv = b.openForOpts(b.constant(.short, 0), @divExact(C, R), 1, full); + { + const cc = pv.iv; + const row = cc.to(.int).mul(R).add(sub); + const p = b.min(if (R == 1) ic.add(cc) else ic.add(row), kv_len.sub(1)); + const pv4 = v0.add(p.to(.ulong).mul(position_bytes)).to(.bfloat4).add(lane); + const w = b.vec(.float4, .{if (R == 1) b.simdBroadcast(vs, cc.to(.ushort)) else b.simdShuffle(vs, row)}); + var dims = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + const ii = dims.iv; + o.store(ii, o.load(ii).add(pv4.load(ii.mul(NL)).to(.float4).mul(w))); + dims.close(); + } + pv.close(); + } + + blocks.next(); + ic0.assign(ic0.add(live)); + blocks.close(); + } + + var mask: i32 = NL; + while (mask < simd_width) : (mask *= 2) { + var sum = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + o.store(sum.iv, o.load(sum.iv).add(b.simdShuffleXor(o.load(sum.iv), b.constant(.ushort, mask)))); + sum.close(); + } + } + var first_row = if (R > 1) b.openIf(sub.eq(0)) else null; + if (NWG == 1) { + const inv_sum = b.ternary(S.eq(0.0), 0.0, b.lift(1.0).div(S)); + const dst4 = a.dst.to(.bfloat4).add(rid.mul(D4)).add(lane); + var out = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + dst4.store(out.iv.mul(NL), o.load(out.iv).mul(inv_sum).to(.bfloat4)); + out.close(); + } + } else { + // [rows][D4][splits] float4 partials, then [rows][splits] (S, M). + const dst4 = a.dst.to(.float4).add(rid.mul(D4).mul(NWG)); + var out = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + const i = lane.to(.int).add(out.iv.to(.int).mul(NL)); + dst4.store(i.mul(NWG).add(iwg), o.load(out.iv)); + out.close(); + } + var lane0 = b.openIf(a.tiisg.eq(0)); + { + const dst1 = a.dst.to(.float).add(nrows.mul(D).mul(NWG)).add(rid.mul(2 * NWG)); + dst1.store(b.lift(2).mul(iwg), S); + dst1.store(b.lift(2).mul(iwg).add(1), M); + lane0.close(); + } + } + if (first_row) |*scope| scope.close(); + } +}; + +fn liveSplits(b: *Builder, kv_len: Value, splits: i32) Value { + return b.min(splits, kv_len.add(simd_width - 1).div(simd_width)); +} + +/// fa_vec_reduce: merges the fa_vec split partials. +pub const FaVecReduce = struct { + pub const Cfg = struct { + splits: i32, + head_dim: i32, + }; + + pub const simdgroups = 8; + + pub const Kernel = metal.Kernel(Cfg, .{ + .name = "fa_vec_reduce", + .inputs = &.{ "htmp", "token_index" }, + .outputs = &.{"dst"}, + .constants = &.{"args"}, + .run = run, + }); + + fn run(b: *Builder, cfg: Cfg) metal.FinishError!void { + const a = try b.declareArgs(.{ + .args = .{ .constant_struct = FaVec.Args }, + .htmp = .{ .buffer = .{ .type = .char, .const_ = true } }, + .token_index = .{ .buffer = .{ .type = .int, .const_ = true } }, + .dst = .{ .buffer = .{ .type = .char } }, + .tgpig = .{ .threadgroup_position_in_grid = .uint2 }, + .tgpg = .{ .threadgroups_per_grid = .uint2 }, + .tiisg = .{ .thread_index_in_simdgroup = .ushort }, + .sgitg = .{ .simdgroup_index_in_threadgroup = .ushort }, + }); + const NWG = cfg.splits; + if (NWG > 4 * simd_width) std.debug.panic("fa_vec_reduce: {d} splits, at most {d}", .{ NWG, 4 * simd_width }); + const NP = @divFloor(NWG + simd_width - 1, simd_width); + const DV = cfg.head_dim; + const DV4 = @divExact(DV, 4); + const full: metal.LoopOpts = .{ .unroll = .full }; + + const rid = a.tgpig.at(0).to(.ulong); + const r = a.tgpig.at(0).div(a.args.field("num_heads").to(.uint)).to(.int); + const i = a.tgpig.at(1).to(.int).mul(simdgroups).add(a.sgitg.to(.int)); + const dst4 = a.dst.to(.bfloat4).add(rid.mul(DV4)); + const live = liveSplits(b, a.token_index.load(r).add(1), NWG); + + const ss = a.htmp.to(.float).add(a.tgpg.at(0).to(.ulong).mul(DV).mul(NWG)).add(rid.mul(2 * NWG)); + const Mj = b.array(.float, @intCast(NP)); + const Sj = b.array(.float, @intCast(NP)); + const mx = b.local(.float, -flt_max / 2.0); + var load = b.openForOpts(b.constant(.short, 0), NP, 1, full); + { + const j = load.iv; + const idx = a.tiisg.to(.int).add(j.to(.int).mul(simd_width)); + const ok = idx.lt(live); + Mj.store(j, b.ternary(ok, ss.load(idx.mul(2).add(1)), b.constant(.float, -flt_max / 2.0))); + Sj.store(j, b.ternary(ok, ss.load(idx.mul(2)), b.constant(.float, 0))); + mx.assign(b.max(mx.rvalue(), Mj.load(j))); + load.close(); + } + const m = b.simdMax(mx.rvalue()); + const sum = b.local(.float, 0.0); + var weigh = b.openForOpts(b.constant(.short, 0), NP, 1, full); + { + const j = weigh.iv; + Mj.store(j, b.exp(Mj.load(j).sub(m))); + sum.assign(sum.add(Sj.load(j).mul(Mj.load(j)))); + weigh.close(); + } + const S = b.simdSum(sum.rvalue()); + const inv_sum = b.ternary(S.eq(0.0), 0.0, b.lift(1.0).div(S)); + + const htmp4 = a.htmp.to(.float4).add(rid.mul(DV4).mul(NWG)); + const acc = b.local(.float4, b.constant(.float4, 0)); + var parts = b.openForOpts(b.constant(.short, 0), NP, 1, full); + { + const j = parts.iv; + const idx = a.tiisg.to(.int).add(j.to(.int).mul(simd_width)); + var ok = b.openIf(idx.lt(live)); + { + acc.assign(acc.add(htmp4.load(i.mul(NWG).add(idx)).mul(Mj.load(j)))); + ok.close(); + } + parts.close(); + } + const v = b.simdSum(acc.rvalue()); + var lane0 = b.openIf(a.tiisg.eq(0)); + { + dst4.store(i, v.mul(inv_sum).to(.bfloat4)); + lane0.close(); + } + } +}; + +/// fa_tiled: causal FlashAttention-2 over simdgroup matrices. Queries past +/// `live` output zeros. +pub const FaTiled = struct { + pub const Cfg = struct { + dtype: Type, + head_size: i32, + bq: i32 = 32, + tile_kv: i32 = 32, + num_threads: i32 = 128, + }; + + pub const Kernel = metal.Kernel(Cfg, .{ + .name = "fa_tiled", + .inputs = &.{ "q", "k", "v", "token_index", "live" }, + .outputs = &.{"out"}, + .constants = &.{ "num_kv_heads", "scale", "q_len", "k_len" }, + .run = run, + }); + + fn supported(cfg: Cfg) bool { + const tile = tileConfig(cfg.head_size) orelse return false; + return (cfg.dtype.eql(.half) or cfg.dtype.eql(.bfloat)) and + cfg.bq == tile[0] and cfg.tile_kv == tile[1] and cfg.num_threads == tile[2]; + } + + pub fn tileConfig(head_size: i64) ?[3]i32 { + return switch (head_size) { + 32, 64, 96, 128 => .{ 32, 32, 128 }, + 256 => .{ 16, 16, 64 }, + 512 => .{ 8, 8, 32 }, + else => null, + }; + } + + pub fn threadgroupMemory(cfg: Cfg) u64 { + const ld = cfg.head_size + @divExact(16, @as(i32, @intCast(cfg.dtype.size()))); + return @intCast((cfg.bq + 2 * cfg.tile_kv) * ld * @as(i32, @intCast(cfg.dtype.size()))); + } + + /// Lane `lane_id`'s (column, row) in an 8x8 simdgroup_matrix fragment. + fn fragCoord(b: *Builder, lane_id: Value) Value { + const qid = lane_id.to(.short).div(4).to(.short); + const fm = qid.bitAnd(4).add(lane_id.to(.short).div(2).rem(4)).to(.short); + const fn_ = qid.bitAnd(2).mul(2).add(lane_id.to(.short).rem(2).mul(2)).to(.short); + return b.vec(.short2, .{ fn_, fm }); + } + + fn fragRowReduce(b: *Builder, comptime op: enum { max, sum }, v: Value) Value { + var x = v.rvalue(); + inline for (.{ 1, 8 }) |mask| { + const y = b.simdShuffleXor(x, b.constant(.ushort, mask)); + x = switch (op) { + .max => b.max(x, y), + .sum => x.add(y), + }; + } + return x; + } + + fn setLane(p: Value, i: Value, lane: anytype, x: Value) void { + p.store(i, p.load(i).insert(lane, x)); + } + + fn run(b: *Builder, cfg: Cfg) metal.FinishError!void { + if (!supported(cfg)) std.debug.panic("fa_tiled: {s} at head size {d} and tiles {d}, {d}, {d} unsupported", .{ cfg.dtype.name(), cfg.head_size, cfg.bq, cfg.tile_kv, cfg.num_threads }); + const T = cfg.dtype; + const HEAD_SIZE = cfg.head_size; + const BQ = cfg.bq; + const TILE_KV = cfg.tile_kv; + const NUM_THREADS = cfg.num_threads; + const int_arg: ArgSpec = .{ .constant = .int }; + const int_in: ArgSpec = .{ .buffer = .{ .type = .int, .const_ = true } }; + const a = try b.declareArgs(.{ + .out = .{ .buffer = .{ .type = T } }, + .q = .{ .buffer = .{ .type = T, .const_ = true } }, + .k = .{ .buffer = .{ .type = T, .const_ = true } }, + .v = .{ .buffer = .{ .type = T, .const_ = true } }, + .token_index = int_in, + .live = int_in, + .num_kv_heads = int_arg, + .scale = .{ .constant = .float }, + .q_len = int_arg, + .k_len = int_arg, + .shared_mem = .{ .threadgroup = .char }, + .tgp = .{ .threadgroup_position_in_grid = .uint3 }, + .tgpg = .{ .threadgroups_per_grid = .uint3 }, + .tpt = .{ .thread_position_in_threadgroup = .uint3 }, + .sg_idx = .{ .simdgroup_index_in_threadgroup = .uint }, + .lane = .{ .thread_index_in_simdgroup = .uint }, + }); + + const NUM_SIMD_LANES = 32; + const NUM_SG = @divExact(NUM_THREADS, NUM_SIMD_LANES); + const TD = @divExact(HEAD_SIZE, 8); + const TK = @divExact(TILE_KV, 8); + if (@divExact(BQ, NUM_SG) != 8) @panic("fa_tiled: ROWS_PER_SG must be 8"); + const unroll: metal.LoopOpts = .{ .unroll = .enable }; + + const thread_idx = a.tpt.x().to(.int); + const head_idx = a.tgp.x().to(.int); + const q_pos_start = a.tgp.y().to(.int).mul(BQ); + const batch = a.tgp.z().to(.int); + const num_heads = a.tgpg.x().to(.int); + const kv_head_idx = head_idx.div(num_heads.div(a.num_kv_heads)); + const q_stride = num_heads.mul(HEAD_SIZE); + const kv_token_stride = a.num_kv_heads.mul(HEAD_SIZE); + + const block_rows = b.min(BQ, a.q_len.sub(q_pos_start)); + const live = a.live.load(0); + const valid_q = b.min(block_rows, b.max(0, live.sub(q_pos_start))); + const out_base = a.out.add(batch.mul(a.q_len).add(q_pos_start).to(.long).mul(q_stride).add(head_idx.mul(HEAD_SIZE))); + + var dead = b.openIf(valid_q.le(0)); + { + var zero = b.openFor(thread_idx, block_rows.mul(HEAD_SIZE), NUM_THREADS); + { + const i = zero.iv; + out_base.store(i.div(HEAD_SIZE).mul(q_stride).add(i.rem(HEAD_SIZE)), b.constant(T, 0)); + zero.close(); + } + b.return_(); + dead.close(); + } + + // Query i of the batch row is at position context_len + i. + const context_len = a.token_index.load(batch); + const seq_len = context_len.add(live); + const kv_base = batch.to(.long).mul(a.k_len).mul(kv_token_stride).add(kv_head_idx.mul(HEAD_SIZE)); + + const SMEM_PAD = @divExact(16, @as(i32, @intCast(T.size()))); + const LD = HEAD_SIZE + SMEM_PAD; + const Q_ELEMS = BQ * LD; + const KV_ELEMS = TILE_KV * LD; + + const Q_smem = a.shared_mem.to(T); + const K_smem = Q_smem.add(Q_ELEMS); + const V_smem = K_smem.add(KV_ELEMS); + + const log2e: f32 = std.math.log2e; + const scale_log2 = a.scale.mul(log2e); + + const q_base = a.q.add(batch.mul(a.q_len).add(q_pos_start).to(.long).mul(q_stride).add(head_idx.mul(HEAD_SIZE))); + var load_q = b.openFor(thread_idx, BQ * HEAD_SIZE, NUM_THREADS); + { + const i = load_q.iv; + const r = i.div(HEAD_SIZE); + const d = i.rem(HEAD_SIZE); + var value = b.openIfElse(r.lt(valid_q), .{T}); + { + value.yieldThen(.{q_base.load(r.mul(q_stride).add(d))}); + } + { + value.yieldElse(.{b.constant(T, 0)}); + } + Q_smem.store(r.mul(LD).add(d), value.results[0]); + load_q.close(); + } + + b.threadgroupBarrier(.mem_threadgroup); + + const T2: Type = .vec(T.scalar, 2); + const q_reg = b.array(T2, @intCast(TD)); + var load_q_reg = b.openForOpts(0, TD, 1, unroll); + { + const d = load_q_reg.iv; + const tmp = b.simdgroupLoad(Q_smem.add(a.sg_idx.mul(8).mul(LD)).add(d.mul(8)), LD, .{}); + q_reg.store(d, tmp.threadElements()); + load_q_reg.close(); + } + + const fc = fragCoord(b, a.lane.to(.ushort)); + const fm = fc.y(); + const fn_ = fc.x(); + + const max_score = b.local(.float, masked_score); + const sum_score = b.local(.float, 0.0); + + const o_reg = b.array(.float2, @intCast(TD)); + var zero_o = b.openForOpts(0, TD, 1, unroll); + { + o_reg.store(zero_o.iv, b.constant(.float2, 0)); + zero_o.close(); + } + + const num_kv_tiles = seq_len.add(TILE_KV).sub(1).div(TILE_KV); + const last_q_pos = context_len.add(q_pos_start).add(valid_q).sub(1); + + var tiles = b.openFor(0, num_kv_tiles, 1); + { + const tile_idx = tiles.iv; + const tile_start = tile_idx.mul(TILE_KV); + + var past = b.openIf(tile_start.gt(last_q_pos)); + { + tiles.breakLoop(); + past.close(); + } + + const VEC_BYTES: i32 = if (HEAD_SIZE >= 256) 16 else 8; + const VEC = @divExact(VEC_BYTES, @as(i32, @intCast(T.size()))); + const LoadUnit: Type = if (VEC_BYTES == 16) .uint4 else .uint2; + std.debug.assert(@rem(LD, VEC) == 0); + var rows = b.openForOpts(0, @divExact(TILE_KV, NUM_SG), 1, unroll); + { + const t = a.sg_idx.mul(@divExact(TILE_KV, NUM_SG)).add(rows.iv).to(.int); + const kv_pos = tile_start.add(t); + var present = b.openIfElse(kv_pos.lt(seq_len), .{}); + { + const off = kv_base.add(kv_pos.to(.long).mul(kv_token_stride)); + const k_ptr = a.k.add(off); + const v_ptr = a.v.add(off); + var loop = b.openForOpts(a.lane.mul(VEC).to(.int), HEAD_SIZE, NUM_SIMD_LANES * VEC, unroll); + { + const d = loop.iv; + K_smem.add(t.mul(LD).add(d)).to(LoadUnit).store(0, k_ptr.add(d).to(LoadUnit).load(0)); + V_smem.add(t.mul(LD).add(d)).to(LoadUnit).store(0, v_ptr.add(d).to(LoadUnit).load(0)); + loop.close(); + } + present.elseBranch(); + } + { + const zero = b.constant(LoadUnit, 0); + var loop = b.openForOpts(a.lane.mul(VEC).to(.int), HEAD_SIZE, NUM_SIMD_LANES * VEC, unroll); + { + const d = loop.iv; + K_smem.add(t.mul(LD).add(d)).to(LoadUnit).store(0, zero); + V_smem.add(t.mul(LD).add(d)).to(LoadUnit).store(0, zero); + loop.close(); + } + present.close(); + } + rows.close(); + } + + b.threadgroupBarrier(.mem_threadgroup); + + const s_reg = b.array(.float2, @intCast(TK)); + var zero_s = b.openForOpts(0, TK, 1, unroll); + { + s_reg.store(zero_s.iv, b.constant(.float2, 0)); + zero_s.close(); + } + + var qk_dims = b.openForOpts(0, TD, 1, unroll); + { + const d = qk_dims.iv; + const q_frag = b.matrix(T.scalar, q_reg.load(d)); + var qk_keys = b.openForOpts(0, TK, 1, unroll); + { + const k = qk_keys.iv; + const k_frag = b.simdgroupLoad(K_smem.add(k.mul(8).mul(LD)).add(d.mul(8)), LD, .{ .transpose = true }); + const s_frag = b.matrix(.float, s_reg.load(k)); + s_reg.store(k, b.simdgroupMultiplyAccumulate(q_frag, k_frag, s_frag).threadElements()); + qk_keys.close(); + } + qk_dims.close(); + } + + const q_abs_pos = context_len.add(q_pos_start).add(a.sg_idx.mul(8)).add(fm).to(.int); + const row_masked = a.sg_idx.mul(8).add(fm).ge(valid_q); + const tile_no_mask = tile_start.add(TILE_KV).le(seq_len).bitAnd(tile_start.add(TILE_KV).sub(1).lt(context_len.add(q_pos_start))); + + var masking = b.openIfElse(tile_no_mask.bitAnd(row_masked.eq(false)), .{}); + { + var scale = b.openForOpts(0, TK, 1, unroll); + { + const k = scale.iv; + setLane(s_reg, k, 0, s_reg.load(k).at(0).mul(scale_log2)); + setLane(s_reg, k, 1, s_reg.load(k).at(1).mul(scale_log2)); + scale.close(); + } + masking.elseBranch(); + } + { + var keys = b.openForOpts(0, TK, 1, unroll); + { + const k = keys.iv; + var pair = b.openForOpts(0, 2, 1, unroll); + { + const jj = pair.iv; + const s = s_reg.load(k).at(jj).mul(scale_log2); + const kv_pos = tile_start.add(k.mul(8)).add(fn_).add(jj); + const masked = row_masked.bitOr(kv_pos.ge(seq_len)).bitOr(kv_pos.gt(q_abs_pos)); + setLane(s_reg, k, jj, b.ternary(masked, masked_score, s)); + pair.close(); + } + keys.close(); + } + masking.close(); + } + + const local_max = b.local(.float, masked_score); + var max_keys = b.openForOpts(0, TK, 1, unroll); + { + const k = max_keys.iv; + local_max.assign(b.max(local_max.rvalue(), b.max(s_reg.load(k).at(0), s_reg.load(k).at(1)))); + max_keys.close(); + } + const row_max = fragRowReduce(b, .max, local_max); + + const prev_max = max_score.rvalue(); + const grown_max = b.max(prev_max, row_max); + const grew = grown_max.gt(prev_max); + const was_empty = prev_max.eq(masked_score); + const factor = b.ternary(grew, b.ternary(was_empty, 0.0, b.fast().exp2(prev_max.sub(grown_max))), 1.0); + const new_max = b.ternary(grew.eq(false).bitAnd(was_empty), 0.0, grown_max); + // A tile masking a whole row leaves the maximum masked: zero would + // pin it, and later scores far below zero would underflow. + max_score.assign(grown_max); + + const local_sum = b.local(.float, 0.0); + var p_keys = b.openForOpts(0, TK, 1, unroll); + { + const k = p_keys.iv; + var pair = b.openForOpts(0, 2, 1, unroll); + { + const jj = pair.iv; + const s = s_reg.load(k).at(jj); + const p = b.ternary(s.eq(masked_score), 0.0, b.fast().exp2(s.sub(new_max))); + setLane(s_reg, k, jj, p); + local_sum.assign(local_sum.add(p)); + pair.close(); + } + p_keys.close(); + } + const row_sum = fragRowReduce(b, .sum, local_sum); + + sum_score.assign(sum_score.mul(factor).add(row_sum)); + + var rescale_o = b.openForOpts(0, TD, 1, unroll); + { + o_reg.store(rescale_o.iv, o_reg.load(rescale_o.iv).mul(factor)); + rescale_o.close(); + } + + var pv_keys = b.openForOpts(0, TK, 1, unroll); + { + const k = pv_keys.iv; + const p_frag = b.matrix(.float, s_reg.load(k)); + var pv_dims = b.openForOpts(0, TD, 1, unroll); + { + const d = pv_dims.iv; + const v_frag = b.simdgroupLoad(V_smem.add(k.mul(8).mul(LD)).add(d.mul(8)), LD, .{}); + const o_frag = b.matrix(.float, o_reg.load(d)); + o_reg.store(d, b.simdgroupMultiplyAccumulate(p_frag, v_frag, o_frag).threadElements()); + pv_dims.close(); + } + pv_keys.close(); + } + + b.threadgroupBarrier(.mem_threadgroup); + tiles.close(); + } + + const inv_sum = b.lift(1.0).div(sum_score.add(@as(f32, 1e-6))); + var normalize = b.openForOpts(0, TD, 1, unroll); + { + o_reg.store(normalize.iv, o_reg.load(normalize.iv).mul(inv_sum)); + normalize.close(); + } + + const O_smem = Q_smem.to(.float); + var store_o = b.openForOpts(0, TD, 1, unroll); + { + const d = store_o.iv; + const o_frag = b.matrix(.float, o_reg.load(d)); + b.simdgroupStore(o_frag, O_smem.add(a.sg_idx.mul(8).mul(LD)).add(d.mul(8)), LD, .{}); + store_o.close(); + } + + b.threadgroupBarrier(.mem_threadgroup); + + var write_out = b.openFor(thread_idx, block_rows.mul(HEAD_SIZE), NUM_THREADS); + { + const i = write_out.iv; + const r = i.div(HEAD_SIZE); + const d = i.rem(HEAD_SIZE); + const value = b.ternary(r.lt(valid_q), O_smem.load(r.mul(LD).add(d)).to(T), b.constant(T, 0)); + out_base.store(r.mul(q_stride).add(d), value); + write_out.close(); + } + } +}; + +test "the flash attention kernels emit verified modules" { + const allocator = std.testing.allocator; + for ([_]i32{ 32, 64, 128, 256, 512 }) |head_dim| { + for ([_][2]i32{ .{ 1, 1 }, .{ 4, 1 }, .{ 2, 8 }, .{ 4, 128 } }) |launch| { + allocator.free(try FaVec.Kernel.emit(allocator, .{ .head_dim = head_dim, .heads = launch[0], .splits = launch[1] })); + } + for ([_]i32{ 8, 128 }) |splits| { + allocator.free(try FaVecReduce.Kernel.emit(allocator, .{ .splits = splits, .head_dim = head_dim })); + } + } + for ([_]Type{ .half, .bfloat }) |dtype| { + for ([_]i32{ 32, 64, 96, 128, 256, 512 }) |head_size| { + const tile = FaTiled.tileConfig(head_size).?; + allocator.free(try FaTiled.Kernel.emit(allocator, .{ .dtype = dtype, .head_size = head_size, .bq = tile[0], .tile_kv = tile[1], .num_threads = tile[2] })); + } + } +} diff --git a/zml/kernels/flash_attention/metal/metal.zig b/zml/kernels/flash_attention/metal/metal.zig index 9327c2a1..bdfb00c0 100644 --- a/zml/kernels/flash_attention/metal/metal.zig +++ b/zml/kernels/flash_attention/metal/metal.zig @@ -1,12 +1,10 @@ const std = @import("std"); -const stdx = @import("stdx"); - const zml = @import("../../../zml.zig"); - -const log = std.log.scoped(.@"zml/attention/metal"); +const flash_attn = @import("flash_attn.zig"); pub const Metadata = struct { + /// Live queries of a single-sequence prefill; 0 means all. num_tokens: zml.Tensor, pub fn init() Metadata { @@ -23,16 +21,123 @@ pub const Metadata = struct { } }; +/// Causal attention: fa_vec for bf16 decode, fa_tiled otherwise, sdpa for +/// unsupported dtypes and head sizes. pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, metadata: Metadata) zml.Tensor { - const qc = q.transpose(.{ .h, .q, .hd }); - const kc = k.transpose(.{ .h, .k, .hd }); - const vc = v.transpose(.{ .h, .k, .hd }); - const tok_i32 = token_index.convert(.i32); + const dtype = q.dtype(); + const head_dim = q.dim(.hd); + if (!((dtype == .bf16 or dtype == .f16) and k.dtype() == dtype and v.dtype() == dtype and flash_attn.FaTiled.tileConfig(head_dim) != null)) { + const seq_len = k.dim(.k); + const mask = zml.nn.causalAttnMask(.{ .q = seq_len, .k = seq_len }, dtype, null) + .gatherSlices(zml.Shape.init(.{ .q = q.dim(.q) }, dtype), token_index.appendAxes(.{.coord}), .{}); + return zml.nn.sdpa(q, k, v, .{ .attn_mask = mask }); + } - const attn = if (q.dim(.q) > 1) - zml.ops.customCall("zml$flash_attn", .{ qc, kc, vc, tok_i32, metadata.num_tokens }, qc.shape(), .{}, .{ .has_side_effect = false }) + const batched = q.shape().hasTag(.b) != null; + const qb = if (batched) q.transpose(.{ .b, .q, .h, .hd }) else q.transpose(.{ .q, .h, .hd }).insertAxes(0, .{.b}); + const kb = if (batched) k.transpose(.{ .b, .k, .h, .hd }) else k.transpose(.{ .k, .h, .hd }).insertAxes(0, .{.b}); + const vb = if (batched) v.transpose(.{ .b, .k, .h, .hd }) else v.transpose(.{ .k, .h, .hd }).insertAxes(0, .{.b}); + const positions = token_index.convert(.i32).broad(.init(.{ .b = qb.dim(.b) }, .i32)); + const scale = 1.0 / @sqrt(@as(f32, @floatFromInt(head_dim))); + + const vec_head_dim = head_dim == 32 or head_dim == 64 or head_dim == 128 or head_dim == 256 or head_dim == 512; + const out = if (qb.dim(.q) == 1 and dtype == .bf16 and vec_head_dim) + decode(qb, kb, vb, positions, scale) else - zml.ops.customCall("zml$flash_attn", .{ qc, kc, vc, tok_i32 }, qc.shape(), .{}, .{ .has_side_effect = false }); + prefill(qb, kb, vb, positions, liveQueries(qb, metadata), scale); + return (if (batched) out else out.squeeze(.b)).transpose(q.shape()); +} - return attn.transpose(q.shape()); +fn liveQueries(q: zml.Tensor, metadata: Metadata) zml.Tensor { + const all: zml.Tensor = .scalar(q.dim(.q), .i32); + if (q.dim(.b) != 1 or q.dim(.q) == 1) return all; + const num_tokens = metadata.num_tokens.convert(.i32); + return num_tokens.cmp(.GE, .scalar(1, .i32)).select(num_tokens.minimum(all), all); +} + +/// `q` is `[b, 1, h, hd]`, `k` and `v` `[b, k, kv_h, hd]`. +fn decode(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, positions: zml.Tensor, scale: f32) zml.Tensor { + const batch: i32 = @intCast(q.dim(.b)); + const num_heads: i32 = @intCast(q.dim(.h)); + const head_dim: i32 = @intCast(q.dim(.hd)); + const num_kv_heads: i32 = @intCast(k.dim(.h)); + const keys: i32 = @intCast(k.dim(.k)); + const nrows = batch * num_heads; + const launch = vecLaunch(nrows, @divExact(num_heads, num_kv_heads), keys, coreCount()); + const args: flash_attn.FaVec.HostArgs = .{ .num_heads = num_heads, .n_kv = num_kv_heads, .k_len = keys, .scale = scale }; + + const dst = if (launch.splits == 1) q.shape() else zml.Shape.init(.{nrows * (head_dim * launch.splits + 2 * launch.splits)}, .f32); + const out = flash_attn.FaVec.Kernel.call(.{ .q = q, .k = k, .v = v, .token_index = positions }, .{ .dst = dst }, .{ + .cfg = .{ .head_dim = head_dim, .heads = launch.heads, .splits = launch.splits }, + .grid = .{ @intCast(@divExact(launch.splits, launch.splits_per_threadgroup)), @intCast(@divExact(num_heads, launch.heads)), @intCast(batch) }, + .threadgroup = .{ 32, @intCast(launch.heads * launch.splits_per_threadgroup), 1 }, + .constants = .{ .args = .of(args) }, + }).dst; + if (launch.splits == 1) return out; + + return flash_attn.FaVecReduce.Kernel.call(.{ .htmp = out, .token_index = positions }, .{ .dst = q.shape() }, .{ + .cfg = .{ .splits = launch.splits, .head_dim = head_dim }, + .grid = .{ @intCast(nrows), @intCast(@divExact(head_dim, 4 * flash_attn.FaVecReduce.simdgroups)), 1 }, + .threadgroup = .{ 32, flash_attn.FaVecReduce.simdgroups, 1 }, + .constants = .{ .args = .of(args) }, + }).dst; +} + +/// `q` is `[b, q, h, hd]`, `k` and `v` `[b, k, kv_h, hd]`. +fn prefill(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, positions: zml.Tensor, live: zml.Tensor, scale: f32) zml.Tensor { + const queries: i32 = @intCast(q.dim(.q)); + const tile = flash_attn.FaTiled.tileConfig(q.dim(.hd)).?; + const cfg: flash_attn.FaTiled.Cfg = .{ + .dtype = zml.kernel.metal.from(q.dtype()), + .head_size = @intCast(q.dim(.hd)), + .bq = tile[0], + .tile_kv = tile[1], + .num_threads = tile[2], + }; + return flash_attn.FaTiled.Kernel.call(.{ .q = q, .k = k, .v = v, .token_index = positions, .live = live }, .{ .out = q.shape() }, .{ + .cfg = cfg, + .grid = .{ @intCast(q.dim(.h)), @intCast(@divFloor(queries + tile[0] - 1, tile[0])), @intCast(q.dim(.b)) }, + .threadgroup = .{ @intCast(tile[2]), 1, 1 }, + .threadgroup_memory = flash_attn.FaTiled.threadgroupMemory(cfg), + .constants = .{ + .num_kv_heads = .of(@as(i32, @intCast(k.dim(.h)))), + .scale = .of(scale), + .q_len = .of(queries), + .k_len = .of(@as(i32, @intCast(k.dim(.k)))), + }, + }).out; +} + +const VecLaunch = struct { + heads: i32, + splits: i32, + splits_per_threadgroup: i32, +}; + +/// At most 256 threads, which every Apple GPU pipeline accepts. Splits past a +/// row's length exit early, so `splits` follows the core and row counts. +fn vecLaunch(rows: i32, group: i32, span: i32, cores: i32) VecLaunch { + var heads: i32 = @min(group, 4); + while (@rem(group, heads) != 0) heads -= 1; + const want: i32 = if (rows <= 2 * cores) 128 else 64; + const chunks = @divFloor(span + 31, 32); + var splits: i32 = 1; + while (splits * 2 <= @min(want, chunks)) splits *= 2; + return .{ .heads = heads, .splits = splits, .splits_per_threadgroup = @min(splits, @divFloor(8, heads)) }; +} + +fn coreCount() i32 { + const platform = zml.Compiler.current().platform; + const attribute = platform.devices[0].pjrt_desc.attribute(platform.pjrt_api, "core_count") orelse + std.debug.panic("Metal device has no core_count attribute", .{}); + return @intCast(attribute.int64); +} + +test vecLaunch { + try std.testing.expectEqual(VecLaunch{ .heads = 4, .splits = 128, .splits_per_threadgroup = 2 }, vecLaunch(16, 4, 128000, 40)); + try std.testing.expectEqual(VecLaunch{ .heads = 1, .splits = 1, .splits_per_threadgroup = 1 }, vecLaunch(8, 1, 20, 40)); +} + +test { + std.testing.refAllDecls(flash_attn); } diff --git a/zml/kernels/flash_attention/nki/nki.zig b/zml/kernels/flash_attention/nki/nki.zig index 282da132..f98a545d 100644 --- a/zml/kernels/flash_attention/nki/nki.zig +++ b/zml/kernels/flash_attention/nki/nki.zig @@ -43,71 +43,17 @@ pub const Parameters = struct { /// Attention for Neuron. Both prefill and decode lower to NKI custom calls. pub fn attention(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor, parameters: Parameters) zml.Tensor { - // TODO: https://github.com/zml/zml/issues/914 - const tp = zml.Compiler.current().resolveMesh(.{.model}); - const q_sharded = q.withPartitioning(tp, .{ .h = .model }); - const k_sharded = k.withPartitioning(tp, .{ .h = .model }); - const v_sharded = v.withPartitioning(tp, .{ .h = .model }); - - if (q.dim(.q) == 1) { - const token_index_2d = token_index.broad(zml.Shape.init(.{ .row = 1, .col = 1 }, token_index.dtype())) - .withPartitioning(tp, .{ .row = .replicated, .col = .replicated }); - - return zml.ops.manualComputation( - (struct { - q: zml.Tensor, - k: zml.Tensor, - v: zml.Tensor, - token_index: zml.Tensor, - parameters: Parameters, - - fn body(self: @This(), output: zml.Shape) zml.Tensor { - const kernel = self.parameters.decode_kernel; - return zml.ops.neuronNki( - .{ self.q, self.k, self.v, self.token_index }, - .{output}, - .{ - .name = kernel.name, - .entrypoint = kernel.entrypoint, - .source_path = kernel.source_path, - .compiler_target = self.parameters.compiler_target, - }, - )[0]; - } - }).body, - .{ - .q = q_sharded, - .k = k_sharded, - .v = v_sharded, - .token_index = token_index_2d, - .parameters = parameters, - }, - q_sharded.shape(), - ).convert(q.dtype()); - } - - return zml.ops.manualComputation( - (struct { - q: zml.Tensor, - k: zml.Tensor, - v: zml.Tensor, - parameters: Parameters, - - fn body(self: @This(), output: zml.Shape) zml.Tensor { - const kernel = self.parameters.prefill_kernel; - return zml.ops.neuronNki( - .{ self.q, self.k, self.v }, - .{output}, - .{ - .name = kernel.name, - .entrypoint = kernel.entrypoint, - .source_path = kernel.source_path, - .compiler_target = self.parameters.compiler_target, - }, - )[0]; - } - }).body, - .{ .q = q_sharded, .k = k_sharded, .v = v_sharded, .parameters = parameters }, - q_sharded.shape(), - ).convert(q.dtype()); + const decode = q.dim(.q) == 1; + const kernel = if (decode) parameters.decode_kernel else parameters.prefill_kernel; + const options: zml.ops.NeuronNkiOps = .{ + .name = kernel.name, + .entrypoint = kernel.entrypoint, + .source_path = kernel.source_path, + .compiler_target = parameters.compiler_target, + }; + const output = if (decode) + zml.ops.neuronNki(.{ q, k, v, token_index.broad(zml.Shape.init(.{ .row = 1, .col = 1 }, token_index.dtype())) }, .{q.shape()}, options)[0] + else + zml.ops.neuronNki(.{ q, k, v }, .{q.shape()}, options)[0]; + return output.convert(q.dtype()); } diff --git a/zml/kernels/flash_attention/tests.zig b/zml/kernels/flash_attention/tests.zig new file mode 100644 index 00000000..c3c3dca2 --- /dev/null +++ b/zml/kernels/flash_attention/tests.zig @@ -0,0 +1,346 @@ +const std = @import("std"); + +const zml = @import("../../zml.zig"); +const flash_attention = @import("flash_attention.zig"); + +test "attention: q=1,qh=64,kh=8" { + try testAttention( + .init(.{ .q = 1, .h = 64, .hd = 64 }, .bf16), + .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{63}, + ); + + try testAttention( + .init(.{ .q = 1, .h = 64, .hd = 64 }, .bf16), + .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{36}, + ); +} + +test "attention: b=4,q=1,qh=64,kh=8" { + // Full attention + try testAttention( + .init(.{ .b = 4, .q = 1, .h = 64, .hd = 64 }, .bf16), + .init(.{ .b = 4, .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{ 63, 63, 63, 63 }, + ); + + // Partial attention + try testAttention( + .init(.{ .b = 4, .q = 1, .h = 64, .hd = 64 }, .bf16), + .init(.{ .b = 4, .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{ 61, 57, 23, 63 }, + ); +} + +test "attention: q=1,qh=8,kh=8" { + try testAttention( + .init(.{ .q = 1, .h = 8, .hd = 64 }, .bf16), + .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{62}, + ); + try testAttention( + .init(.{ .q = 1, .h = 8, .hd = 64 }, .bf16), + .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{63}, + ); +} + +test "attention: q=1,k=4096,qh=32,kh=8,hd=128" { + try testAttention( + .init(.{ .q = 1, .h = 32, .hd = 128 }, .bf16), + .init(.{ .k = 4096, .h = 8, .hd = 128 }, .bf16), + &.{3000}, + ); +} + +test "attention: q=1,k=1000,qh=8,kh=2,hd=32" { + try testAttention( + .init(.{ .q = 1, .h = 8, .hd = 32 }, .bf16), + .init(.{ .k = 1000, .h = 2, .hd = 32 }, .bf16), + &.{900}, + ); +} + +test "attention: q=8,qh=64,kh=8" { + try testAttention( + .init(.{ .q = 8, .h = 64, .hd = 64 }, .bf16), + .init(.{ .k = 64, .h = 8, .hd = 64 }, .bf16), + &.{56}, + ); +} + +test "attention: prefill, q=k=300,qh=8,kh=4" { + // Neither length is a multiple of the query (128) or key (64) tiles. + try testAttention( + .init(.{ .q = 300, .h = 8, .hd = 128 }, .bf16), + .init(.{ .k = 300, .h = 4, .hd = 128 }, .bf16), + &.{0}, + ); +} + +test "attention: prefill, q=200,k=1000,qh=4,kh=4,hd=64" { + // Queries start at 800: every query tile sees full key tiles before the diagonal. + try testAttention( + .init(.{ .q = 200, .h = 4, .hd = 64 }, .bf16), + .init(.{ .k = 1000, .h = 4, .hd = 64 }, .bf16), + &.{800}, + ); +} + +test "attention: prefill, q=k=200,qh=4,kh=2,hd=32" { + // Rows of 32 features only hold 4 16-byte vectors: the shared memory swizzle must stay inside them. + try testAttention( + .init(.{ .q = 200, .h = 4, .hd = 32 }, .bf16), + .init(.{ .k = 200, .h = 2, .hd = 32 }, .bf16), + &.{0}, + ); +} + +test "attention: prefill, q=k=130,qh=4,kh=2,hd=256" { + try testAttention( + .init(.{ .q = 130, .h = 4, .hd = 256 }, .bf16), + .init(.{ .k = 130, .h = 2, .hd = 256 }, .bf16), + &.{0}, + ); +} + +test "attention: prefill, b=2,q=150,k=300,qh=4,kh=2" { + try testAttention( + .init(.{ .b = 2, .q = 150, .h = 4, .hd = 128 }, .bf16), + .init(.{ .b = 2, .k = 300, .h = 2, .hd = 128 }, .bf16), + &.{ 150, 100 }, + ); +} + +test "attention: non-causal prefill, q=k=300,qh=8,kh=4" { + try testAttentionOpts( + .init(.{ .q = 300, .h = 8, .hd = 128 }, .bf16), + .init(.{ .k = 300, .h = 4, .hd = 128 }, .bf16), + &.{0}, + .{ .causal = false }, + ); +} + +test "attention: non-causal prefill, b=2,q=200,k=1000,qh=4,kh=4,hd=64" { + // Queries are a suffix of the keys, as for the image tokens of a joint text/image attention. + try testAttentionOpts( + .init(.{ .b = 2, .q = 200, .h = 4, .hd = 64 }, .bf16), + .init(.{ .b = 2, .k = 1000, .h = 4, .hd = 64 }, .bf16), + &.{ 0, 0 }, + .{ .causal = false }, + ); +} + +test "attention: non-causal prefill, more queries than keys, q=300,k=100,qh=4,kh=2" { + try testAttentionOpts( + .init(.{ .q = 300, .h = 4, .hd = 128 }, .bf16), + .init(.{ .k = 100, .h = 2, .hd = 128 }, .bf16), + &.{0}, + .{ .causal = false }, + ); +} + +test "attention: non-causal decode ignores token_index, q=1,k=1000,qh=64,kh=8" { + try testAttentionOpts( + .init(.{ .q = 1, .h = 64, .hd = 64 }, .bf16), + .init(.{ .k = 1000, .h = 8, .hd = 64 }, .bf16), + &.{5}, + .{ .causal = false }, + ); +} + +pub fn testAttention(q_shape: zml.Shape, k_shape: zml.Shape, token_index_h: []const u32) !void { + return testAttentionOpts(q_shape, k_shape, token_index_h, .{}); +} + +fn testAttentionOpts(q_shape: zml.Shape, k_shape: zml.Shape, token_index_h: []const u32, opts: zml.attention.Options) !void { + return testAttentionBackends(q_shape, k_shape, token_index_h, opts, std.enums.values(flash_attention.Backend)); +} + +fn testAttentionBackends(q_shape: zml.Shape, k_shape: zml.Shape, token_index_h: []const u32, opts: zml.attention.Options, backends: []const flash_attention.Backend) !void { + const io = std.testing.io; + const allocator = std.testing.allocator; + + const platform = zml.testing.env(); + + var arena_state: std.heap.ArenaAllocator = .init(allocator); + defer arena_state.deinit(); + const arena = arena_state.allocator(); + + const token_index_shape: zml.Shape = if (token_index_h.len > 1) b: { + std.debug.assert(token_index_h.len == q_shape.dim(.b)); + break :b .init(.{ .b = q_shape.dim(.b) }, .u32); + } else .init(.{}, .u32); + + const max_k_idx: i64 = k_shape.dim(.k) - 1; + const max_q_idx: i64 = q_shape.dim(.q) - 1; + if (opts.causal) { + // Check for out of bound reads, token_index is ignored by non-causal attention. + for (token_index_h) |index| { + std.debug.assert(index + max_q_idx <= max_k_idx); + } + } + + const tensors: struct { q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, token_index: zml.Tensor } = .{ + .q = .fromShape(q_shape), + .k = .fromShape(k_shape), + .v = .fromShape(k_shape), + .token_index = .fromShape(token_index_shape), + }; + + var q = try attentionTestInput(q_shape, 17); + defer q.deinit(); + var k = try attentionTestInput(k_shape, 29); + defer k.deinit(); + var v = try attentionTestInput(k_shape, 43); + defer v.deinit(); + var token_index = try zml.Buffer.fromBytes(io, platform, token_index_shape, @ptrCast(token_index_h)); + defer token_index.deinit(); + + const meshes = platform.meshes.values(); + const mesh = zml.testing.tpMesh(); + const partitioning: zml.attention.Partitioning = if (@rem(k_shape.dim(.h), mesh.numPartitionsForLogicalAxis(.model)) == 0) .head(.model) else .{ + .in_specs = .{ .q = .replicated, .k = .replicated, .v = .replicated }, + .out_specs = .replicated, + }; + const AttentionCall = struct { + fn forward(q_: zml.Tensor, k_: zml.Tensor, v_: zml.Tensor, position: zml.Tensor, metadata: zml.attention.Metadata, parameters: zml.attention.Parameters, opts_: zml.attention.Options, mesh_: *const zml.Mesh, partitioning_: zml.attention.Partitioning) zml.Tensor { + return zml.attention.attentionOpts(q_, k_, v_, position, metadata, parameters, opts_, mesh_, partitioning_); + } + }; + const reference_metadata: zml.attention.Metadata = .{ .stablehlo = {} }; + const reference_parameters: zml.attention.Parameters = .{ .stablehlo = {} }; + const reference_exe = try platform.compileFn(allocator, io, AttentionCall.forward, .{ tensors.q, tensors.k, tensors.v, tensors.token_index, reference_metadata, reference_parameters, opts, mesh, partitioning }, .{ + .program_name = "attention_stablehlo", + .meshes = meshes, + }); + defer reference_exe.deinit(); + + var reference_d = try reference_exe.eval(allocator, io, .{ q, k, v, token_index, reference_metadata }); + defer reference_d.deinit(); + try reference_d.await(io); + try std.testing.expectEqualDeep(q_shape.withPartitioning(mesh, partitioning.out_specs)._sharding, reference_d.shape()._sharding); + const reference_h: zml.Slice = try reference_d.toSliceAlloc(allocator, io); + defer reference_h.free(allocator); + + for (backends) |backend| { + if (!backend.isAvailable(platform)) continue; + + const metadata: flash_attention.Metadata = .init(.fromBackend(backend)); + const parameters: flash_attention.Parameters = .init(.fromBackend(backend)); + const exe = try platform.compileFn( + allocator, + io, + AttentionCall.forward, + .{ tensors.q, tensors.k, tensors.v, tensors.token_index, .{ .flash_attention = metadata }, .{ .flash_attention = parameters }, opts, mesh, partitioning }, + .{ + .program_name = try std.fmt.allocPrint(arena, "attention_{t}", .{backend}), + .meshes = meshes, + }, + ); + defer exe.deinit(); + + var metadata_d = try metadata.initBuffer(io, platform); + defer flash_attention.Metadata.deinitBuffer(&metadata_d); + + var output_d = try exe.eval(allocator, io, .{ q, k, v, token_index, metadata_d }); + defer output_d.deinit(); + try output_d.await(io); + const output_h = try output_d.toSliceAlloc(allocator, io); + defer output_h.free(allocator); + + errdefer std.log.err( + \\ Attention test failed, {0t} output doesn't match reference. + \\ - reference: {1d} + \\ - {0t}: {2d} + , .{ backend, reference_h, output_h }); + try zml.testing.expectClose(io, reference_h, output_h, .{ + .absolute_tolerance = 5e-3, + .relative_tolerance = 1e-2, + .epsilon_relative = 1e-3, + .minimum_close_fraction = 0.99, + }); + } +} + +fn attentionTestInput(shape: zml.Shape, seed: u64) !zml.Buffer { + const host = try zml.Slice.alloc(std.testing.allocator, shape); + defer host.free(std.testing.allocator); + var prng: std.Random.DefaultPrng = .init(seed); + const random = prng.random(); + switch (shape.dtype()) { + inline .bf16, .f16, .f32 => |dtype| { + for (host.items(dtype.toZigType())) |*value| { + const x = random.float(f32) * 4 - 2; + value.* = if (dtype == .bf16) .fromF32(x) else @floatCast(x); + } + }, + else => unreachable, + } + return zml.Buffer.fromSlice(std.testing.io, zml.testing.env(), host); +} + +/// Independent f64 oracle: the SDPA reference uses fast dot precision, so its +/// f32 result alone cannot establish the accuracy of the kernel's f32 arithmetic. +fn checkAttentionF32(q: zml.Slice, k: zml.Slice, v: zml.Slice, positions: []const u32, output: zml.Slice) !void { + const Layout = struct { + b: usize, + s: usize, + h: usize, + d: usize, + + fn stride(shape: zml.Shape, tag: anytype) usize { + const axis = shape.hasTag(tag) orelse return 0; + var result: usize = 1; + for (shape.dims()[axis + 1 ..]) |dim| result *= @intCast(dim); + return result; + } + + fn init(shape: zml.Shape) @This() { + return .{ .b = stride(shape, .b), .s = if (shape.hasTag(.q) != null) stride(shape, .q) else stride(shape, .k), .h = stride(shape, .h), .d = stride(shape, .hd) }; + } + + fn offset(self: @This(), b: usize, s: usize, h: usize, d: usize) usize { + return b * self.b + s * self.s + h * self.h + d * self.d; + } + }; + const ql = Layout.init(q.shape); + const kl = Layout.init(k.shape); + const vl = Layout.init(v.shape); + const batch: usize = if (q.shape.hasTag(.b) != null) @intCast(q.shape.dim(.b)) else 1; + const queries: usize = @intCast(q.shape.dim(.q)); + const heads: usize = @intCast(q.shape.dim(.h)); + const dim: usize = @intCast(q.shape.dim(.hd)); + const group: usize = @intCast(@divExact(q.shape.dim(.h), k.shape.dim(.h))); + const scores = try std.testing.allocator.alloc(f64, @intCast(k.shape.dim(.k))); + defer std.testing.allocator.free(scores); + const scale = 1.0 / @sqrt(@as(f64, @floatFromInt(dim))); + for (0..batch) |b| { + const start: usize = positions[if (positions.len == 1) 0 else b]; + for (0..queries) |i| { + for (0..heads) |h| { + var maximum: f64 = -std.math.inf(f64); + const active = scores[0 .. start + i + 1]; + for (active, 0..) |*score, j| { + var dot: f64 = 0; + for (0..dim) |d| dot += @as(f64, q.items(f32)[ql.offset(b, i, h, d)]) * @as(f64, k.items(f32)[kl.offset(b, j, h / group, d)]); + score.* = dot * scale; + maximum = @max(maximum, score.*); + } + var sum: f64 = 0; + for (active) |*score| { + score.* = @exp(score.* - maximum); + sum += score.*; + } + for (0..dim) |d| { + var expected: f64 = 0; + for (active, 0..) |weight, j| expected += weight * @as(f64, v.items(f32)[vl.offset(b, j, h / group, d)]); + expected /= sum; + const actual = output.items(f32)[ql.offset(b, i, h, d)]; + try std.testing.expectApproxEqAbs(expected, @as(f64, actual), 2e-5 + 2e-5 * @abs(expected)); + } + } + } + } +} diff --git a/zml/kernels/moe/api.zig b/zml/kernels/moe/api.zig new file mode 100644 index 00000000..b66593c1 --- /dev/null +++ b/zml/kernels/moe/api.zig @@ -0,0 +1,164 @@ +const std = @import("std"); + +const stdx = @import("stdx"); + +const Tensor = @import("../../tensor.zig").Tensor; + +const log = std.log.scoped(.@"zml/moe"); + +pub const Options = struct { + activation: Activation, + /// Quantize activations for Triton FP8 GEMMs; false keeps BF16 activations. + quantize_input: bool, + /// Where routing weights are applied; FlashInfer, Mosaic, Metal and TCL require after_down. + routing_weight_placement: RoutingWeightPlacement, +}; + +pub const ActivationKind = enum { + /// Gelu activation function + gelu, + /// ReLU activation function + relu, + /// SiLU activation function + silu, + /// SwiGLU activation function (including clamped/clipped/scaled/biased variants) + swiglu, + /// SwiGLU activation function with step function + swiglu_step, + /// GeGlu activation function + geglu, + /// GeGlu activation function with tanh + geglu_tanh, +}; + +pub const Activation = union(ActivationKind) { + gelu: void, + relu: void, + silu: void, + swiglu: struct { + limit: ?f32 = null, + scale: ?f32 = null, + bias: ?f32 = null, + }, + swiglu_step: struct { + limit: ?f32 = null, + }, + geglu: void, + geglu_tanh: void, + + pub fn applyActivation(self: Activation, input: Tensor, layout: ProjectionLayout) Tensor { + const x = input.convert(.f32); + return switch (self) { + .gelu => x.gelu(), + .relu => x.relu(), + .silu => x.silu(), + .swiglu, .swiglu_step, .geglu, .geglu_tanh => b: { + const mid = @divFloor(x.dim(.out), 2); + var gate, var up = switch (layout) { + .concatenated => .{ x.slice(.out, .{ .end = mid }), x.slice(.out, .{ .start = mid }) }, + .row_interleaved_1 => .{ x.slice(.out, .{ .start = 0, .step = 2 }), x.slice(.out, .{ .start = 1, .step = 2 }) }, + .row_interleaved_8 => blk: { + const grouped = x.splitAxis(.out, .{ .group = @divExact(mid, 8), .pair = 2, .row = 8 }); + break :blk .{ + grouped.slice(.pair, .{ .end = 1 }).merge(.{ .out = .{ .group, .pair, .row } }), + grouped.slice(.pair, .{ .start = 1 }).merge(.{ .out = .{ .group, .pair, .row } }), + }; + }, + }; + + break :b switch (self) { + .swiglu => |parameters| { + stdx.debug.assert(parameters.scale == null, "triton and fly moe backend don't support swiglu scale", .{}); + + const limit: ?Tensor = if (parameters.limit) |limit| .scalar(limit, x.dtype()) else null; + + // Apply limit on gate and clamp up + gate = if (limit) |l| gate.minimum(l) else gate; + up = if (limit) |l| up.clamp(l.negate(), l) else up; + + // Apply bias + up = if (parameters.bias) |bias| up.addConstant(bias) else up; + + break :b gate.silu().mul(up); + }, + .swiglu_step => |parameters| { + gate = gate.silu(); + + const limit: ?Tensor = if (parameters.limit) |limit| .scalar(limit, x.dtype()) else null; + gate = if (limit) |l| gate.minimum(l) else gate; + up = if (limit) |l| up.clamp(l.negate(), l) else up; + break :b gate.mul(up); + }, + .geglu_tanh => gate.gelu().mul(up), + .geglu => { + log.warn("The geglu activation function was requested but we only support the tanh approximation", .{}); + break :b gate.gelu().mul(up); + }, + else => unreachable, // already treated by the top-level switch + }; + }, + }; + } +}; + +pub const RoutingWeightPlacement = enum { before_down, after_down }; + +pub const ProjectionLayout = enum { + concatenated, // G0 G1 … Gn U0 U1 … Un + row_interleaved_1, // G0 U0 G1 U1 … + row_interleaved_8, // G0…G7 U0…U7 G8…G15 U8…U15 … +}; + +/// How a backend expects the expert weights to be stored +/// The packer writes this layout and forwardMoe reads it +pub const ExpertsLayout = struct { + /// Gate/up ordering before applying the backend-specific packing. + gate_up: ProjectionLayout, + /// Backend specific transformation of the weights and scales. + packing: Packing, + + pub const Packing = enum { + /// Experts stacked as stored in the checkpoint. + plain, + /// Block scales swizzled to the 128x4 tensor-core layout (rows tiled by 4x32, columns by 4). + swizzled_scales, + /// Weights, block scales and global scales in the FlashInfer CUTLASS NVFP4 layout. + flashinfer_nvfp4, + /// MXFP4 gate/up rows interleaved by 8. E8M0 scales use 512-byte + /// 128x4 atoms in MN-major order `[expert, k/128, rows/128, 512]`, + /// represented as `[expert, k/128, rows]` i32 words of four scales. + mxfp4_interleaved8, + }; +}; + +/// Routing for a local token batch and a contiguous range of experts. +/// IDs retain the model's global expert numbering, while the projection weights +/// contain only the local experts. Kernels compute those experts contributions and +/// `forwardMoe` sums them along the expert axis. +pub const LocalRouting = struct { + /// Global expert IDs for local tokens, with shape {.t, .topk} + ids: Tensor, + weights: Tensor, + global_num_experts: u32, + local_num_experts: u32, + /// Scalar i32 containing the first local expert's global ID. + /// Null when all experts are local and global IDs already index the weights. + expert_start: ?Tensor = null, + + /// Maps routing IDs to local weight indices, using -1 for nonlocal experts. + /// For a shard owning experts 4–7, routing IDs [1, 6] become [-1, 2]. + pub fn localIds(self: LocalRouting) Tensor { + const start = self.expert_start orelse return self.ids; + const local = self.ids.cmp(.GE, start).logical(.AND, self.ids.cmp(.LT, start.addConstant(self.local_num_experts))); + return local.select(self.ids.sub(start), .scalar(-1, .i32)); + } + + /// Returns a lookup table from every global expert ID to its local weight + /// index or -1 if nonlocal. Returns null when every expert is local. + pub fn expertMap(self: LocalRouting) ?Tensor { + const start = self.expert_start orelse return null; + const ids = Tensor.arange(.{ .end = self.global_num_experts }, .i32).withTags(.{.expert}); + const local = ids.cmp(.GE, start).logical(.AND, ids.cmp(.LT, start.addConstant(self.local_num_experts))); + return local.select(ids.sub(start), .scalar(-1, .i32)); + } +}; diff --git a/zml/kernels/moe/cute/cute.zig b/zml/kernels/moe/cute/cute.zig index c0156720..8bc03c9c 100644 --- a/zml/kernels/moe/cute/cute.zig +++ b/zml/kernels/moe/cute/cute.zig @@ -3,6 +3,9 @@ const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); +const Options = @import("../api.zig").Options; +const RoutingWeightPlacement = @import("../api.zig").RoutingWeightPlacement; +const LocalRouting = @import("../api.zig").LocalRouting; pub const kernels = @import("sm10x/moe.zig"); // `moe.zig`'s `refAllDecls` reaches this file but not its imports, so the @@ -24,7 +27,7 @@ pub fn isNvfp4Supported(platform: *const zml.Platform) bool { return cc.major == 10 and cc.minor == 3; } -pub fn validateOptions(opts: zml.moe.Options) void { +pub fn validateOptions(opts: Options) void { stdx.debug.assert(opts.activation == .swiglu, "cute backend only accepts swiglu activation, got {}", .{opts.activation}); stdx.debug.assert(opts.activation.swiglu.bias == null, "cute backend requires swiglu bias to be null", .{}); stdx.debug.assert(opts.activation.swiglu.scale == null, "cute backend requires swiglu scale to be null", .{}); @@ -35,14 +38,14 @@ pub fn validateOptions(opts: zml.moe.Options) void { /// is the per-expert multiplier dequantizing each GEMM output (it folds in /// the inverse of the input's global scale), and their input scale is the /// multiplier applied to the GEMM's input before NVFP4 quantization. Both -/// expect block scales in the 128x4 tensor-core layout. +/// expect block scales in the 128x4 tensor-core layout. MXFP4 gate/up rows +/// interleave by 8 and its scale atoms are MN-major packed i32 words. pub fn fusedExperts( input: zml.Tensor, - ids: zml.Tensor, - weights: zml.Tensor, + routing: LocalRouting, gate_up: zml.nn.Linear, down: zml.nn.Linear, - options: zml.moe.Options, + options: Options, ) zml.Tensor { validateOptions(options); stdx.debug.assert(input.dtype() == .bf16, "cute backend only supports bf16 inputs, got {}", .{input.dtype()}); @@ -83,24 +86,43 @@ pub fn fusedExperts( else => stdx.debug.panic("cute backend only supports mxfp4 and nvfp4 quantization, got {}", .{gq.scheme}), }; - const context: Context = .{ - .input = input, - .ids = ids, - .weights = weights, - .w1 = gate_up.weight, - .s1 = gq.scales, - .w2 = down.weight, - .s2 = dq.scales, - .topk = ids.dim(.topk), - .swiglu_limit = options.activation.swiglu.limit, - .routing_weight_placement = options.routing_weight_placement, - .format = format, - .expert_parallel = gate_up.weight.shape().partition(.expert).isSharded(), - }; - return if (context.expert_parallel) - zml.ops.manualComputation(Context.body, context, input.shape()) - else - context.body(input.shape()); + const hidden = down.weight.dim(1); + const intermediate = down.weight.dim(2) * 2; + const topk = routing.ids.dim(.topk); + const experts = gate_up.weight.dim(.expert); + const ids = routing.localIds().convert(.i32).reshape(.{ .token = .auto, .topk = topk }); + const tokens = ids.dim(.token); + const routing_weights = routing.weights.convert(.f32).reshape(.{ tokens, topk }); + + return switch (format) { + .mxfp4 => kernels.forwardMxfp4(tokens, hidden, intermediate, experts, topk, options.activation.swiglu.limit, .{ + .x = input.reshape(.{ tokens, hidden }), + .routing_weights = routing_weights, + .w1 = gate_up.weight, + .s1 = gq.scales, + .w2 = down.weight, + .s2 = dq.scales, + .ids = ids, + }, null), + .nvfp4 => |f| kernels.forwardNvfp4(tokens, hidden, intermediate, experts, topk, .{ + .x = input.reshape(.{ tokens, hidden }), + .routing_weights = routing_weights, + .ids = ids, + .w1 = gate_up.weight, + .s1 = gq.scales, + .w2 = down.weight, + .s2 = dq.scales, + // `alpha` is per expert; `act` is one scale shared by every + // expert, which the kernels read per expert. + .alpha1 = f.alpha1.reshape(.{experts}), + .alpha2 = f.alpha2.reshape(.{experts}), + .act1 = f.act1.reshape(.{}).broad(.init(.{experts}, .f32)), + .act2 = f.act2.reshape(.{}).broad(.init(.{experts}, .f32)), + }, .{ + .swiglu_limit = options.activation.swiglu.limit, + .routing_weight_placement = options.routing_weight_placement, + }), + }.reshape(input.shape().dims()).withTags(input.shape()); } /// The format-specific tensors: NVFP4's per-expert global scales. They stay @@ -115,69 +137,3 @@ const Format = union(enum) { act2: zml.Tensor, }, }; - -const Context = struct { - input: zml.Tensor, - ids: zml.Tensor, - weights: zml.Tensor, - w1: zml.Tensor, - s1: zml.Tensor, - w2: zml.Tensor, - s2: zml.Tensor, - topk: i64, - swiglu_limit: ?f32, - routing_weight_placement: zml.moe.RoutingWeightPlacement, - format: Format, - expert_parallel: bool, - - fn body(self: Context, _: zml.Shape) zml.Tensor { - const experts = self.w1.dim(.expert); - const hidden = self.w2.dim(1); - const intermediate = self.w2.dim(2) * 2; - const tokens = self.input.dim(.t); - var ids = self.ids.convert(.i32); - const routing_weights = self.weights.convert(.f32); - if (self.expert_parallel) { - const partition_id = zml.ops.partitionId().convert(.i32); - const expert_start = partition_id.scale(experts); - const expert_end = expert_start.addConstant(experts); - // Routes of other ranks get expert -1 and are not computed here. - const local = ids.cmp(.GE, expert_start).logical(.AND, ids.cmp(.LT, expert_end)); - ids = local.select(ids.sub(expert_start), .scalar(-1, .i32)); - } - const result = switch (self.format) { - .mxfp4 => kernels.forwardMxfp4(tokens, hidden, intermediate, experts, self.topk, self.swiglu_limit, .{ - .x = self.input, - .routing_weights = routing_weights, - .w1 = self.w1, - .s1 = self.s1, - .w2 = self.w2, - .s2 = self.s2, - .ids = ids, - }, null), - .nvfp4 => |f| kernels.forwardNvfp4(tokens, hidden, intermediate, experts, self.topk, .{ - .x = self.input, - .routing_weights = routing_weights, - .ids = ids, - .w1 = self.w1, - .s1 = self.s1, - .w2 = self.w2, - .s2 = self.s2, - // `alpha` is per expert; `act` is one scale shared by every - // expert, which the kernels read per expert. - .alpha1 = f.alpha1.reshape(.{experts}), - .alpha2 = f.alpha2.reshape(.{experts}), - .act1 = f.act1.reshape(.{}).broad(.init(.{experts}, .f32)), - .act2 = f.act2.reshape(.{}).broad(.init(.{experts}, .f32)), - }, .{ - .swiglu_limit = self.swiglu_limit, - .routing_weight_placement = self.routing_weight_placement, - }), - }.withTags(self.input.shape()); - - return if (self.expert_parallel) - zml.ops.allReduce(result, zml.Tensor.add) - else - result; - } -}; diff --git a/zml/kernels/moe/cute/sm10x/common.zig b/zml/kernels/moe/cute/sm10x/common.zig index 47e05d37..da639c75 100644 --- a/zml/kernels/moe/cute/sm10x/common.zig +++ b/zml/kernels/moe/cute/sm10x/common.zig @@ -138,7 +138,7 @@ pub fn advancePipeline(b: *B, index: V, phase: V, comptime stages: usize) Pipeli pub fn peekBarrierIfRemaining(b: *B, base: V, count: V, limit: usize, state: PipelinePosition) V { var peek = b.openIfElse(count.lt(limit), .{.i1}); { - peek.yieldThen(.{b.mbarrierWaitParity(barrierAt(b, base, state.index), state.phase)}); + peek.yieldThen(.{b.mbarrierTryWait(barrierAt(b, base, state.index), state.phase, .{ .scope = .cta })}); } { peek.yieldElse(.{b.constant(.i1, true)}); diff --git a/zml/kernels/moe/cute/sm10x/moe.zig b/zml/kernels/moe/cute/sm10x/moe.zig index 479e6516..ff6edec2 100644 --- a/zml/kernels/moe/cute/sm10x/moe.zig +++ b/zml/kernels/moe/cute/sm10x/moe.zig @@ -4,26 +4,31 @@ //! rows in route order, and the top-k reduction. MXFP4 applies routing weights //! before the down projection, with the reference model's BF16 rounding. const std = @import("std"); + const stdx = @import("stdx"); const zml = @import("../../../../zml.zig"); +const RoutingWeightPlacement = @import("../../api.zig").RoutingWeightPlacement; +const triton_kernels = @import("../../triton/moe.zig"); pub const mxfp4 = @import("mxfp4.zig"); pub const mxfp4_persistent = @import("persistent_mxfp4.zig"); pub const nvfp4 = @import("nvfp4.zig"); pub const nvfp4_persistent = @import("persistent_nvfp4.zig"); const reduce = @import("reduce.zig"); -const triton_kernels = @import("../../triton/moe.zig"); pub const Mxfp4Inputs = struct { /// `[tokens, hidden]` BF16. x: zml.Tensor, /// `[tokens, topk]` FP32. routing_weights: zml.Tensor, + /// `[experts, 2 * intermediate, hidden / 2]` packed FP4, gate/up rows interleaved by 8. w1: zml.Tensor, - /// `w1` scales in the 128x4 tensor-core layout (rows tiled by 4x32, columns by 4). + /// `[experts, hidden / 128, 2 * intermediate]` i32 words of four E8M0 scales, + /// in MN-major 128x4 atoms, with rows interleaved like `w1`. s1: zml.Tensor, + /// `[experts, hidden, intermediate / 2]` packed FP4. w2: zml.Tensor, - /// `w2` scales in the 128x4 tensor-core layout (rows tiled by 4x32, columns by 4). + /// `[experts, intermediate / 128, hidden]` i32 words in MN-major 128x4 atoms. s2: zml.Tensor, /// `[tokens, topk]` local expert ids. ids: zml.Tensor, @@ -207,14 +212,14 @@ pub fn forwardMxfp4(tokens: i64, hidden: i64, intermediate: i64, experts: i64, t .weight = a.w1, .schedule = routing.schedule, .input_quant = q1.q, - .weight_scale = a.s1.bitCast(.u8), + .weight_scale = a.s1, .input_scale = q1.s, }, routing.route_inverse, a.routing_weights.reshape(.{routes})); const down = mxfp4_persistent.downRows(down_cfg, .{ .weight = a.w2, .schedule = routing.schedule, .input_quant = q2.q, - .weight_scale = a.s2.bitCast(.u8), + .weight_scale = a.s2, .input_scale = q2.s, }, routing.route_inverse); return reduce.topk(down, routing.route_map orelse routing.schedule, .{ @@ -252,7 +257,7 @@ pub const Nvfp4Inputs = struct { pub const Nvfp4Options = struct { swiglu_limit: ?f32 = null, - routing_weight_placement: zml.moe.RoutingWeightPlacement = .after_down, + routing_weight_placement: RoutingWeightPlacement = .after_down, /// GEMM tile N; null selects it from the batch size. tile_n: ?i64 = null, }; diff --git a/zml/kernels/moe/cute/sm10x/persistent_mxfp4.zig b/zml/kernels/moe/cute/sm10x/persistent_mxfp4.zig index e8901fb1..30062a3a 100644 --- a/zml/kernels/moe/cute/sm10x/persistent_mxfp4.zig +++ b/zml/kernels/moe/cute/sm10x/persistent_mxfp4.zig @@ -550,7 +550,7 @@ fn buildTiledProgram(b: *B, cfg: Config, comptime tile: Tile, comptime epilogue: .weight = .{ .ptr = cute.DType.i8 }, .schedule = .{ .ptr = cute.DType.i32 }, .input_quant = .{ .ptr = cute.DType.f8e4m3fn }, - .weight_scale = .{ .ptr = cute.DType.i8 }, + .weight_scale = .{ .ptr = cute.DType.i32 }, .input_scale = .{ .ptr = cute.DType.i8 }, }; const shape = .{ @@ -762,10 +762,10 @@ fn issueTmaStage( b.mbarrierArriveExpectTx(stage.load_barrier, tile.transactionBytes()); elected.close(); } - b.copy(b.setTmaBarrier(exec_a, stage.load_barrier), .{a_source}, .{stage.a_destination}, null); - b.copy(b.setTmaBarrier(exec_b, stage.load_barrier), .{b_source}, .{stage.b_destination}, null); - b.copy(b.setTmaBarrier(exec_sfa, stage.load_barrier), .{sfa_source}, .{stage.sfa_destination}, null); - b.copy(b.setTmaBarrier(exec_sfb, stage.load_barrier), .{sfb_source}, .{stage.sfb_destination}, null); + b.copy(b.setTmaBarrier(exec_a, stage.load_barrier), .{a_source}, .{stage.a_destination}, .{}); + b.copy(b.setTmaBarrier(exec_b, stage.load_barrier), .{b_source}, .{stage.b_destination}, .{}); + b.copy(b.setTmaBarrier(exec_sfa, stage.load_barrier), .{sfa_source}, .{stage.sfa_destination}, .{}); + b.copy(b.setTmaBarrier(exec_sfb, stage.load_barrier), .{sfb_source}, .{stage.sfb_destination}, .{}); } fn accumulatorTmemAt(b: *B, comptime tile: Tile, tmem: V, stage: V) V { @@ -826,10 +826,7 @@ fn runTmaProducer( .index = tiles.after_carried[5], .phase = tiles.after_carried[6], }; - const first_ready = b.mbarrierWaitParity( - barrierAt(b, storage.mma_barriers, first_state.index), - first_state.phase, - ); + const first_ready = b.mbarrierTryWait(barrierAt(b, storage.mma_barriers, first_state.index), first_state.phase, .{ .scope = .cta }); var stages = b.openFor( 0, k_tiles, @@ -845,7 +842,7 @@ fn runTmaProducer( const buffers = stageBuffersAt(b, tile, storage, state.index); var pending = b.openIf(stages.carried[0].eq(false)); { - b.mbarrierTryWaitParity(buffers.mma_barrier, state.phase, 10_000_000); + b.mbarrierWait(buffers.mma_barrier, state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); pending.close(); } issueTmaStage(b, tile, k_tiles, partitioned_a_target, partitioned_b_target, buffers, k, tile_m, group, expert, exec_a, exec_b, exec_sfa, exec_sfb); @@ -893,11 +890,7 @@ fn runTmaProducer( .phase = tiles.results[6], }; inline for (0..tile.stages) |stage| { - b.mbarrierTryWaitParity( - barrierAt(b, storage.mma_barriers, tail_state.index), - tail_state.phase, - 10_000_000, - ); + b.mbarrierWait(barrierAt(b, storage.mma_barriers, tail_state.index), tail_state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); if (stage + 1 < tile.stages) { tail_state = advancePipeline(b, tail_state.index, tail_state.phase, tile.stages); } @@ -922,8 +915,8 @@ const Mxfp4Mainloop = struct { sfb_target: V, pub fn copyScales(self: Mxfp4Mainloop, builder: *B) void { - builder.copy(self.sfa_copy, .{self.sfa_source}, .{self.sfa_target}, null); - builder.copy(self.sfb_copy, .{self.sfb_source}, .{self.sfb_target}, null); + builder.copy(self.sfa_copy, .{self.sfa_source}, .{self.sfa_target}, .{}); + builder.copy(self.sfb_copy, .{self.sfb_source}, .{self.sfb_target}, .{}); } }; @@ -1164,17 +1157,14 @@ fn runMmaConsumer( .index = tiles.after_carried[5], .phase = tiles.after_carried[6], }; - const first_ready = b.mbarrierWaitParity( - barrierAt(b, storage.load_barriers, first_state.index), - first_state.phase, - ); + const first_ready = b.mbarrierTryWait(barrierAt(b, storage.load_barriers, first_state.index), first_state.phase, .{ .scope = .cta }); const accumulator_state: PipelinePosition = .{ .index = tiles.after_carried[9], .phase = tiles.after_carried[10], }; const accumulator_stage = accumulator_state.index; const accumulator_empty = barrierAt(b, storage.accumulator_empty_barriers, accumulator_stage); - b.mbarrierTryWaitParity(accumulator_empty, accumulator_state.phase, 10_000_000); + b.mbarrierWait(accumulator_empty, accumulator_state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); const persistent_mma: cute.Atom = .{ .inner = tiles.after_carried[7].inner, .kernel = b }; const initial_mma = b.setMmaAccumulate(persistent_mma, b.constant(.i1, false)); @@ -1192,7 +1182,7 @@ fn runMmaConsumer( const state: PipelinePosition = .{ .index = stages.carried[3], .phase = stages.carried[4] }; var pending = b.openIf(stages.carried[0].eq(false)); { - b.mbarrierTryWaitParity(barrierAt(b, storage.load_barriers, state.index), state.phase, 10_000_000); + b.mbarrierWait(barrierAt(b, storage.load_barriers, state.index), state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); pending.close(); } const accumulator_tmem = accumulatorTmemAt(b, tile, tmem, accumulator_stage); @@ -1209,7 +1199,7 @@ fn runMmaConsumer( ); var commit_lane = b.openIf(b.electSync()); { - b.tcgen05Commit(barrierAt(b, storage.mma_barriers, state.index)); + b.tcgen05Commit(barrierAt(b, storage.mma_barriers, state.index), .{}); commit_lane.close(); } const next_count = stages.carried[2].add(1); @@ -1220,7 +1210,7 @@ fn runMmaConsumer( var full_lane = b.openIf(b.electSync()); { - b.tcgen05Commit(barrierAt(b, storage.accumulator_full_barriers, accumulator_stage)); + b.tcgen05Commit(barrierAt(b, storage.accumulator_full_barriers, accumulator_stage), .{}); full_lane.close(); } const next_accumulator_state = advancePipeline(b, accumulator_state.index, accumulator_state.phase, accumulator_stages); @@ -1271,11 +1261,7 @@ fn runMmaConsumer( inline for (0..accumulator_stages - 1) |_| { accumulator_tail = advancePipeline(b, accumulator_tail.index, accumulator_tail.phase, accumulator_stages); } - b.mbarrierTryWaitParity( - barrierAt(b, storage.accumulator_empty_barriers, accumulator_tail.index), - accumulator_tail.phase, - 10_000_000, - ); + b.mbarrierWait(barrierAt(b, storage.accumulator_empty_barriers, accumulator_tail.index), accumulator_tail.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); } /// Epilogue operands; which fields are set depends on the `Epilogue`. @@ -1349,7 +1335,7 @@ fn runEpilogue( .phase = tiles.after_carried[5], }; const stage = accumulator_state.index; - b.mbarrierTryWaitParity(barrierAt(b, storage.accumulator_full_barriers, stage), accumulator_state.phase, 10_000_000); + b.mbarrierWait(barrierAt(b, storage.accumulator_full_barriers, stage), accumulator_state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); const accumulator_tmem = accumulatorTmemAt(b, tile, tmem, stage); const warp_tmem = b.addOffset( @@ -1375,7 +1361,7 @@ fn runEpilogue( if (chunk + 1 == tile_chunks) { // Registers now own the accumulator values, so all 128 epilogue // threads release this TMEM stage before the stores. - b.mbarrierArrive(barrierAt(b, storage.accumulator_empty_barriers, stage), 1); + _ = b.mbarrierArrive(barrierAt(b, storage.accumulator_empty_barriers, stage), .{}); } var columns: [@intCast(@min(n, 32))]V = undefined; inline for (&columns, 0..) |*column, i| column.* = b.vectorExtract(values, i, .i32).bitCast(.f32); @@ -1405,7 +1391,7 @@ fn runEpilogue( b.relinquishTmemAllocPermit(.{}); release_warp.close(); } - b.namedBarrier(1, 128); + b.barrierCtaSyncAligned(1, 128); var dealloc_warp = b.openIf(warp.eq(0)); { b.deallocTmem(tmem, tile.tmem_columns, .{}); @@ -1453,11 +1439,10 @@ fn routedLanes(b: *B, comptime tile: Tile, operands: EpilogueOperands, group: V, return .{ .count = count, .route = route, .weight = weight }; } -/// Python `fused_epilogue`: thread `t` owns output row `128 * tile_m + t` -/// of the interleaved projection, so even threads hold a gate and their odd -/// neighbour the matching up value. Each even thread computes one SwiGLU -/// activation per routed row; 32 consecutive activations span two warps, -/// whose partial maxima meet in shared memory. +/// Thread `t` owns output row `128 * tile_m + t` of the interleave-8 +/// projection. Gate and up lanes are paired by XOR 8. Each gate thread +/// computes one SwiGLU activation per routed row; 32 consecutive activations +/// span two warps, whose partial maxima meet in shared memory. /// /// At decode sizes most rows of a group are padding, whose outputs are never /// read. `count` is uniform across the CTA, so whole 8-row chunks of padding @@ -1484,7 +1469,7 @@ fn swigluRows(b: *B, comptime tile: Tile, storage: StageStorage, operands: Epilo const tid = b.threadIdx().x; const lane = tid.rem(32); const warp = b.makeWarpUniform(b.makeWarpUniform(tid.div(32))); - const is_gate = tid.rem(2).eq(0); + const is_gate = tid.rem(16).lt(8); const first_row = group.mul(operands.group_rows); const f32_type = cute.DType.f32.toMlir(b.ctx); // The TMA C staging buffer is unused by this epilogue. @@ -1511,7 +1496,7 @@ fn swigluRows(b: *B, comptime tile: Tile, storage: StageStorage, operands: Epilo if (operands.swiglu_limit) |limit| { gate = gate.minimum(limit); } - var up_value = b.shuffleXor(value, 1).to(.bf16).to(.f32); + var up_value = b.shuffleXor(value, 8).to(.bf16).to(.f32); if (operands.swiglu_limit) |limit| { up_value = up_value.maximum(-limit).minimum(limit); } @@ -1520,7 +1505,7 @@ fn swigluRows(b: *B, comptime tile: Tile, storage: StageStorage, operands: Epilo const exp_neg = b.unaryF32("ex2.approx.ftz.f32", gate.mul(-std.math.log2e)); var result = up_value.mul(gate).mul(b.unaryF32("rcp.approx.ftz.f32", exp_neg.add(1.0))); result = result.mul(b.shuffleIdx(routed.weight.?, c)).to(.bf16).to(.f32); - // Odd threads hold no activation and must not raise the maxima. + // Up lanes hold no activation and must not raise the maxima. activation.* = b.select(is_gate, result, b.constant(.f32, 0)); const maximum = b.warpMaxNonNegative(activation.*.abs()); chunk_maximum = b.select(lane.eq(c), maximum, chunk_maximum); @@ -1545,16 +1530,17 @@ fn swigluRows(b: *B, comptime tile: Tile, storage: StageStorage, operands: Epilo scratch.set(.{ warp, lane }, lane_maximum); owner.close(); } - b.namedBarrier(1, 128); + b.barrierCtaSyncAligned(1, 128); // Lane `c` turns the two half-block maxima of column `c` into its scale. const pair = warp.bitAnd(~@as(i32, 1)); const lane_column = b.select(lane.lt(n), lane, b.constant(.i32, 0)); const scale = mxScale(b, scratch.get(.{ pair, lane_column }, .f32).maximum(scratch.get(.{ pair.add(1), lane_column }, .f32))); // The next tile reuses the scratch maxima. - b.namedBarrier(1, 128); + b.barrierCtaSyncAligned(1, 128); - const activation_index = tile_m.mul(64).add(tid.div(2)); + const activation_row = tid.div(16).mul(8).add(tid.rem(8)); + const activation_index = tile_m.mul(64).add(activation_row); for (0..row_chunks) |row_chunk| { const first = row_chunk * swiglu_chunk; const last = @min(first + swiglu_chunk, live); @@ -1767,7 +1753,7 @@ fn buildUmmaDevice(b: *B, cfg: Config, comptime tile: Tile, comptime epilogue: E var mma_warp = b.openIf(warp.eq(4)); { // Python's MMA role reaches its own TmemAllocator.wait_for_alloc site. - b.namedBarrier(2, 160); + b.barrierCtaSyncAligned(2, 160); switch (k_tiles) { inline 16, 18, 32, 40 => |kt| runMmaConsumer(b, tile, kt, storage, tiled_mma, tmem_holding, active, d.tiles_m), else => unreachable, @@ -1784,7 +1770,7 @@ fn buildUmmaDevice(b: *B, cfg: Config, comptime tile: Tile, comptime epilogue: E b.allocTmem(tile.tmem_columns, tmem_holding, .{}); alloc_warp.close(); } - b.namedBarrier(2, 160); + b.barrierCtaSyncAligned(2, 160); switch (epilogue) { .swiglu_mxfp8 => runEpilogue(b, tile, epilogue, storage, tmem_holding, active, d.tiles_m, direct, .{ .group_sizes = view(b, d.group_sizes), @@ -1844,7 +1830,13 @@ fn buildHostLaunch(b: *B, cfg: Config, comptime tile: Tile, comptime epilogue: E const sf_order = b.makeIntTuple(.{ 1, 0, 2 }); const sfa_shape = b.makeShapeValues(.{ b.arg(host_arg.problem_m), b.arg(host_arg.problem_k), b.arg(host_arg.expert_count) }, b.shapeType(.{ cute.AlgebraToken.dynamic, cute.AlgebraToken.dynamic, cute.AlgebraToken.dynamic })); const sfb_shape = b.makeShapeValues(.{ b.arg(host_arg.problem_n), b.arg(host_arg.problem_k), b.arg(host_arg.problem_l) }, b.shapeType(.{ cute.AlgebraToken.dynamic, cute.AlgebraToken.dynamic, cute.AlgebraToken.dynamic })); - const dynamic_sfa_layout = b.tileToShape(sf_atom_layout, sfa_shape, sf_order, b.layoutType(dynamic_sf_layout)); + // Weight scales step M atoms first; activation scales retain K-major order. + const sfa_order = b.makeIntTuple(.{ 0, 1, 2 }); + const sfa_layout = b.layoutSpec( + .{ .{ .{ 32, 4 }, cute.AlgebraToken.dynamic }, .{ .{ 32, 4 }, cute.AlgebraToken.dynamic }, .{ 1, cute.AlgebraToken.dynamic } }, + .{ .{ .{ 16, 4 }, 512 }, .{ .{ 0, 1 }, sf_stride }, .{ 0, sf_stride } }, + ); + const dynamic_sfa_layout = b.tileToShape(sf_atom_layout, sfa_shape, sfa_order, b.layoutType(sfa_layout)); const dynamic_sfb_layout = b.tileToShape(sf_atom_layout, sfb_shape, sf_order, b.layoutType(dynamic_sf_layout)); const gSFA = b.makeTensorView(b.recastPointer(b.arg(host_arg.weight_scale), .f8e8m0fnu, .gmem, 16), dynamic_sfa_layout); const gSFB = b.makeTensorView(b.recastPointer(b.arg(host_arg.input_scale), .f8e8m0fnu, .gmem, 16), dynamic_sfb_layout); diff --git a/zml/kernels/moe/cute/sm10x/persistent_nvfp4.zig b/zml/kernels/moe/cute/sm10x/persistent_nvfp4.zig index ac59735c..0de45007 100644 --- a/zml/kernels/moe/cute/sm10x/persistent_nvfp4.zig +++ b/zml/kernels/moe/cute/sm10x/persistent_nvfp4.zig @@ -699,10 +699,10 @@ fn issueTmaStage( var elected = b.openIf(b.electSync()); b.mbarrierArriveExpectTx(stage.load_barrier, tile.transactionBytes()); elected.close(); - b.copy(b.setTmaBarrier(exec_a, stage.load_barrier), .{a_source}, .{stage.a_destination}, null); - b.copy(b.setTmaBarrier(exec_b, stage.load_barrier), .{b_source}, .{stage.b_destination}, null); - b.copy(b.setTmaBarrier(exec_sfa, stage.load_barrier), .{sfa_source}, .{stage.sfa_destination}, null); - b.copy(b.setTmaBarrier(exec_sfb, stage.load_barrier), .{sfb_source}, .{stage.sfb_destination}, null); + b.copy(b.setTmaBarrier(exec_a, stage.load_barrier), .{a_source}, .{stage.a_destination}, .{}); + b.copy(b.setTmaBarrier(exec_b, stage.load_barrier), .{b_source}, .{stage.b_destination}, .{}); + b.copy(b.setTmaBarrier(exec_sfa, stage.load_barrier), .{sfa_source}, .{stage.sfa_destination}, .{}); + b.copy(b.setTmaBarrier(exec_sfb, stage.load_barrier), .{sfb_source}, .{stage.sfb_destination}, .{}); } fn accumulatorTmemAt(b: *B, comptime tile: Tile, tmem: V, stage: V) V { @@ -756,10 +756,7 @@ fn runTmaProducer( .index = tiles.after_carried[5], .phase = tiles.after_carried[6], }; - const first_ready = b.mbarrierWaitParity( - barrierAt(b, storage.mma_barriers, first_state.index), - first_state.phase, - ); + const first_ready = b.mbarrierTryWait(barrierAt(b, storage.mma_barriers, first_state.index), first_state.phase, .{ .scope = .cta }); var stages = b.openFor( 0, @as(i32, @intCast(k_tiles)), @@ -773,7 +770,7 @@ fn runTmaProducer( }; const buffers = stageBuffersAt(b, tile, storage, state.index); var pending = b.openIf(stages.carried[0].eq(false)); - b.mbarrierTryWaitParity(buffers.mma_barrier, state.phase, 10_000_000); + b.mbarrierWait(buffers.mma_barrier, state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); pending.close(); issueTmaStage(b, tile, k_tiles, partitioned_a_target, partitioned_b_target, buffers, k, tile_m, group, expert, exec_a, exec_b, exec_sfa, exec_sfb); const next_count = stages.carried[1].add(1); @@ -817,11 +814,7 @@ fn runTmaProducer( .phase = tiles.results[6], }; inline for (0..tile.stages) |stage| { - b.mbarrierTryWaitParity( - barrierAt(b, storage.mma_barriers, tail_state.index), - tail_state.phase, - 10_000_000, - ); + b.mbarrierWait(barrierAt(b, storage.mma_barriers, tail_state.index), tail_state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); if (stage + 1 < tile.stages) { tail_state = advancePipeline(b, tail_state.index, tail_state.phase, tile.stages); } @@ -843,8 +836,8 @@ const Mainloop = struct { sfb_target: V, pub fn copyScales(self: Mainloop, builder: *B) void { - builder.copy(self.sfa_copy, .{self.sfa_source}, .{self.sfa_target}, null); - builder.copy(self.sfb_copy, .{self.sfb_source}, .{self.sfb_target}, null); + builder.copy(self.sfa_copy, .{self.sfa_source}, .{self.sfa_target}, .{}); + builder.copy(self.sfb_copy, .{self.sfb_source}, .{self.sfb_target}, .{}); } }; @@ -1020,17 +1013,14 @@ fn runMmaConsumer( .index = tiles.after_carried[5], .phase = tiles.after_carried[6], }; - const first_ready = b.mbarrierWaitParity( - barrierAt(b, storage.load_barriers, first_state.index), - first_state.phase, - ); + const first_ready = b.mbarrierTryWait(barrierAt(b, storage.load_barriers, first_state.index), first_state.phase, .{ .scope = .cta }); const accumulator_state: PipelinePosition = .{ .index = tiles.after_carried[9], .phase = tiles.after_carried[10], }; const accumulator_stage = accumulator_state.index; const accumulator_empty = barrierAt(b, storage.accumulator_empty_barriers, accumulator_stage); - b.mbarrierTryWaitParity(accumulator_empty, accumulator_state.phase, 10_000_000); + b.mbarrierWait(accumulator_empty, accumulator_state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); const persistent_mma: cute.Atom = .{ .inner = tiles.after_carried[7].inner, .kernel = b }; const initial_mma = b.setMmaAccumulate(persistent_mma, b.constant(.i1, false)); @@ -1043,7 +1033,7 @@ fn runMmaConsumer( const k = stages.carried[2]; const state: PipelinePosition = .{ .index = stages.carried[3], .phase = stages.carried[4] }; var pending = b.openIf(stages.carried[0].eq(false)); - b.mbarrierTryWaitParity(barrierAt(b, storage.load_barriers, state.index), state.phase, 10_000_000); + b.mbarrierWait(barrierAt(b, storage.load_barriers, state.index), state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); pending.close(); const accumulator_tmem = accumulatorTmemAt(b, tile, tmem, accumulator_stage); const mainloop = staged_mainloop.atStage(b, state.index, accumulator_tmem); @@ -1058,7 +1048,7 @@ fn runMmaConsumer( mainloop.accumulator, ); var commit_lane = b.openIf(b.electSync()); - b.tcgen05Commit(barrierAt(b, storage.mma_barriers, state.index)); + b.tcgen05Commit(barrierAt(b, storage.mma_barriers, state.index), .{}); commit_lane.close(); const next_count = stages.carried[2].add(1); const next_state = advancePipeline(b, state.index, state.phase, tile.stages); @@ -1066,7 +1056,7 @@ fn runMmaConsumer( stages.yield(.{ next_ready, mma.value(), next_count, next_state.index, next_state.phase }); var full_lane = b.openIf(b.electSync()); - b.tcgen05Commit(barrierAt(b, storage.accumulator_full_barriers, accumulator_stage)); + b.tcgen05Commit(barrierAt(b, storage.accumulator_full_barriers, accumulator_stage), .{}); full_lane.close(); const next_accumulator_state = advancePipeline(b, accumulator_state.index, accumulator_state.phase, accumulator_stages); const next_acc_count = tiles.after_carried[8].add(1); @@ -1113,11 +1103,7 @@ fn runMmaConsumer( inline for (0..accumulator_stages - 1) |_| { accumulator_tail = advancePipeline(b, accumulator_tail.index, accumulator_tail.phase, accumulator_stages); } - b.mbarrierTryWaitParity( - barrierAt(b, storage.accumulator_empty_barriers, accumulator_tail.index), - accumulator_tail.phase, - 10_000_000, - ); + b.mbarrierWait(barrierAt(b, storage.accumulator_empty_barriers, accumulator_tail.index), accumulator_tail.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); } /// Epilogue operands; which fields are set depends on the `Epilogue`. @@ -1193,7 +1179,7 @@ fn runEpilogue( const expert = operands.expert_ids.get(.{group}, .i32); const alpha = operands.alpha.get(.{expert}, .f32); const act_scale = if (operands.act_scale) |scales| scales.get(.{expert}, .f32) else null; - b.mbarrierTryWaitParity(barrierAt(b, storage.accumulator_full_barriers, stage), accumulator_state.phase, 10_000_000); + b.mbarrierWait(barrierAt(b, storage.accumulator_full_barriers, stage), accumulator_state.phase, .{ .scope = .cta, .time_limit = 10_000_000, .test_first = false }); const accumulator_tmem = accumulatorTmemAt(b, tile, tmem, stage); const warp_tmem = b.addOffset( @@ -1213,7 +1199,7 @@ fn runEpilogue( const values = b.tmemLoadVector(chunk_tmem, @intCast(chunk_rows), .{ .num_dp = 32, .num_b = 32, .num_rep = @intCast(chunk_rows) }); b.fenceTmemLoad(); if (chunk + 1 == tile_chunks) { - b.mbarrierArrive(barrierAt(b, storage.accumulator_empty_barriers, stage), 1); + _ = b.mbarrierArrive(barrierAt(b, storage.accumulator_empty_barriers, stage), .{}); } var columns: [@intCast(@min(n, 32))]V = undefined; inline for (&columns, 0..) |*column, i| column.* = b.vectorExtract(values, i, .i32).bitCast(.f32).mul(alpha); @@ -1239,7 +1225,7 @@ fn runEpilogue( var release_warp = b.openIf(warp.eq(0)); b.relinquishTmemAllocPermit(.{}); release_warp.close(); - b.namedBarrier(1, 128); + b.barrierCtaSyncAligned(1, 128); var dealloc_warp = b.openIf(warp.eq(0)); b.deallocTmem(tmem, tile.tmem_columns, .{}); dealloc_warp.close(); @@ -1543,7 +1529,7 @@ fn buildUmmaDevice(b: *B, cfg: Config, comptime tile: Tile, comptime epilogue: E tma_warp.close(); var mma_warp = b.openIf(warp.eq(4)); - b.namedBarrier(2, 160); + b.barrierCtaSyncAligned(2, 160); runMmaConsumer(b, tile, reduction_tiles, storage, tiled_mma, tmem_holding, active, d.tiles_m); mma_warp.close(); @@ -1551,7 +1537,7 @@ fn buildUmmaDevice(b: *B, cfg: Config, comptime tile: Tile, comptime epilogue: E var alloc_warp = b.openIf(warp.eq(0)); b.allocTmem(tile.tmem_columns, tmem_holding, .{}); alloc_warp.close(); - b.namedBarrier(2, 160); + b.barrierCtaSyncAligned(2, 160); const shared_operands: EpilogueOperands = .{ .expert_ids = expert_ids, .alpha = view(b, d.alpha), diff --git a/zml/kernels/moe/cutlass_flashinfer/BUILD.bazel b/zml/kernels/moe/cutlass_flashinfer/BUILD.bazel deleted file mode 100644 index 54b8604c..00000000 --- a/zml/kernels/moe/cutlass_flashinfer/BUILD.bazel +++ /dev/null @@ -1,8 +0,0 @@ -package(default_visibility = ["//zml:__pkg__"]) - -filegroup( - name = "cutlass_flashinfer", - srcs = [ - "cutlass_flashinfer.zig", - ], -) diff --git a/zml/kernels/moe/cutlass_flashinfer/cutlass_flashinfer.zig b/zml/kernels/moe/cutlass_flashinfer/cutlass_flashinfer.zig deleted file mode 100644 index 2c7ad342..00000000 --- a/zml/kernels/moe/cutlass_flashinfer/cutlass_flashinfer.zig +++ /dev/null @@ -1,633 +0,0 @@ -const std = @import("std"); -const stdx = @import("stdx"); - -const fi_cutlass_moe = @import("platforms/cuda/flashinfer_cutlass_moe"); -const platforms = @import("platforms"); -const zml = @import("../../../zml.zig"); - -const log = std.log.scoped(.@"zml/moe/cutlass_flashinfer"); - -pub const auto_tactic: i32 = -1; - -pub const Options = struct { - /// Device used at graph-construction time to select the architecture - /// library and size the XLA-owned scratch buffer. - workspace_query_device: i32, - activation: zml.moe.Activation, - enable_pdl: bool, - gemm1_tactic: i32, - /// GEMM2 uses FlashInfer's absolute tactic index. Query tacticCounts() to - /// obtain the first valid GEMM2 index. - gemm2_tactic: i32, - - pub fn fromGlobalOptions(opts: zml.moe.Options) Options { - return .{ - .workspace_query_device = 0, - .activation = opts.activation, - .enable_pdl = false, - .gemm1_tactic = auto_tactic, - .gemm2_tactic = auto_tactic, - }; - } -}; - -pub fn validateOptions(opts: zml.moe.Options) void { - stdx.debug.assert(!opts.quantize_input, "Optional FP8 input quantization requires the Triton MoE backend", .{}); - stdx.debug.assert(opts.routing_weight_placement == .after_down, "Non-Triton MoE backends require routing weights after the down projection", .{}); - stdx.debug.assert(opts.activation == .swiglu, "flashinfer_cutlass backend only accepts swiglu activation, got {}", .{opts.activation}); - stdx.debug.assert(opts.activation.swiglu.limit == null, "Activation thresholds require the Triton MoE backend", .{}); - stdx.debug.assert(opts.activation.swiglu.bias == null, "flashinfer_cutlass backend requires swiglu bias to be null", .{}); - stdx.debug.assert(opts.activation.swiglu.scale == null, "flashinfer_cutlass backend requires swiglu scale to be null", .{}); -} - -const Input = struct { - hidden_states: zml.Tensor, - fc1_weights: zml.Tensor, - fc2_weights: zml.Tensor, - topk_weights: zml.Tensor, - topk_ids: zml.Tensor, - fc1_act_global: zml.Tensor, - fc1_weight_block: zml.Tensor, - fc1_global: zml.Tensor, - fc2_act_global: zml.Tensor, - fc2_weight_block: zml.Tensor, - fc2_global: zml.Tensor, -}; - -const Output = struct { - output: zml.Shape, - workspace: zml.Shape, -}; - -const Attributes = struct { - runners: u64 = 0, - num_tokens: i64, - hidden_size: i64, - intermediate_size: i64, - num_experts: i32, - top_k: i32, - activation: i32, - enable_pdl: bool, - fc1_act_per_expert: bool, - fc2_act_per_expert: bool, - gemm1_tactic: i32, - gemm2_tactic: i32, -}; - -const DeviceRunner = struct { - api: *fi_cutlass_moe.Api, - runner: *fi_cutlass_moe.Runner, -}; - -const max_num_devices = zml.platform.Platform.MAX_NUM_DEVICES; - -pub const Runners = struct { - runners: [max_num_devices]?DeviceRunner = @splat(null), - - pub fn deinit(self: *Runners) void { - for (self.runners) |device_runner| { - if (device_runner) |runner| { - _ = runner.api.runnerDestroy(runner.runner); - } - } - } - - fn ensureRunner(self: *Runners, device: i32) !DeviceRunner { - if (device < 0 or device >= max_num_devices) return error.UnsupportedDevice; - - const index: usize = @intCast(device); - if (self.runners[index]) |runner| return runner; - - const api = try fi_cutlass_moe.apiForDevice(device); - const options = runnerOptions(device); - var runner: ?*fi_cutlass_moe.Runner = null; - try checkStatus(api, api.runnerCreate(&options, &runner)); - const result: DeviceRunner = .{ - .api = api, - .runner = runner orelse return error.RunnerInitializationFailed, - }; - self.runners[index] = result; - return result; - } -}; - -fn checkStatus(api: *const fi_cutlass_moe.Api, status: fi_cutlass_moe.Status) !void { - if (status == fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_SUCCESS) return; - if (api.lastError()) |message| { - log.err("FlashInfer CUTLASS MoE failed (status {d}): {s}", .{ - status, - std.mem.span(message), - }); - } else { - log.err("FlashInfer CUTLASS MoE failed with status {d}", .{status}); - } - return switch (status) { - fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_INVALID_ARGUMENT => error.InvalidArgument, - fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_UNSUPPORTED => error.UnsupportedArchitecture, - fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_CUDA_ERROR => error.Cuda, - else => error.FlashinferCutlassMoe, - }; -} - -fn runnerOptions(device: i32) fi_cutlass_moe.RunnerOptions { - var options = std.mem.zeroes(fi_cutlass_moe.RunnerOptions); - options.struct_size = @sizeOf(fi_cutlass_moe.RunnerOptions); - options.activation_dtype = fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_DTYPE_BF16; - options.weight_dtype = fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_DTYPE_PACKED_FP4; - options.output_dtype = fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_DTYPE_BF16; - options.device = device; - options.use_fused_finalize = 1; - return options; -} - -fn runnersFromAttributes(attributes: Attributes) !*Runners { - if (attributes.runners == 0) return error.RunnersNotLoaded; - return @ptrFromInt(attributes.runners); -} - -fn currentRunners() !*Runners { - const platform = zml.Compiler.current().platform; - const runners = platform.state.cuda.fi_cutlass_moe_runners; - return runners orelse error.RunnersNotLoaded; -} - -fn cutlassActivation(activation: zml.moe.Activation) i32 { - return switch (activation) { - .gelu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_GELU, - .relu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_RELU, - .silu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_SILU, - .swiglu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_SWIGLU, - .swiglu_step => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_SWIGLU_STEP, - .geglu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_GEGLU, - .geglu_tanh => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_GEGLU_TANH, - }; -} - -fn makeContext(attributes: Attributes) fi_cutlass_moe.Context { - var context = std.mem.zeroes(fi_cutlass_moe.Context); - context.struct_size = @sizeOf(fi_cutlass_moe.Context); - context.num_tokens = attributes.num_tokens; - context.hidden_size = attributes.hidden_size; - context.intermediate_size = attributes.intermediate_size; - context.num_experts = attributes.num_experts; - context.num_experts_on_rank = attributes.num_experts; - context.top_k = attributes.top_k; - context.tp_size = 1; - context.ep_size = 1; - context.activation = @intCast(attributes.activation); - context.enable_pdl = @intFromBool(attributes.enable_pdl); - context.swizzled_input_sf = 1; - return context; -} - -fn ffiCallNvfp4( - call_frame: *zml.pjrt.ffi.CallFrame, - input: zml.pjrtx.TensorToCustomCallBuffer(Input), - output: zml.pjrtx.ShapeToCustomCallBuffer(Output), - attributes: Attributes, -) !?*zml.pjrt.ffi.Error { - const device = try call_frame.ctx.getDeviceOrdinal(call_frame.api); - const runners = try runnersFromAttributes(attributes); - const deviceRunner = try runners.ensureRunner(device); - const context = makeContext(attributes); - - var io = std.mem.zeroes(fi_cutlass_moe.Io); - io.struct_size = @sizeOf(fi_cutlass_moe.Io); - io.input = input.hidden_states.ptr; - io.token_selected_experts = @ptrCast(@alignCast(input.topk_ids.ptr)); - io.token_final_scales = @ptrCast(@alignCast(input.topk_weights.ptr)); - io.fc1_expert_weights = input.fc1_weights.ptr; - io.fc2_expert_weights = input.fc2_weights.ptr; - io.quant_scales[0] = input.fc1_act_global.ptr; - io.quant_scales[1] = input.fc1_weight_block.ptr; - io.quant_scales[2] = input.fc1_global.ptr; - io.quant_scales[3] = input.fc2_act_global.ptr; - io.quant_scales[4] = input.fc2_weight_block.ptr; - io.quant_scales[5] = input.fc2_global.ptr; - io.quant_scale_count = 6; - io.quant_scale_per_expert_mask = - (@as(u32, @intFromBool(attributes.fc1_act_per_expert)) << 0) | - (@as(u32, @intFromBool(attributes.fc2_act_per_expert)) << 3); - io.output = output.output.ptr; - - var workspace = std.mem.zeroes(fi_cutlass_moe.Workspace); - workspace.struct_size = @sizeOf(fi_cutlass_moe.Workspace); - workspace.data = output.workspace.ptr; - workspace.data_bytes = output.workspace.shape.byteSize(); - - try checkStatus( - deviceRunner.api, - deviceRunner.api.run( - deviceRunner.runner, - &context, - &io, - &workspace, - @ptrCast(call_frame.api.stream(call_frame.ctx)), - attributes.gemm1_tactic, - attributes.gemm2_tactic, - ), - ); - return null; -} - -const routedNvfp4Call = zml.ops.CustomCall(Input, Output, Attributes, ffiCallNvfp4, .{ - .name = "flashinfer_cutlass_nvfp4_routed_moe", - // Expert mesh is owned by forwardMoe outer manual computation. - .sharding_aware = false, - .has_side_effect = false, -}); - -fn computeCapability(platform: *const zml.Platform) !u16 { - const cc = zml.platform.cuda.computeCapability(platform) orelse return error.UnsupportedPlatform; - return switch (cc.sm()) { - 100, 103, 120 => |sm| sm, - else => error.UnsupportedArchitecture, - }; -} - -pub fn load( - allocator: std.mem.Allocator, - io: std.Io, - platform: *zml.Platform, -) !void { - if (comptime platforms.isEnabled(.cuda)) { - var cuda_state = &platform.state.cuda; - if (cuda_state.fi_cutlass_moe_runners != null) return; - try fi_cutlass_moe.load(allocator, io, try computeCapability(platform)); - const runners = try allocator.create(Runners); - runners.* = .{}; - cuda_state.fi_cutlass_moe_runners = runners; - return; - } - return error.UnsupportedPlatform; -} - -pub fn register(platform: *const zml.Platform) !void { - if (comptime platforms.isEnabled(.cuda)) { - try routedNvfp4Call.register(platform); - return; - } - return error.UnsupportedPlatform; -} - -pub fn isAvailable(platform: *const zml.Platform) bool { - if (platform.target != .cuda) return false; - if (platform.state.cuda.fi_cutlass_moe_runners == null) return false; - _ = computeCapability(platform) catch return false; - return true; -} - -pub fn isNvfp4Supported(platform: *const zml.Platform) bool { - if (platform.state.cuda.fi_cutlass_moe_runners == null) return false; - const compute_capability = computeCapability(platform) catch return false; - return switch (compute_capability) { - 100, 103, 120 => true, - else => false, - }; -} - -pub fn tacticCounts( - platform: *const zml.Platform, - device: i32, -) !struct { gemm1: i32, gemm2: i32 } { - const runners = platform.state.cuda.fi_cutlass_moe_runners orelse return error.RunnersNotLoaded; - const deviceRunner = try runners.ensureRunner(device); - var gemm1: i32 = 0; - var gemm2: i32 = 0; - try checkStatus( - deviceRunner.api, - deviceRunner.api.getTacticCounts(deviceRunner.runner, &gemm1, &gemm2), - ); - return .{ .gemm1 = gemm1, .gemm2 = gemm2 }; -} - -fn roundUp(value: i64, alignment: i64) i64 { - return @divTrunc(value + alignment - 1, alignment) * alignment; -} - -pub fn fc1BlockScaleShape( - num_experts: i64, - hidden_size: i64, - intermediate_size: i64, - activation: zml.moe.Activation, -) zml.Shape { - const rows = switch (activation) { - .silu, .gelu, .relu => intermediate_size, - .swiglu, .geglu, .geglu_tanh, .swiglu_step => 2 * intermediate_size, - }; - return .init( - .{ num_experts, roundUp(rows, 128), roundUp(@divExact(hidden_size, 16), 4) }, - .f8e4m3fn, - ); -} - -pub fn fc2BlockScaleShape( - num_experts: i64, - hidden_size: i64, - intermediate_size: i64, -) zml.Shape { - return .init( - .{ num_experts, roundUp(hidden_size, 128), roundUp(@divExact(intermediate_size, 16), 4) }, - .f8e4m3fn, - ); -} - -fn isGlobalOrPerExpertScale(tensor: zml.Tensor, num_experts: i64) bool { - return tensor.dtype() == .f32 and - (tensor.rank() == 0 or (tensor.rank() == 1 and tensor.dim(0) == num_experts)); -} - -fn validateInputs( - hidden_states: zml.Tensor, - fc1_weights: zml.Tensor, - fc2_weights: zml.Tensor, - topk_weights: zml.Tensor, - topk_ids: zml.Tensor, - fc1_act_global: zml.Tensor, - fc1_weight_block: zml.Tensor, - fc1_global: zml.Tensor, - fc2_act_global: zml.Tensor, - fc2_weight_block: zml.Tensor, - fc2_global: zml.Tensor, - options: Options, -) !Attributes { - if (hidden_states.dtype() != .bf16 or - fc1_weights.dtype() != .f4e2m1 or - fc2_weights.dtype() != .f4e2m1 or - topk_weights.dtype() != .f32 or - topk_ids.dtype() != .i32 or - fc1_weight_block.dtype() != .f8e4m3fn or - fc2_weight_block.dtype() != .f8e4m3fn or - fc1_global.dtype() != .f32 or - fc2_global.dtype() != .f32) - { - return error.UnsupportedType; - } - if (hidden_states.rank() != 2 or - fc1_weights.rank() != 3 or - fc2_weights.rank() != 3 or - topk_weights.rank() != 2 or - topk_ids.rank() != 2) - { - return error.InvalidShape; - } - - const tokens = hidden_states.dim(.t); - const hidden_size = hidden_states.dim(.d); - const num_experts = fc1_weights.dim(0); - const fc1_rows = fc1_weights.dim(1); - const intermediate_size = switch (options.activation) { - .silu, .gelu, .relu => fc1_rows, - .swiglu, .geglu, .geglu_tanh, .swiglu_step => @divExact(fc1_rows, 2), - }; - const top_k = topk_ids.dim(.topk); - - if (tokens <= 0 or hidden_size <= 0 or - num_experts <= 0 or intermediate_size <= 0 or - @mod(hidden_size, 16) != 0 or @mod(intermediate_size, 16) != 0 or - top_k <= 0 or top_k > num_experts) - { - return error.InvalidShape; - } - if (fc1_weights.dim(2) != hidden_size or - fc2_weights.dim(0) != num_experts or - fc2_weights.dim(1) != hidden_size or - fc2_weights.dim(2) != intermediate_size or - topk_weights.dim(.t) != tokens or - topk_weights.dim(.topk) != top_k or - topk_ids.dim(.t) != tokens or - !fc1_weight_block.shape().eql(fc1BlockScaleShape( - num_experts, - hidden_size, - intermediate_size, - options.activation, - )) or - !fc2_weight_block.shape().eql(fc2BlockScaleShape( - num_experts, - hidden_size, - intermediate_size, - )) or - !fc1_global.shape().eql(.init(.{num_experts}, .f32)) or - !fc2_global.shape().eql(.init(.{num_experts}, .f32)) or - !isGlobalOrPerExpertScale(fc1_act_global, num_experts) or - !isGlobalOrPerExpertScale(fc2_act_global, num_experts)) - { - return error.InvalidShape; - } - - return .{ - .num_tokens = tokens, - .hidden_size = hidden_size, - .intermediate_size = intermediate_size, - .num_experts = @intCast(num_experts), - .top_k = @intCast(top_k), - .activation = cutlassActivation(options.activation), - .enable_pdl = options.enable_pdl, - .fc1_act_per_expert = fc1_act_global.rank() == 1, - .fc2_act_per_expert = fc2_act_global.rank() == 1, - .gemm1_tactic = options.gemm1_tactic, - .gemm2_tactic = options.gemm2_tactic, - }; -} - -/// Routed NVFP4 CUTLASS MoE. Weights are logical E2M1 tensors in [expert, output, input] order. -/// Block scales must use FlashInfer nvfp4_block_scale_interleave layout -/// BF16 activations are dynamically quantized to NVFP4 by the fused runner before each expert GEMM. -pub fn fusedExpertsNvfp4( - hidden_states: zml.Tensor, - fc1_weights: zml.Tensor, - fc2_weights: zml.Tensor, - topk_weights: zml.Tensor, - topk_ids: zml.Tensor, - fc1_act_global: zml.Tensor, - fc1_weight_block: zml.Tensor, - fc1_global: zml.Tensor, - fc2_act_global: zml.Tensor, - fc2_weight_block: zml.Tensor, - fc2_global: zml.Tensor, - options: Options, -) zml.Tensor { - const runners = currentRunners() catch |e| stdx.debug.panic("Failed to get runners: {}", .{e}); - var attributes = validateInputs( - hidden_states, - fc1_weights, - fc2_weights, - topk_weights, - topk_ids, - fc1_act_global, - fc1_weight_block, - fc1_global, - fc2_act_global, - fc2_weight_block, - fc2_global, - options, - ) catch |e| stdx.debug.panic("Invalid inputs: {}", .{e}); - - attributes.runners = @intFromPtr(runners); - const deviceRunner = runners.ensureRunner(options.workspace_query_device) catch |e| stdx.debug.panic("Failed to ensure runner: {}", .{e}); - const context = makeContext(attributes); - var requirements = std.mem.zeroes(fi_cutlass_moe.WorkspaceRequirements); - requirements.struct_size = @sizeOf(fi_cutlass_moe.WorkspaceRequirements); - checkStatus( - deviceRunner.api, - deviceRunner.api.getWorkspaceRequirements( - deviceRunner.runner, - &context, - &requirements, - ), - ) catch |e| stdx.debug.panic("Failed to get workspace requirements: {}", .{e}); - - const result = routedNvfp4Call.call( - .{ - .hidden_states = hidden_states, - .fc1_weights = fc1_weights, - .fc2_weights = fc2_weights, - .topk_weights = topk_weights, - .topk_ids = topk_ids, - .fc1_act_global = fc1_act_global, - .fc1_weight_block = fc1_weight_block, - .fc1_global = fc1_global, - .fc2_act_global = fc2_act_global, - .fc2_weight_block = fc2_weight_block, - .fc2_global = fc2_global, - }, - .{ - .output = hidden_states.shape(), - .workspace = .init(.{@as(i64, @intCast(requirements.total_bytes))}, .u8), - }, - attributes, - ); - return result.output; -} - -pub fn fusedExperts( - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - opts: zml.moe.Options, -) zml.Tensor { - if (comptime !platforms.isEnabled(.cuda)) { - @panic("FlashInfer CUTLASS MoE is only supported on CUDA platforms"); - } - - if (gate_up.bias != null or down.bias != null) { - @panic("FlashInfer CUTLASS MoE does not support bias in gate_up or down linear layers"); - } - - const runner_options: Options = .fromGlobalOptions(opts); - const expert_partition = gate_up.weight.shape().partition(.expert); - - const quant_scheme: ?zml.Quantization.Scheme = if (gate_up.quantization) |q| q.scheme else null; - stdx.debug.assert(quant_scheme == .nvfp4, "FlashInfer CUTLASS MoE only supports NVFP4, got {?}", .{quant_scheme}); - - const gate_up_weight_unpacked = zml.moe.unpackedWeight(gate_up); - const down_weight_unpacked = zml.moe.unpackedWeight(down); - - // TODO(Corentin): Do error checking on nvfp4 - // Also, maybe pass `zml.nn.Linear` directly - if (expert_partition.isSharded()) { - return zml.ops.manualComputation( - (struct { - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - gate_up_weight_unpacked: zml.Tensor, - down_weight_unpacked: zml.Tensor, - gate_up_input_scale: zml.Tensor, - gate_up_scales: zml.Tensor, - gate_up_global_scale: zml.Tensor, - down_input_scale: zml.Tensor, - down_scales: zml.Tensor, - down_global_scale: zml.Tensor, - activation: zml.moe.Activation, - enable_pdl: bool, - gemm1_tactic: i32, - gemm2_tactic: i32, - workspace_query_device: i32, - - fn body( - self: @This(), - _: zml.Shape, - ) zml.Tensor { - const local_num_experts = self.gate_up_weight_unpacked.dim(.expert); - const partition_id = zml.ops.partitionId().convert(.i32); - const expert_start = partition_id.scale(local_num_experts).convert(.i32); - const expert_end = expert_start.addConstant(local_num_experts); - - const local_route_mask = self.topk_ids - .cmp(.GE, expert_start) - .logical(.AND, self.topk_ids.cmp(.LT, expert_end)); - const local_topk_ids = local_route_mask.select( - self.topk_ids.sub(expert_start), - zml.Tensor.scalar(-1, .i32), - ); - const local_topk_weights = local_route_mask.select( - self.topk_weights, - zml.Tensor.scalar(-1, self.topk_weights.dtype()), - ); - - const local_output = fusedExpertsNvfp4( - self.input, - self.gate_up_weight_unpacked, - self.down_weight_unpacked, - local_topk_weights, - local_topk_ids, - self.gate_up_input_scale, - self.gate_up_scales, - self.gate_up_global_scale, - self.down_input_scale, - self.down_scales, - self.down_global_scale, - .{ - .workspace_query_device = self.workspace_query_device, - .activation = self.activation, - .enable_pdl = self.enable_pdl, - .gemm1_tactic = self.gemm1_tactic, - .gemm2_tactic = self.gemm2_tactic, - }, - ); - const local_reshaped = local_output - .reshape(self.input.shape().dims()) - .withTags(.{ .t, .d }); - return zml.ops.allReduce(local_reshaped, zml.Tensor.add); - } - }).body, - .{ - .input = input, - .topk_ids = topk_ids, - .topk_weights = topk_weights, - .gate_up_weight_unpacked = gate_up_weight_unpacked, - .down_weight_unpacked = down_weight_unpacked, - .gate_up_input_scale = gate_up.quantization.?.input_scale.?.asMultiplier(), - .gate_up_scales = gate_up.quantization.?.scales, - .gate_up_global_scale = gate_up.quantization.?.global_scale.?.asMultiplier(), - .down_input_scale = down.quantization.?.input_scale.?.asMultiplier(), - .down_scales = down.quantization.?.scales, - .down_global_scale = down.quantization.?.global_scale.?.asMultiplier(), - .activation = runner_options.activation, - .enable_pdl = runner_options.enable_pdl, - .gemm1_tactic = runner_options.gemm1_tactic, - .gemm2_tactic = runner_options.gemm2_tactic, - .workspace_query_device = runner_options.workspace_query_device, - }, - input.shape(), - ); - } - - return fusedExpertsNvfp4( - input, - gate_up_weight_unpacked, - down_weight_unpacked, - topk_weights, - topk_ids, - gate_up.quantization.?.input_scale.?.asMultiplier(), - gate_up.quantization.?.scales, - gate_up.quantization.?.global_scale.?.asMultiplier(), - down.quantization.?.input_scale.?.asMultiplier(), - down.quantization.?.scales, - down.quantization.?.global_scale.?.asMultiplier(), - runner_options, - ); -} diff --git a/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel b/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel new file mode 100644 index 00000000..c9c5d9a6 --- /dev/null +++ b/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel @@ -0,0 +1,10 @@ +package(default_visibility = ["//zml:__pkg__"]) + +filegroup( + name = "flashinfer_cutlass", + srcs = [ + "contract.zig", + "flashinfer_cutlass.zig", + "runtime.zig", + ], +) diff --git a/zml/kernels/moe/flashinfer_cutlass/contract.zig b/zml/kernels/moe/flashinfer_cutlass/contract.zig new file mode 100644 index 00000000..b4b77eaa --- /dev/null +++ b/zml/kernels/moe/flashinfer_cutlass/contract.zig @@ -0,0 +1,40 @@ +pub fn Inputs(comptime T: type) type { + return struct { + hidden_states: T, + fc1_weights: T, + fc2_weights: T, + topk_weights: T, + topk_ids: T, + fc1_act_global: T, + fc1_weight_block: T, + fc1_global: T, + fc2_act_global: T, + fc2_weight_block: T, + fc2_global: T, + }; +} + +pub fn Outputs(comptime T: type) type { + return struct { + output: T, + workspace: T, + }; +} + +pub const Attributes = struct { + runners: u64 = 0, + num_tokens: i64, + hidden_size: i64, + intermediate_size: i64, + num_experts: i32, + top_k: i32, + activation: i32, + enable_pdl: bool, + fc1_act_per_expert: bool, + fc2_act_per_expert: bool, + gemm1_tactic: i32, + gemm2_tactic: i32, +}; + +pub const compute_on_host: bool = false; +pub const name = "flashinfer_cutlass_nvfp4_routed_moe"; diff --git a/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig b/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig new file mode 100644 index 00000000..92317c7c --- /dev/null +++ b/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig @@ -0,0 +1,355 @@ +const std = @import("std"); + +const fi_cutlass_moe = @import("platforms/cuda/flashinfer_cutlass_moe"); +const platforms = @import("platforms"); +const stdx = @import("stdx"); + +const zml = @import("../../../zml.zig"); +const GlobalOptions = @import("../api.zig").Options; +const Activation = @import("../api.zig").Activation; +const LocalRouting = @import("../api.zig").LocalRouting; +const unpackedWeight = @import("../weights.zig").unpackedWeight; +const contract = @import("contract.zig"); +const Attributes = contract.Attributes; +const runtime = @import("runtime.zig"); + +const log = std.log.scoped(.@"zml/moe/flashinfer_cutlass"); + +pub const auto_tactic: i32 = -1; + +pub const Options = struct { + /// Device used at graph-construction time to select the architecture + /// library and size the XLA-owned scratch buffer. + workspace_query_device: i32, + activation: Activation, + enable_pdl: bool, + gemm1_tactic: i32, + /// GEMM2 uses FlashInfer's absolute tactic index. Query tacticCounts() to + /// obtain the first valid GEMM2 index. + gemm2_tactic: i32, + + pub fn fromGlobalOptions(opts: GlobalOptions) Options { + return .{ + .workspace_query_device = 0, + .activation = opts.activation, + .enable_pdl = false, + .gemm1_tactic = auto_tactic, + .gemm2_tactic = auto_tactic, + }; + } +}; + +pub fn validateOptions(opts: GlobalOptions) void { + stdx.debug.assert(!opts.quantize_input, "Optional FP8 input quantization requires the Triton MoE backend", .{}); + stdx.debug.assert(opts.routing_weight_placement == .after_down, "Non-Triton MoE backends require routing weights after the down projection", .{}); + stdx.debug.assert(opts.activation == .swiglu, "flashinfer_cutlass backend only accepts swiglu activation, got {}", .{opts.activation}); + stdx.debug.assert(opts.activation.swiglu.limit == null, "Activation thresholds require the Triton MoE backend", .{}); + stdx.debug.assert(opts.activation.swiglu.bias == null, "flashinfer_cutlass backend requires swiglu bias to be null", .{}); + stdx.debug.assert(opts.activation.swiglu.scale == null, "flashinfer_cutlass backend requires swiglu scale to be null", .{}); +} + +const Inputs = contract.Inputs(zml.Tensor); + +const Outputs = contract.Outputs(zml.Shape); + +fn currentRunners() !*runtime.Runners { + const platform = zml.Compiler.current().platform; + const runners = platform.state.cuda.flashinfer_cutlass_moe_runners; + return runners orelse error.RunnersNotLoaded; +} + +fn cutlassActivation(activation: Activation) i32 { + return switch (activation) { + .gelu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_GELU, + .relu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_RELU, + .silu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_SILU, + .swiglu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_SWIGLU, + .swiglu_step => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_SWIGLU_STEP, + .geglu => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_GEGLU, + .geglu_tanh => fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_ACTIVATION_GEGLU_TANH, + }; +} + +pub fn call(input_tensors: Inputs, output_shapes: Outputs, attributes: Attributes) zml.ops.ShapeToTensor(Outputs) { + const opts: zml.ops.CustomCallOptions = .{ + .has_side_effect = false, + .output_operand_aliases = comptime zml.ops.customCallOutputOperandAliases(Inputs, Outputs, null), + .compute_on_host = contract.compute_on_host, + }; + + return zml.ops.typedCustomCall( + contract.name, + opts, + input_tensors, + output_shapes, + attributes, + ); +} + +fn computeCapability(platform: *const zml.Platform) !u16 { + const cc = zml.platform.cuda.computeCapability(platform) orelse return error.UnsupportedPlatform; + return switch (cc.sm()) { + 100, 103, 120 => |sm| sm, + else => error.UnsupportedArchitecture, + }; +} + +pub fn isAvailable(platform: *const zml.Platform) bool { + if (platform.target != .cuda) return false; + if (platform.state.cuda.flashinfer_cutlass_moe_runners == null) return false; + _ = computeCapability(platform) catch return false; + return true; +} + +pub fn isNvfp4Supported(platform: *const zml.Platform) bool { + if (platform.state.cuda.flashinfer_cutlass_moe_runners == null) return false; + const compute_capability = computeCapability(platform) catch return false; + return switch (compute_capability) { + 100, 103, 120 => true, + else => false, + }; +} + +fn roundUp(value: i64, alignment: i64) i64 { + return @divTrunc(value + alignment - 1, alignment) * alignment; +} + +pub fn fc1BlockScaleShape( + num_experts: i64, + hidden_size: i64, + intermediate_size: i64, + activation: Activation, +) zml.Shape { + const rows = switch (activation) { + .silu, .gelu, .relu => intermediate_size, + .swiglu, .geglu, .geglu_tanh, .swiglu_step => 2 * intermediate_size, + }; + return .init( + .{ num_experts, roundUp(rows, 128), roundUp(@divExact(hidden_size, 16), 4) }, + .f8e4m3fn, + ); +} + +pub fn fc2BlockScaleShape( + num_experts: i64, + hidden_size: i64, + intermediate_size: i64, +) zml.Shape { + return .init( + .{ num_experts, roundUp(hidden_size, 128), roundUp(@divExact(intermediate_size, 16), 4) }, + .f8e4m3fn, + ); +} + +fn isGlobalOrPerExpertScale(tensor: zml.Tensor, num_experts: i64) bool { + return tensor.dtype() == .f32 and + (tensor.rank() == 0 or (tensor.rank() == 1 and tensor.dim(0) == num_experts)); +} + +fn validateInputs( + hidden_states: zml.Tensor, + fc1_weights: zml.Tensor, + fc2_weights: zml.Tensor, + topk_weights: zml.Tensor, + topk_ids: zml.Tensor, + fc1_act_global: zml.Tensor, + fc1_weight_block: zml.Tensor, + fc1_global: zml.Tensor, + fc2_act_global: zml.Tensor, + fc2_weight_block: zml.Tensor, + fc2_global: zml.Tensor, + options: Options, +) !Attributes { + if (hidden_states.dtype() != .bf16 or + fc1_weights.dtype() != .f4e2m1 or + fc2_weights.dtype() != .f4e2m1 or + topk_weights.dtype() != .f32 or + topk_ids.dtype() != .i32 or + fc1_weight_block.dtype() != .f8e4m3fn or + fc2_weight_block.dtype() != .f8e4m3fn or + fc1_global.dtype() != .f32 or + fc2_global.dtype() != .f32) + { + return error.UnsupportedType; + } + if (hidden_states.rank() != 2 or + fc1_weights.rank() != 3 or + fc2_weights.rank() != 3 or + topk_weights.rank() != 2 or + topk_ids.rank() != 2) + { + return error.InvalidShape; + } + + const tokens = hidden_states.dim(.t); + const hidden_size = hidden_states.dim(.d); + const num_experts = fc1_weights.dim(0); + const fc1_rows = fc1_weights.dim(1); + const intermediate_size = switch (options.activation) { + .silu, .gelu, .relu => fc1_rows, + .swiglu, .geglu, .geglu_tanh, .swiglu_step => @divExact(fc1_rows, 2), + }; + const top_k = topk_ids.dim(.topk); + + if (tokens <= 0 or hidden_size <= 0 or + num_experts <= 0 or intermediate_size <= 0 or + @mod(hidden_size, 16) != 0 or @mod(intermediate_size, 16) != 0 or + top_k <= 0 or top_k > num_experts) + { + return error.InvalidShape; + } + if (fc1_weights.dim(2) != hidden_size or + fc2_weights.dim(0) != num_experts or + fc2_weights.dim(1) != hidden_size or + fc2_weights.dim(2) != intermediate_size or + topk_weights.dim(.t) != tokens or + topk_weights.dim(.topk) != top_k or + topk_ids.dim(.t) != tokens or + !fc1_weight_block.shape().eql(fc1BlockScaleShape( + num_experts, + hidden_size, + intermediate_size, + options.activation, + )) or + !fc2_weight_block.shape().eql(fc2BlockScaleShape( + num_experts, + hidden_size, + intermediate_size, + )) or + !fc1_global.shape().eql(.init(.{num_experts}, .f32)) or + !fc2_global.shape().eql(.init(.{num_experts}, .f32)) or + !isGlobalOrPerExpertScale(fc1_act_global, num_experts) or + !isGlobalOrPerExpertScale(fc2_act_global, num_experts)) + { + return error.InvalidShape; + } + + return .{ + .num_tokens = tokens, + .hidden_size = hidden_size, + .intermediate_size = intermediate_size, + .num_experts = @intCast(num_experts), + .top_k = @intCast(top_k), + .activation = cutlassActivation(options.activation), + .enable_pdl = options.enable_pdl, + .fc1_act_per_expert = fc1_act_global.rank() == 1, + .fc2_act_per_expert = fc2_act_global.rank() == 1, + .gemm1_tactic = options.gemm1_tactic, + .gemm2_tactic = options.gemm2_tactic, + }; +} + +/// Routed NVFP4 CUTLASS MoE. Weights are logical E2M1 tensors in [expert, output, input] order. +/// Block scales must use FlashInfer nvfp4_block_scale_interleave layout +/// BF16 activations are dynamically quantized to NVFP4 by the fused runner before each expert GEMM. +pub fn fusedExpertsNvfp4( + hidden_states: zml.Tensor, + fc1_weights: zml.Tensor, + fc2_weights: zml.Tensor, + topk_weights: zml.Tensor, + topk_ids: zml.Tensor, + fc1_act_global: zml.Tensor, + fc1_weight_block: zml.Tensor, + fc1_global: zml.Tensor, + fc2_act_global: zml.Tensor, + fc2_weight_block: zml.Tensor, + fc2_global: zml.Tensor, + options: Options, +) zml.Tensor { + const runners = currentRunners() catch |e| stdx.debug.panic("Failed to get runners: {}", .{e}); + var attributes = validateInputs( + hidden_states, + fc1_weights, + fc2_weights, + topk_weights, + topk_ids, + fc1_act_global, + fc1_weight_block, + fc1_global, + fc2_act_global, + fc2_weight_block, + fc2_global, + options, + ) catch |e| stdx.debug.panic("Invalid inputs: {}", .{e}); + + attributes.runners = @intFromPtr(runners); + const deviceRunner = runners.ensureRunner(options.workspace_query_device) catch |e| stdx.debug.panic("Failed to ensure runner: {}", .{e}); + const context = runtime.makeContext(attributes); + var requirements = std.mem.zeroes(fi_cutlass_moe.WorkspaceRequirements); + requirements.struct_size = @sizeOf(fi_cutlass_moe.WorkspaceRequirements); + runtime.checkStatus( + deviceRunner.api, + deviceRunner.api.getWorkspaceRequirements( + deviceRunner.runner, + &context, + &requirements, + ), + ) catch |e| stdx.debug.panic("Failed to get workspace requirements: {}", .{e}); + + const result = call( + .{ + .hidden_states = hidden_states, + .fc1_weights = fc1_weights, + .fc2_weights = fc2_weights, + .topk_weights = topk_weights, + .topk_ids = topk_ids, + .fc1_act_global = fc1_act_global, + .fc1_weight_block = fc1_weight_block, + .fc1_global = fc1_global, + .fc2_act_global = fc2_act_global, + .fc2_weight_block = fc2_weight_block, + .fc2_global = fc2_global, + }, + .{ + .output = hidden_states.shape(), + .workspace = .init(.{@as(i64, @intCast(requirements.total_bytes))}, .u8), + }, + attributes, + ); + return result.output; +} + +pub fn fusedExperts( + input: zml.Tensor, + routing: LocalRouting, + gate_up: zml.nn.Linear, + down: zml.nn.Linear, + opts: GlobalOptions, +) zml.Tensor { + if (comptime !platforms.isEnabled(.cuda)) { + @panic("FlashInfer CUTLASS MoE is only supported on CUDA platforms"); + } + + if (gate_up.bias != null or down.bias != null) { + @panic("FlashInfer CUTLASS MoE does not support bias in gate_up or down linear layers"); + } + + const topk_ids = routing.localIds(); + // The runner ignores routes whose local expert id is -1. + const topk_weights = topk_ids.cmp(.GE, .scalar(0, .i32)).select(routing.weights, .scalar(-1, routing.weights.dtype())); + + const runner_options: Options = .fromGlobalOptions(opts); + + const quant_scheme: ?zml.Quantization.Scheme = if (gate_up.quantization) |q| q.scheme else null; + stdx.debug.assert(quant_scheme == .nvfp4, "FlashInfer CUTLASS MoE only supports NVFP4, got {?}", .{quant_scheme}); + + const gate_up_weight_unpacked = unpackedWeight(gate_up); + const down_weight_unpacked = unpackedWeight(down); + + // TODO(Corentin): Do error checking on nvfp4 + // Also, maybe pass `zml.nn.Linear` directly + return fusedExpertsNvfp4( + input, + gate_up_weight_unpacked, + down_weight_unpacked, + topk_weights, + topk_ids, + gate_up.quantization.?.input_scale.?.asMultiplier(), + gate_up.quantization.?.scales, + gate_up.quantization.?.global_scale.?.asMultiplier(), + down.quantization.?.input_scale.?.asMultiplier(), + down.quantization.?.scales, + down.quantization.?.global_scale.?.asMultiplier(), + runner_options, + ); +} diff --git a/zml/kernels/moe/flashinfer_cutlass/runtime.zig b/zml/kernels/moe/flashinfer_cutlass/runtime.zig new file mode 100644 index 00000000..028ef892 --- /dev/null +++ b/zml/kernels/moe/flashinfer_cutlass/runtime.zig @@ -0,0 +1,223 @@ +const std = @import("std"); + +const fi_cutlass_moe = @import("platforms/cuda/flashinfer_cutlass_moe"); +const pjrt = @import("pjrt"); + +const DataType = @import("../../../dtype.zig").DataType; +const contract = @import("contract.zig"); + +const log = std.log.scoped(.@"zml/kernels/moe/flashinfer_cutlass/runtime"); + +// TODO(Corentin): This was coming from Platform but we are now downstream. +// Find a way to solve that. +// Maybe the max_num_devices could be its own module because it only depends on Target. +const max_num_devices = 64; + +// TODO(Corentin): replace that with Shape when its dependencies are untangled and +// it doesn't depend on the platform anymore. +fn count(buffer: *const pjrt.ffi.Buffer) usize { + if (buffer.rank == 0) return @intCast(buffer.dims()[0]); + + var c: usize = 1; + for (buffer.dims()) |dim| c = c * @as(usize, @intCast(dim)); + + return c; +} + +fn byteSize(buffer: *const pjrt.ffi.Buffer) usize { + const dt: DataType = switch (buffer.dtype) { + .invalid => @panic("Found an invalid pjrt buffer"), + .token => @panic("Found an token pjrt buffer"), + inline else => |tag| @field(DataType, @tagName(tag)), + }; + return std.math.divCeil(usize, count(buffer) * dt.bitSizeOf(), 8) catch unreachable; +} + +pub fn handler(call_frame: *pjrt.ffi.CallFrame) callconv(.c) ?*pjrt.ffi.Error { + if (call_frame.registeringHook()) return null; + + // TODO(Corentin): Find a place to abstract that + var input_buffers: contract.Inputs(*const pjrt.ffi.Buffer) = undefined; + inline for (@typeInfo(contract.Inputs(*const pjrt.ffi.Buffer)).@"struct".field_names, 0..) |field_name, i| { + const buf = call_frame.args.buffers()[i]; + @field(input_buffers, field_name) = buf; + } + + var output_buffers: contract.Outputs(*const pjrt.ffi.Buffer) = undefined; + inline for (@typeInfo(contract.Outputs(*const pjrt.ffi.Buffer)).@"struct".field_names, 0..) |field_name, i| { + const buf = call_frame.results.buffers()[i]; + @field(output_buffers, field_name) = buf; + } + + const attributes: contract.Attributes = switch (@typeInfo(contract.Attributes)) { + .void => {}, + .@"struct" => blk: { + var attrs: contract.Attributes = undefined; + inline for (@typeInfo(contract.Attributes).@"struct".field_names, @typeInfo(contract.Attributes).@"struct".field_types) |field_name, Field| { + // TODO: Use getByIndex instead? + const attribute = call_frame.attrs.getByName(.scalar, field_name) orelse + @panic("Attribute not found: " ++ field_name); + @field(attrs, field_name) = attribute.get(Field); + } + break :blk attrs; + }, + else => @compileError("Unsupported attributes type: " ++ @typeName(contract.Attributes)), + }; + + return handlerInner(call_frame, input_buffers, output_buffers, attributes) catch |err| + pjrt.ffi.Error.create(call_frame.api, .unknown, @errorName(err)); +} + +fn handlerInner( + call_frame: *pjrt.ffi.CallFrame, + inputs: contract.Inputs(*const pjrt.ffi.Buffer), + outputs: contract.Outputs(*const pjrt.ffi.Buffer), + attributes: contract.Attributes, +) !?*pjrt.ffi.Error { + const device = try call_frame.ctx.getDeviceOrdinal(call_frame.api); + const runners = try runnersFromAttributes(attributes); + const deviceRunner = try runners.ensureRunner(device); + const context = makeContext(attributes); + + var io = std.mem.zeroes(fi_cutlass_moe.Io); + io.struct_size = @sizeOf(fi_cutlass_moe.Io); + io.input = inputs.hidden_states.data; + io.token_selected_experts = @ptrCast(@alignCast(inputs.topk_ids.data)); + io.token_final_scales = @ptrCast(@alignCast(inputs.topk_weights.data)); + io.fc1_expert_weights = inputs.fc1_weights.data; + io.fc2_expert_weights = inputs.fc2_weights.data; + io.quant_scales[0] = inputs.fc1_act_global.data; + io.quant_scales[1] = inputs.fc1_weight_block.data; + io.quant_scales[2] = inputs.fc1_global.data; + io.quant_scales[3] = inputs.fc2_act_global.data; + io.quant_scales[4] = inputs.fc2_weight_block.data; + io.quant_scales[5] = inputs.fc2_global.data; + io.quant_scale_count = 6; + io.quant_scale_per_expert_mask = + (@as(u32, @intFromBool(attributes.fc1_act_per_expert)) << 0) | + (@as(u32, @intFromBool(attributes.fc2_act_per_expert)) << 3); + io.output = outputs.output.data; + + var workspace = std.mem.zeroes(fi_cutlass_moe.Workspace); + workspace.struct_size = @sizeOf(fi_cutlass_moe.Workspace); + workspace.data = outputs.workspace.data; + workspace.data_bytes = byteSize(outputs.workspace); + + try checkStatus( + deviceRunner.api, + deviceRunner.api.run( + deviceRunner.runner, + &context, + &io, + &workspace, + @ptrCast(call_frame.api.stream(call_frame.ctx)), + attributes.gemm1_tactic, + attributes.gemm2_tactic, + ), + ); + return null; +} + +pub fn makeContext(attributes: contract.Attributes) fi_cutlass_moe.Context { + var context = std.mem.zeroes(fi_cutlass_moe.Context); + context.struct_size = @sizeOf(fi_cutlass_moe.Context); + context.num_tokens = attributes.num_tokens; + context.hidden_size = attributes.hidden_size; + context.intermediate_size = attributes.intermediate_size; + context.num_experts = attributes.num_experts; + context.num_experts_on_rank = attributes.num_experts; + context.top_k = attributes.top_k; + context.tp_size = 1; + context.ep_size = 1; + context.activation = @intCast(attributes.activation); + context.enable_pdl = @intFromBool(attributes.enable_pdl); + context.swizzled_input_sf = 1; + return context; +} + +pub fn checkStatus(api: *const fi_cutlass_moe.Api, status: fi_cutlass_moe.Status) !void { + if (status == fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_SUCCESS) return; + if (api.lastError()) |message| { + log.err("FlashInfer CUTLASS MoE failed (status {d}): {s}", .{ + status, + std.mem.span(message), + }); + } else { + log.err("FlashInfer CUTLASS MoE failed with status {d}", .{status}); + } + return switch (status) { + fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_INVALID_ARGUMENT => error.InvalidArgument, + fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_UNSUPPORTED => error.UnsupportedArchitecture, + fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_STATUS_CUDA_ERROR => error.Cuda, + else => error.FlashinferCutlassMoe, + }; +} + +fn runnersFromAttributes(attributes: contract.Attributes) !*Runners { + if (attributes.runners == 0) return error.RunnersNotLoaded; + return @ptrFromInt(attributes.runners); +} + +pub const Runners = struct { + runners: [max_num_devices]?DeviceRunner = @splat(null), + + pub fn deinit(self: *Runners) void { + for (self.runners) |device_runner| { + if (device_runner) |runner| { + _ = runner.api.runnerDestroy(runner.runner); + } + } + } + + pub fn ensureRunner(self: *Runners, device: i32) !DeviceRunner { + if (device < 0 or device >= max_num_devices) return error.UnsupportedDevice; + + const index: usize = @intCast(device); + if (self.runners[index]) |runner| return runner; + + const api = try fi_cutlass_moe.apiForDevice(device); + const options = runnerOptions(device); + var runner: ?*fi_cutlass_moe.Runner = null; + try checkStatus(api, api.runnerCreate(&options, &runner)); + const result: DeviceRunner = .{ + .api = api, + .runner = runner orelse return error.RunnerInitializationFailed, + }; + self.runners[index] = result; + return result; + } +}; + +const DeviceRunner = struct { + api: *fi_cutlass_moe.Api, + runner: *fi_cutlass_moe.Runner, +}; + +fn runnerOptions(device: i32) fi_cutlass_moe.RunnerOptions { + var options = std.mem.zeroes(fi_cutlass_moe.RunnerOptions); + options.struct_size = @sizeOf(fi_cutlass_moe.RunnerOptions); + options.activation_dtype = fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_DTYPE_BF16; + options.weight_dtype = fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_DTYPE_PACKED_FP4; + options.output_dtype = fi_cutlass_moe.c.ZML_FI_CUTLASS_MOE_DTYPE_BF16; + options.device = device; + options.use_fused_finalize = 1; + return options; +} + +pub fn tacticCounts( + runners: *const Runners, + device: i32, +) !struct { gemm1: i32, gemm2: i32 } { + const deviceRunner = try runners.ensureRunner(device); + var gemm1: i32 = 0; + var gemm2: i32 = 0; + try checkStatus( + deviceRunner.api, + deviceRunner.api.getTacticCounts(deviceRunner.runner, &gemm1, &gemm2), + ); + return .{ .gemm1 = gemm1, .gemm2 = gemm2 }; +} + +pub fn load(allocator: std.mem.Allocator, io: std.Io, sm: u16) !void { + try fi_cutlass_moe.load(allocator, io, sm); +} diff --git a/zml/kernels/moe/fly/fly.zig b/zml/kernels/moe/fly/fly.zig index e5c5b6de..ba3d3c00 100644 --- a/zml/kernels/moe/fly/fly.zig +++ b/zml/kernels/moe/fly/fly.zig @@ -6,9 +6,13 @@ const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; const Shape = zml.Shape; const DataType = zml.DataType; +const ExpertsLayout = @import("../api.zig").ExpertsLayout; +const ProjectionLayout = @import("../api.zig").ProjectionLayout; +const RoutingWeightPlacement = @import("../api.zig").RoutingWeightPlacement; +const Activation = @import("../api.zig").Activation; +const Options = @import("../api.zig").Options; +const LocalRouting = @import("../api.zig").LocalRouting; const fly = @import("moe.zig"); -const ProjectionLayout = zml.moe.ProjectionLayout; -const RoutingWeightPlacement = zml.moe.RoutingWeightPlacement; const log = std.log.scoped(.@"zml/moe/fused_experts"); @@ -17,90 +21,19 @@ test { } pub fn fusedExperts( - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - opts: zml.moe.Options, -) zml.Tensor { - const gate_up_scheme: ?zml.Quantization.Scheme = if (gate_up.quantization) |q| q.scheme else null; - const layout: zml.moe.ExpertsLayout = if (gate_up_scheme == .mxfp4) - .{ .gate_up = .interleaved, .packing = .plain } - else - .{ .gate_up = .concatenated, .packing = .plain }; - - const args: FusedExpertsArgs = .{ - .hidden_states = input, - .gate_up = gate_up, - .down = down, - .topk_weights = topk_weights, - .topk_ids = topk_ids, - .activation = opts.activation, - .quantize_input = opts.quantize_input, - .gate_up_layout = layout.gate_up, - .routing_weight_placement = opts.routing_weight_placement, - }; - const expert_partition = gate_up.weight.shape().partition(.expert); - - if (!expert_partition.isSharded()) { - return fusedExpertsImpl(args); - } - - return zml.ops.manualComputation( - (struct { - args: FusedExpertsArgs, - global_num_experts: i64, - - fn call(self: @This(), _: zml.Shape) zml.Tensor { - const local_args = self.args; - const local_num_experts = local_args.gate_up.weight.dim(.expert); - const partition_id = zml.ops.partitionId().convert(.i32); - const expert_start = partition_id.scale(local_num_experts).convert(.i32); - const global_expert_ids = zml.Tensor.arange(.{ .end = self.global_num_experts }, .i32).withTags(.{.expert}); - - // Map global expert ids to local ids, or -1 for experts outside this partition. - const local_expert_mask = global_expert_ids.cmp(.GE, expert_start) - .logical(.AND, global_expert_ids.cmp(.LT, expert_start.addConstant(local_num_experts))); - var mapped_args = local_args; - mapped_args.expert_map = local_expert_mask.select( - global_expert_ids.sub(expert_start), - zml.Tensor.scalar(-1, .i32), - ); - - const local_output = fusedExpertsImpl(mapped_args); - const local_reshaped = local_output.reshape(local_args.hidden_states.shape()).withTags(.{ .t, .d }); - return zml.ops.allReduce(local_reshaped, zml.Tensor.add); - } - }).call, - .{ .args = args, .global_num_experts = gate_up.weight.dim(.expert) }, - input.shape(), - ); -} - -pub const FusedExpertsArgs = struct { hidden_states: Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - topk_weights: Tensor, - topk_ids: Tensor, - activation: zml.moe.Activation, - expert_map: ?Tensor = null, - /// Use FP8 activations for FP8 weights; false keeps BF16 activations. - quantize_input: bool, - gate_up_layout: ProjectionLayout, - routing_weight_placement: RoutingWeightPlacement, -}; - -pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { - const hidden_states = opts.hidden_states; - const topk_weights = opts.topk_weights; - const topk_ids = opts.topk_ids; + routing_: LocalRouting, + gate_up_: zml.nn.Linear, + down_: zml.nn.Linear, + opts: Options, +) Tensor { + const expert_map_ = routing_.expertMap(); + const layout: ProjectionLayout = if (gate_up_.quantizationScheme() == .mxfp4) .row_interleaved_1 else .concatenated; const num_tokens = hidden_states.dim(.t); var launch_config = launchConfigForTokens(num_tokens); - const gate_up_scheme = opts.gate_up.quantizationScheme(); - const down_scheme = opts.down.quantizationScheme(); + const gate_up_scheme = gate_up_.quantizationScheme(); + const down_scheme = down_.quantizationScheme(); if (gate_up_scheme) |scheme| switch (scheme) { .nvfp4 => stdx.debug.panic("fly backend doesn't support gate_up nvfp4 quantization scheme", .{}), .fp8_block128 => launch_config.block_size_k = 128, @@ -117,10 +50,10 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { }; const hidden = hidden_states.reshape(.{ .token = num_tokens, .in = hidden_states.dim(.d) }).withTags(.{ .token, .in }); - const gate_up = opts.gate_up.weight.withTags(.{ .expert, .out, .in }); - const down = opts.down.weight.withTags(.{ .expert, .out, .mid }); - const routing_weights = topk_weights.reshape(.{ .token = num_tokens, .in = topk_weights.dim(.topk) }).withTags(.{ .token, .topk }); - const ids = topk_ids.reshape(.{ .token = num_tokens, .in = topk_ids.dim(.topk) }).withTags(.{ .token, .topk }); + const gate_up = gate_up_.weight.withTags(.{ .expert, .out, .in }); + const down = down_.weight.withTags(.{ .expert, .out, .mid }); + const routing_weights = routing_.weights.reshape(.{ .token = num_tokens, .in = routing_.weights.dim(.topk) }).withTags(.{ .token, .topk }); + const ids = routing_.ids.reshape(.{ .token = num_tokens, .in = routing_.ids.dim(.topk) }).withTags(.{ .token, .topk }); stdx.debug.assert(hidden.dtype() == .bf16, "fly backend expected BF16 hidden states, got {}", .{hidden.dtype()}); if (gate_up_scheme == .mxfp4) { @@ -148,13 +81,13 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { stdx.debug.assert(ids.dim(.token) == hidden.dim(.token) and routing_weights.dim(.token) == hidden.dim(.token), "fly backend expected routing ids and weights to match hidden token count, got {}, {} and {}", .{ ids.dim(.token), routing_weights.dim(.token), hidden.dim(.token) }); stdx.debug.assert(ids.dim(.topk) == routing_weights.dim(.topk), "fly backend expected routing ids and weights to have matching top-k dimensions, got {} and {}", .{ ids.dim(.topk), routing_weights.dim(.topk) }); stdx.debug.assert(gate_up.dim(.expert) == down.dim(.expert), "fly backend expected gate/up and down expert counts to match, got {} and {}", .{ gate_up.dim(.expert), down.dim(.expert) }); - if (opts.expert_map) |expert_map| + if (expert_map_) |expert_map| stdx.debug.assert(expert_map.dtype() == .i32 and expert_map.rank() == 1, "fly backend expected a rank-1 i32 expert map, got rank {} and dtype {}", .{ expert_map.rank(), expert_map.dtype() }); - const num_experts = if (opts.expert_map) |expert_map| expert_map.dim(.expert) else gate_up.dim(.expert); + const num_experts = if (expert_map_) |expert_map| expert_map.dim(.expert) else gate_up.dim(.expert); const routing = prepareRouting(ids, num_experts, @intCast(launch_config.block_size_m)); - const expert_ids = if (opts.expert_map) |expert_map| + const expert_ids = if (expert_map_) |expert_map| expert_map.gather(.{ .expert = routing.expert_ids }, .{}).withTags(.{.g}) else routing.expert_ids; @@ -164,9 +97,9 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { const gate_up_out = gemm(.{ .input = hidden_quant, .weight = gate_up, - .bias = opts.gate_up.bias, + .bias = gate_up_.bias, .input_scale = input_scale, - .weight_scale = opts.gate_up.quantizationScales(), + .weight_scale = gate_up_.quantizationScales(), .routing = routing, .expert_ids = expert_ids, .launch_config = launch_config, @@ -175,7 +108,7 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { .output_shape = Shape.init(.{ .token = routing.num_assignments, .out = gate_up.dim(.out) }, .bf16), }); - var activated = zml.moe.applyActivation(gate_up_out, opts.activation, opts.gate_up_layout); + var activated = opts.activation.applyActivation(gate_up_out, layout); if (opts.routing_weight_placement == .before_down) { const weights = routing_weights.reshape(.{ .token = routing.num_assignments }).convert(.f32); activated = activated.mul(weights.broad(activated.shape())); @@ -185,9 +118,9 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { const down_out = gemm(.{ .input = activated_quant, .weight = down, - .bias = opts.down.bias, + .bias = down_.bias, .input_scale = down_input_scale, - .weight_scale = opts.down.quantizationScales(), + .weight_scale = down_.quantizationScales(), .routing_weights = if (opts.routing_weight_placement == .after_down) routing_weights else null, .routing = routing, .expert_ids = expert_ids, @@ -203,18 +136,18 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { } // TODO(Corentin): unused for now -pub fn supports(opts: FusedExpertsArgs) bool { - if (opts.hidden_states.dtype() != .bf16 or opts.gate_up.quantizationScheme() != .mxfp4 or - opts.down.quantizationScheme() != .mxfp4 or opts.gate_up.bias != null or opts.down.bias != null or +pub fn supports(hidden_states: Tensor, routing: LocalRouting, gate_up_: zml.nn.Linear, down_: zml.nn.Linear, opts: Options) bool { + if (hidden_states.dtype() != .bf16 or gate_up_.quantizationScheme() != .mxfp4 or + down_.quantizationScheme() != .mxfp4 or gate_up_.bias != null or down_.bias != null or opts.routing_weight_placement != .before_down) return false; - const gate_up = opts.gate_up.weight; - const down = opts.down.weight; + const gate_up = gate_up_.weight; + const down = down_.weight; if (gate_up.rank() != 3 or down.rank() != 3) return false; - const gate_up_scale = opts.gate_up.quantizationScales() orelse return false; - const down_scale = opts.down.quantizationScales() orelse return false; - const experts = if (opts.expert_map) |map| map.dim(.expert) else gate_up.dim(0); + const gate_up_scale = gate_up_.quantizationScales() orelse return false; + const down_scale = down_.quantizationScales() orelse return false; + const experts = routing.global_num_experts; if (experts < 1 or experts > 512) return false; - var max_width: i64 = opts.hidden_states.dim(.d); + var max_width: i64 = hidden_states.dim(.d); inline for (.{ .{ gate_up, gate_up_scale }, .{ down, down_scale } }) |projection| { const weight, const scale = projection; const packing_factor: i64 = switch (weight.dtype()) { @@ -233,7 +166,7 @@ pub fn supports(opts: FusedExpertsArgs) bool { else => return false, } } - const routes = opts.hidden_states.dim(.t) * opts.topk_ids.dim(.topk); + const routes = hidden_states.dim(.t) * routing.ids.dim(.topk); return routes > 0 and routes <= @divTrunc(std.math.maxInt(i32), max_width); } diff --git a/zml/kernels/moe/metal/metal.zig b/zml/kernels/moe/metal/metal.zig index e129c76c..f835c889 100644 --- a/zml/kernels/moe/metal/metal.zig +++ b/zml/kernels/moe/metal/metal.zig @@ -1,21 +1,17 @@ const std = @import("std"); + const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; +const Activation = @import("../api.zig").Activation; +const GlobalOptions = @import("../api.zig").Options; +const LocalRouting = @import("../api.zig").LocalRouting; +const unpackedWeight = @import("../weights.zig").unpackedWeight; const log = std.log.scoped(.@"zml/moe/metal"); -pub const Options = struct { - activation: zml.moe.Activation, - global_num_experts: i64 = -1, - w1_scale: ?Tensor = null, - w2_scale: ?Tensor = null, - w1_global_scale: ?Tensor = null, - w2_global_scale: ?Tensor = null, -}; - -fn validateOptions(opts: zml.moe.Options) void { +fn validateOptions(opts: GlobalOptions) void { stdx.debug.assert(!opts.quantize_input, "Optional FP8 input quantization requires the Triton MoE backend", .{}); stdx.debug.assert(opts.routing_weight_placement == .after_down, "Non-Triton MoE backends require routing weights after the down projection", .{}); } @@ -72,57 +68,30 @@ fn applyDownGlobalScale(output: Tensor, global_scale: ?Tensor, expert_ids: Tenso } pub fn fusedExperts( - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - opts: zml.moe.Options, + hidden_states: zml.Tensor, + routing: LocalRouting, + gate_up_: zml.nn.Linear, + down_: zml.nn.Linear, + opts: GlobalOptions, ) zml.Tensor { validateOptions(opts); - stdx.debug.assert(gate_up.bias == null, "metal backend doesn't support gate_up bias", .{}); - stdx.debug.assert(down.bias == null, "metal backend doesn't support down bias", .{}); + stdx.debug.assert(gate_up_.bias == null, "metal backend doesn't support gate_up bias", .{}); + stdx.debug.assert(down_.bias == null, "metal backend doesn't support down bias", .{}); - const gate_up_scales: ?zml.Tensor = if (gate_up.quantization) |q| q.scales else null; - const gate_up_global_scale: ?zml.Tensor = if (gate_up.quantization) |q| (if (q.global_scale) |scale| scale.asMultiplier() else null) else null; + const gate_up_scales: ?zml.Tensor = if (gate_up_.quantization) |q| q.scales else null; + const gate_up_global_scale: ?zml.Tensor = if (gate_up_.quantization) |q| (if (q.global_scale) |scale| scale.asMultiplier() else null) else null; - const down_scales: ?zml.Tensor = if (down.quantization) |q| q.scales else null; - const down_global_scale: ?zml.Tensor = if (down.quantization) |q| (if (q.global_scale) |scale| scale.asMultiplier() else null) else null; + const down_scales: ?zml.Tensor = if (down_.quantization) |q| q.scales else null; + const down_global_scale: ?zml.Tensor = if (down_.quantization) |q| (if (q.global_scale) |scale| scale.asMultiplier() else null) else null; - const gate_up_weight_unpacked = zml.moe.unpackedWeight(gate_up); - const down_weight_unpacked = zml.moe.unpackedWeight(down); - - return fusedExpertsImpl( - input, - gate_up_weight_unpacked, - down_weight_unpacked, - topk_weights, - topk_ids, - .{ - .activation = opts.activation, - .global_num_experts = gate_up_weight_unpacked.dim(.expert), - .w1_scale = gate_up_scales, - .w2_scale = down_scales, - .w1_global_scale = gate_up_global_scale, - .w2_global_scale = down_global_scale, - }, - ); -} - -pub fn fusedExpertsImpl( - hidden_states: Tensor, - gate_up: Tensor, - down: Tensor, - topk_weights: Tensor, - topk_ids: Tensor, - opts: Options, -) Tensor { + const gate_up = unpackedWeight(gate_up_); + const down = unpackedWeight(down_); stdx.debug.assert(gate_up.dtype() == down.dtype(), "metal backend expected gate_up and down data type to match, got {} and {}", .{ gate_up.dtype(), down.dtype() }); const mode = quantMode(gate_up.dtype()); switch (mode) { .none => { - stdx.debug.assert(opts.w1_scale == null and opts.w2_scale == null, "metal backend expected gate_up and down scale to be null when the weights are not quantized", .{}); - stdx.debug.assert(opts.w1_global_scale == null and opts.w2_global_scale == null, "metal backend expected gate_up and down global scale to be null when the weights are not quantized", .{}); + stdx.debug.assert(gate_up_scales == null and down_scales == null, "metal backend expected gate_up and down scale to be null when the weights are not quantized", .{}); + stdx.debug.assert(gate_up_global_scale == null and down_global_scale == null, "metal backend expected gate_up and down global scale to be null when the weights are not quantized", .{}); const is_hidden_states_valid_dtype = switch (hidden_states.dtype()) { .bf16, .f16, .f32 => true, else => false, @@ -130,28 +99,28 @@ pub fn fusedExpertsImpl( stdx.debug.assert(is_hidden_states_valid_dtype, "metal backend expected hidden_states data type to be bf16, f16 or f32, got {}", .{hidden_states.dtype()}); }, .fp8 => { - stdx.debug.assert(opts.w1_scale != null and opts.w2_scale != null, "metal backend expected gate_up and down scale to be non null", .{}); - stdx.debug.assert(opts.w1_scale.?.dtype() == .bf16 and opts.w2_scale.?.dtype() == .bf16, "metal backend expected gate_up and down scale data type to be bf16, got {} and {}", .{ opts.w1_scale.?.dtype(), opts.w2_scale.?.dtype() }); - stdx.debug.assert(opts.w1_global_scale == null and opts.w2_global_scale == null, "metal backend expected gate_up and down global scale to be null", .{}); + stdx.debug.assert(gate_up_scales != null and down_scales != null, "metal backend expected gate_up and down scale to be non null", .{}); + stdx.debug.assert(gate_up_scales.?.dtype() == .bf16 and down_scales.?.dtype() == .bf16, "metal backend expected gate_up and down scale data type to be bf16, got {} and {}", .{ gate_up_scales.?.dtype(), down_scales.?.dtype() }); + stdx.debug.assert(gate_up_global_scale == null and down_global_scale == null, "metal backend expected gate_up and down global scale to be null", .{}); stdx.debug.assert(hidden_states.dtype() == .bf16, "metal backend expected hidden_states data type to be bf16, got {}", .{hidden_states.dtype()}); }, .nvfp4 => { - stdx.debug.assert(opts.w1_scale != null and opts.w2_scale != null, "metal backend expected gate_up and down scale to be non null", .{}); - stdx.debug.assert(opts.w1_scale.?.dtype() == .f8e4m3fn and opts.w2_scale.?.dtype() == .f8e4m3fn, "metal backend expected gate_up and down scale data type to be f8e4m3fn, got {} and {}", .{ opts.w1_scale.?.dtype(), opts.w2_scale.?.dtype() }); + stdx.debug.assert(gate_up_scales != null and down_scales != null, "metal backend expected gate_up and down scale to be non null", .{}); + stdx.debug.assert(gate_up_scales.?.dtype() == .f8e4m3fn and down_scales.?.dtype() == .f8e4m3fn, "metal backend expected gate_up and down scale data type to be f8e4m3fn, got {} and {}", .{ gate_up_scales.?.dtype(), down_scales.?.dtype() }); stdx.debug.assert(hidden_states.dtype() == .bf16, "metal backend expected hidden_states data type to be bf16, got {}", .{hidden_states.dtype()}); }, } - const is_topk_weights_valid_dtype = switch (topk_weights.dtype()) { + const is_topk_weights_valid_dtype = switch (routing.weights.dtype()) { .bf16, .f16, .f32 => true, else => false, }; - stdx.debug.assert(is_topk_weights_valid_dtype, "metal backend expected topk_weights data type to be bf16, f16 or f32, got {}", .{topk_weights.dtype()}); - stdx.debug.assert(topk_ids.dtype() == .i32, "metal backend expected topk_ids data type to be i32, got {}", .{topk_ids.dtype()}); + stdx.debug.assert(is_topk_weights_valid_dtype, "metal backend expected topk_weights data type to be bf16, f16 or f32, got {}", .{routing.weights.dtype()}); + stdx.debug.assert(routing.ids.dtype() == .i32, "metal backend expected topk_ids data type to be i32, got {}", .{routing.ids.dtype()}); const d = hidden_states.dim(.d); const num_tokens = hidden_states.dim(.t); - const topk = topk_ids.dim(.topk); + const topk = routing.ids.dim(.topk); const num_routes = num_tokens * topk; const act_dtype: zml.DataType = switch (mode) { @@ -163,29 +132,29 @@ pub fn fusedExpertsImpl( const x_rows = hidden.insertAxes(.d, .{.topk}) .broad(zml.Shape.init(.{ .token = num_tokens, .topk = topk, .d = d }, act_dtype)) .merge(.{ .r = .{ .token, .topk } }); - const expert_ids = topk_ids.reshape(.{ .token = num_tokens, .topk = topk }).withTags(.{ .token, .topk }) + const expert_ids = routing.ids.reshape(.{ .token = num_tokens, .topk = topk }).withTags(.{ .token, .topk }) .merge(.{ .r = .{ .token, .topk } }).convert(.i32); const gate_up_out = applyGateUpGlobalScale(moeGemm( x_rows, gate_up, - opts.w1_scale, + gate_up_scales, expert_ids, zml.Shape.init(.{ .r = num_routes, .out = gate_up.dim(.dout) }, act_dtype), mode, - ), opts.w1_global_scale, expert_ids); - const activated = zml.moe.applyActivation(gate_up_out, opts.activation, .concatenated).convert(act_dtype); + ), gate_up_global_scale, expert_ids); + const activated = opts.activation.applyActivation(gate_up_out, .concatenated).convert(act_dtype); const down_out = applyDownGlobalScale(moeGemm( activated, down, - opts.w2_scale, + down_scales, expert_ids, zml.Shape.init(.{ .r = num_routes, .d = down.dim(.d) }, act_dtype), mode, - ), opts.w2_global_scale, expert_ids); + ), down_global_scale, expert_ids); - const weights = topk_weights.reshape(.{ .token = num_tokens, .topk = topk }).withTags(.{ .token, .topk }) + const weights = routing.weights.reshape(.{ .token = num_tokens, .topk = topk }).withTags(.{ .token, .topk }) .merge(.{ .r = .{ .token, .topk } }); const weighted = down_out.mul(weights.convert(down_out.dtype()).broad(down_out.shape())); const combined = weighted.splitAxis(.r, .{ .token = num_tokens, .topk = topk }) diff --git a/zml/kernels/moe/moe.zig b/zml/kernels/moe/moe.zig index 25ce26d6..eb697f36 100644 --- a/zml/kernels/moe/moe.zig +++ b/zml/kernels/moe/moe.zig @@ -4,14 +4,21 @@ const platforms = @import("platforms"); const zml = @import("../../zml.zig"); const stdx = zml.stdx; -pub const tcl = @import("tcl/tcl.zig"); -pub const stablehlo = @import("stablehlo/stablehlo.zig"); +pub const Activation = @import("api.zig").Activation; +pub const ActivationKind = @import("api.zig").ActivationKind; pub const cute = @import("cute/cute.zig"); -pub const cutlass_flashinfer = @import("cutlass_flashinfer/cutlass_flashinfer.zig"); +pub const ExpertsLayout = @import("api.zig").ExpertsLayout; +pub const flashinfer_cutlass = @import("flashinfer_cutlass/flashinfer_cutlass.zig"); +pub const fly = @import("fly/fly.zig"); +pub const LocalRouting = @import("api.zig").LocalRouting; pub const metal = @import("metal/metal.zig"); pub const mosaic_tpu = @import("mosaic_tpu/mosaic_tpu.zig"); +pub const Options = @import("api.zig").Options; +pub const ProjectionLayout = @import("api.zig").ProjectionLayout; +pub const RoutingWeightPlacement = @import("api.zig").RoutingWeightPlacement; +pub const stablehlo = @import("stablehlo/stablehlo.zig"); +pub const tcl = @import("tcl/tcl.zig"); pub const triton = @import("triton/triton.zig"); -pub const fly = @import("fly/fly.zig"); const log = std.log.scoped(.@"zml/moe"); @@ -20,60 +27,6 @@ test { _ = @import("tests.zig"); } -pub const ProjectionLayout = enum { concatenated, interleaved }; -pub const RoutingWeightPlacement = enum { before_down, after_down }; - -/// How a backend expects the expert weights to be stored -/// The packer writes this layout and forwardMoe reads it -pub const ExpertsLayout = struct { - /// Whether the two projections occupy contiguous halves or alternate. - gate_up: ProjectionLayout, - /// Backend specific transformation of the weights and scales. - packing: Packing, - - pub const Packing = enum { - /// Experts stacked as stored in the checkpoint. - plain, - /// Block scales swizzled to the 128x4 tensor-core layout (rows tiled by 4x32, columns by 4). - swizzled_scales, - /// Weights, block scales and global scales in the FlashInfer CUTLASS NVFP4 layout. - flashinfer_nvfp4, - }; -}; - -pub const ActivationKind = enum { - /// Gelu activation function - gelu, - /// ReLU activation function - relu, - /// SiLU activation function - silu, - /// SwiGLU activation function (including clamped/clipped/scaled/biased variants) - swiglu, - /// SwiGLU activation function with step function - swiglu_step, - /// GeGlu activation function - geglu, - /// GeGlu activation function with tanh - geglu_tanh, -}; - -pub const Activation = union(ActivationKind) { - gelu: void, - relu: void, - silu: void, - swiglu: struct { - limit: ?f32 = null, - scale: ?f32 = null, - bias: ?f32 = null, - }, - swiglu_step: struct { - limit: ?f32 = null, - }, - geglu: void, - geglu_tanh: void, -}; - pub const Backend = enum { stablehlo, cute, @@ -85,8 +38,6 @@ pub const Backend = enum { tcl, pub fn auto(platform: *const zml.Platform, scheme: ?zml.Quantization.Scheme, dtype: zml.DataType) !Backend { - // Keep the dtype as an argument because non scheme-specific backends may depend on it later - _ = dtype; return switch (platform.target) { .cuda => b: { const s = scheme orelse break :b .triton; @@ -97,7 +48,7 @@ pub const Backend = enum { .triton, .nvfp4 => if (cute.isNvfp4Supported(platform)) .cute - else if (cutlass_flashinfer.isNvfp4Supported(platform)) + else if (flashinfer_cutlass.isNvfp4Supported(platform)) .flashinfer_cutlass else error.UnsupportedQuantization, @@ -120,7 +71,13 @@ pub const Backend = enum { }; }, .tpu => if (scheme == null) .mosaic_tpu else error.UnsupportedQuantization, - .furiosa => if (scheme == null) .tcl else error.UnsupportedQuantization, + .furiosa => b: { + const s = scheme orelse break :b .tcl; + break :b switch (s) { + .fp8_block128 => if (dtype == .f8e4m3fn) .tcl else error.UnsupportedQuantization, + .mxfp4, .nvfp4, .mxfp8, .fp8_per_channel, .fp8_per_tensor, .fp8_block32 => error.UnsupportedQuantization, + }; + }, .metal => b: { const s = scheme orelse break :b .metal; break :b switch (s) { @@ -136,7 +93,7 @@ pub const Backend = enum { return switch (backend) { .stablehlo => true, .cute => cute.isAvailable(platform), - .flashinfer_cutlass => cutlass_flashinfer.isAvailable(platform), + .flashinfer_cutlass => flashinfer_cutlass.isAvailable(platform), .fly => switch (platform.target) { .rocm => zml.platform.rocm.computeCapability(platform) == .gfx942, else => false, @@ -155,13 +112,17 @@ pub const Backend = enum { pub fn expertsLayout(backend: Backend, scheme: ?zml.Quantization.Scheme) !ExpertsLayout { return switch (backend) { .stablehlo => .{ .gate_up = .concatenated, .packing = .plain }, - .cute => if (scheme == .mxfp4 or scheme == .nvfp4) .{ .gate_up = .interleaved, .packing = .swizzled_scales } else error.UnsupportedQuantization, + .cute => if (scheme) |s| switch (s) { + .mxfp4 => .{ .gate_up = .row_interleaved_8, .packing = .mxfp4_interleaved8 }, + .nvfp4 => .{ .gate_up = .row_interleaved_1, .packing = .swizzled_scales }, + else => error.UnsupportedQuantization, + } else error.UnsupportedQuantization, .flashinfer_cutlass => if (scheme) |s| switch (s) { .nvfp4 => .{ .gate_up = .concatenated, .packing = .flashinfer_nvfp4 }, else => error.UnsupportedQuantization, } else error.UnsupportedQuantization, .mosaic_tpu, .metal, .tcl => .{ .gate_up = .concatenated, .packing = .plain }, - .triton, .fly => if (scheme == .mxfp4) .{ .gate_up = .interleaved, .packing = .plain } else .{ .gate_up = .concatenated, .packing = .plain }, + .triton, .fly => if (scheme == .mxfp4) .{ .gate_up = .row_interleaved_1, .packing = .plain } else .{ .gate_up = .concatenated, .packing = .plain }, }; } @@ -169,7 +130,7 @@ pub const Backend = enum { return switch (backend) { .stablehlo => {}, .cute => {}, - .flashinfer_cutlass => cutlass_flashinfer.register(platform), + .flashinfer_cutlass => flashinfer_cutlass.register(platform), .triton, .fly => {}, .mosaic_tpu => {}, .metal => {}, @@ -178,14 +139,6 @@ pub const Backend = enum { } }; -pub const Options = struct { - activation: Activation, - /// Quantize activations for Triton FP8 GEMMs; false keeps BF16 activations. - quantize_input: bool, - /// Where routing weights are applied; FlashInfer, Mosaic, Metal and TCL require after_down. - routing_weight_placement: RoutingWeightPlacement, -}; - /// Check the shared token-wise contract for activations and routing tensors. pub fn validateRoutingInputs(input: zml.Tensor, ids: zml.Tensor, weights: zml.Tensor) void { stdx.debug.assert(input.rank() == 2 and ids.rank() == 2 and weights.rank() == 2, "MoE expects input {{.t, .d}} and routing {{.t, .topk}}, got {f}, {f}, {f}", .{ input.shape(), ids.shape(), weights.shape() }); @@ -196,6 +149,35 @@ pub fn validateRoutingInputs(input: zml.Tensor, ids: zml.Tensor, weights: zml.Te weights.shape().hasTag(.t) == 0 and weights.shape().hasTag(.topk) == 1, "MoE expects ordered tags {{.t, .d}} for input and {{.t, .topk}} for routing", .{}); } +pub const Partitioning = struct { + const Logical = zml.Sharding.Partitioning.Logical; + + in_specs: struct { + input: Logical, + ids: Logical, + weights: Logical, + gate_up: Logical, + down: Logical, + }, + out_specs: Logical, + + /// Shards the `.expert` dimension of both projections over `mesh_axis`. + /// Other dimensions, input tokens, routing IDs/weights and output are replicated. + pub fn expert(comptime mesh_axis: @EnumLiteral()) Partitioning { + const experts: Logical = .parse(.{ .expert = mesh_axis }); + return .{ + .in_specs = .{ + .input = .replicated, + .ids = .replicated, + .weights = .replicated, + .gate_up = experts, + .down = experts, + }, + .out_specs = .replicated, + }; + } +}; + /// Input has shape {.t, .d}; routing IDs and weights have shape {.t, .topk}. /// Returns {.t, .d}. Flatten batched input with merge(.{ .t = .{ .b, .s } }). pub fn forwardMoe( @@ -205,72 +187,112 @@ pub fn forwardMoe( gate_up: zml.nn.Linear, down: zml.nn.Linear, backend: Backend, + mesh: *const zml.Mesh, + partitioning: Partitioning, opts: Options, ) zml.Tensor { validateRoutingInputs(input, topk_ids, topk_weights); - return switch (backend) { - .stablehlo => stablehlo.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .cute => cute.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .flashinfer_cutlass => cutlass_flashinfer.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .triton => triton.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .fly => fly.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .mosaic_tpu => mosaic_tpu.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .metal => metal.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), - .tcl => tcl.fusedExperts(input, topk_ids, topk_weights, gate_up, down, opts), + + const shapes = .{ + .input = input.shape().withPartitioning(mesh, partitioning.in_specs.input), + .ids = topk_ids.shape().withPartitioning(mesh, partitioning.in_specs.ids), + .weights = topk_weights.shape().withPartitioning(mesh, partitioning.in_specs.weights), + .gate_up = gate_up.weight.shape().withPartitioning(mesh, partitioning.in_specs.gate_up), + .down = down.weight.shape().withPartitioning(mesh, partitioning.in_specs.down), }; -} + const output_shape = input.shape().withPartitioning(mesh, partitioning.out_specs); -pub fn unpackedWeight(linear: zml.nn.Linear) zml.Tensor { - const quantization = linear.quantization orelse return linear.weight; - return if (zml.nn.isPackedFp4(quantization.scheme, linear.weight.dtype())) - zml.nn.unpackFp4(linear.weight, linear.tag, linear.tag) - else - linear.weight; -} + const tokens = shapes.input.partition(.t); + const experts = shapes.gate_up.partition(.expert); + // Local kernels require complete feature vectors and top-k routing lists. + stdx.debug.assert(shapes.input.partition(.d) == .replicated and shapes.ids.partition(.topk) == .replicated and shapes.weights.partition(.topk) == .replicated, "MoE feature and top-k dimensions must be replicated", .{}); + // The expert-axis reduction must add contributions for the same token rows. + // Sharding tokens along that axis would mix contributions for different tokens. + stdx.debug.assert(shapes.ids.partition(.t) == tokens and shapes.weights.partition(.t) == tokens and output_shape.partition(.t) == tokens and output_shape.partition(.d) == .replicated, "MoE input, routing and output must have matching token partitioning", .{}); + stdx.debug.assert(shapes.down.partition(.expert) == experts and (!experts.isSharded() or experts != tokens), "MoE projections must agree on expert partitioning; tokens must be replicated along the expert axis", .{}); + for ([_]zml.Shape{ shapes.gate_up, shapes.down }) |shape| { + stdx.debug.assert(!shape.containsPartitionSpec(.open), "MoE kernel weight partitioning must be closed", .{}); + for (0..shape.rank()) |axis| { + stdx.debug.assert(axis == shape.axis(.expert) or shape.partition(axis) == .replicated, "MoE kernel weights can only be partitioned along experts", .{}); + } + } + const expert_axis = experts.meshAxis(); + const parallel = if (expert_axis) |axis| mesh.grid.axis_sizes[axis] > 1 else false; + if (backend == .metal) stdx.debug.assert(!parallel, "This MoE backend does not support expert parallelism", .{}); -pub fn applyActivation(input: zml.Tensor, activation: zml.moe.Activation, layout: ProjectionLayout) zml.Tensor { - const x = input.convert(.f32); - return switch (activation) { - .gelu => x.gelu(), - .relu => x.relu(), - .silu => x.silu(), - .swiglu, .swiglu_step, .geglu, .geglu_tanh => b: { - const mid = @divFloor(x.dim(.out), 2); - var gate, var up = switch (layout) { - .concatenated => .{ x.slice(.out, .{ .end = mid }), x.slice(.out, .{ .start = mid }) }, - .interleaved => .{ x.slice(.out, .{ .start = 0, .step = 2 }), x.slice(.out, .{ .start = 1, .step = 2 }) }, + const Call = struct { + input: zml.Tensor, + ids: zml.Tensor, + weights: zml.Tensor, + gate_up: zml.nn.Linear, + down: zml.nn.Linear, + backend: Backend, + opts: Options, + mesh: *const zml.Mesh, + expert_axis: ?u3, + global_experts: u32, + + fn forward(args: @This(), _: zml.Shape) zml.Tensor { + const routing: LocalRouting = .{ + .ids = args.ids, + .weights = args.weights, + .global_num_experts = args.global_experts, + .local_num_experts = @intCast(args.gate_up.weight.dim(.expert)), + .expert_start = if (args.expert_axis) |axis| zml.ops.axisIndex(args.mesh, args.mesh.logical.axes.get(axis)).convert(.i32).scale(args.gate_up.weight.dim(.expert)) else null, }; - - break :b switch (activation) { - .swiglu => |parameters| { - stdx.debug.assert(parameters.scale == null, "triton and fly moe backend don't support swiglu scale", .{}); - - const limit: ?zml.Tensor = if (parameters.limit) |limit| .scalar(limit, x.dtype()) else null; - - // Apply limit on gate and clamp up - gate = if (limit) |l| gate.minimum(l) else gate; - up = if (limit) |l| up.clamp(l.negate(), l) else up; - - // Apply bias - up = if (parameters.bias) |bias| up.addConstant(bias) else up; - - break :b gate.silu().mul(up); - }, - .swiglu_step => |parameters| { - gate = gate.silu(); - - const limit: ?zml.Tensor = if (parameters.limit) |limit| .scalar(limit, x.dtype()) else null; - gate = if (limit) |l| gate.minimum(l) else gate; - up = if (limit) |l| up.clamp(l.negate(), l) else up; - break :b gate.mul(up); - }, - .geglu_tanh => gate.gelu().mul(up), - .geglu => { - log.warn("The geglu activation function was requested but we only support the tanh approximation", .{}); - break :b gate.gelu().mul(up); - }, - else => unreachable, // already treated by the top-level switch + const output = switch (args.backend) { + .stablehlo => stablehlo.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .cute => cute.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .flashinfer_cutlass => flashinfer_cutlass.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .triton => triton.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .fly => fly.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .mosaic_tpu => mosaic_tpu.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .metal => metal.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), + .tcl => tcl.fusedExperts(args.input, routing, args.gate_up, args.down, args.opts), }; + return if (args.expert_axis) |axis| zml.ops.allReduceOnAxis(output, zml.Tensor.add, args.mesh, args.mesh.logical.axes.get(axis)) else output; + } + }; + const args: Call = .{ + .input = input, + .ids = topk_ids, + .weights = topk_weights, + .gate_up = gate_up, + .down = down, + .backend = backend, + .opts = opts, + .mesh = mesh, + .expert_axis = if (parallel) expert_axis else null, + .global_experts = @intCast(gate_up.weight.dim(.expert)), + }; + const Logical = zml.Sharding.Partitioning.Logical; + var projection_specs: [2]zml.ops.PartitionSpecs(zml.nn.Linear) = undefined; + for ([_]zml.nn.Linear{ gate_up, down }, &projection_specs) |linear, *specs| { + zml.meta.mapAlloc(struct { + fn spec(ctx: struct { experts: i64, axis: ?zml.Shape.Tag }, tensor: zml.Tensor) Logical { + const axis = ctx.axis orelse return .replicated; + if (tensor.count() == 1) return .replicated; + // Packed scales may be untagged; experts are their leading dimension. + const dim = tensor.shape().hasTag(.expert) orelse 0; + stdx.debug.assert(tensor.dim(dim) == ctx.experts, "MoE tensor must contain {} experts, got {f}", .{ ctx.experts, tensor.shape() }); + return .{ .mapping = .{ + .src = .init(&.{if (tensor.shape().hasTag(.expert) != null) zml.Shape.toTag(.expert) else zml.Shape.toTag(._0)}), + .tgt = .init(&.{axis}), + } }; + } + }.spec, zml.Compiler.current().arena.allocator(), .{ + .experts = linear.weight.dim(.expert), + .axis = if (expert_axis) |axis| mesh.logical.axes.get(axis) else null, + }, linear, specs) catch @panic("OOM"); + } + return zml.ops.manualComputation(Call.forward, args, output_shape, mesh, .{ + .in_specs = .{ + .input = partitioning.in_specs.input, + .ids = partitioning.in_specs.ids, + .weights = partitioning.in_specs.weights, + .gate_up = projection_specs[0], + .down = projection_specs[1], }, - }; + .out_specs = partitioning.out_specs, + }); } diff --git a/zml/kernels/moe/mosaic_tpu/mosaic_tpu.zig b/zml/kernels/moe/mosaic_tpu/mosaic_tpu.zig index 09dc360a..bb60594b 100644 --- a/zml/kernels/moe/mosaic_tpu/mosaic_tpu.zig +++ b/zml/kernels/moe/mosaic_tpu/mosaic_tpu.zig @@ -5,6 +5,9 @@ const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; +const Activation = @import("../api.zig").Activation; +const GlobalOptions = @import("../api.zig").Options; +const LocalRouting = @import("../api.zig").LocalRouting; pub const gmm_ep = @import("gmm_ep.zig"); const log = std.log.scoped(.@"zml/moe/mosaic_tpu"); @@ -12,22 +15,9 @@ const log = std.log.scoped(.@"zml/moe/mosaic_tpu"); var gate_up_transpose: std.atomic.Value(bool) = .init(false); var down_transpose: std.atomic.Value(bool) = .init(false); -pub const Options = struct { - activation: zml.moe.Activation, - global_num_experts: i64 = -1, - expert_map: ?Tensor = null, - w1_scale: ?Tensor = null, - w2_scale: ?Tensor = null, - w1_bias: ?Tensor = null, - w2_bias: ?Tensor = null, -}; - -fn validateOptions(opts: zml.moe.Options) void { +fn validateOptions(opts: GlobalOptions) void { stdx.debug.assert(!opts.quantize_input, "Optional FP8 input quantization requires the Triton MoE backend", .{}); stdx.debug.assert(opts.routing_weight_placement == .after_down, "Non-Triton MoE backends require routing weights after the down projection", .{}); - // TODO(Corentin): find where to put that - //if (opts.expert_map != null and opts.global_num_experts == -1) return error.InvalidShape; - //if (opts.w1_scale != null or opts.w2_scale != null) return error.UnsupportedQuantization; } fn validateInputs(hidden: Tensor, gate_up: Tensor, down: Tensor, weights: Tensor, ids: Tensor) !void { @@ -224,120 +214,25 @@ pub fn callGmmEp( } pub fn fusedExperts( - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - opts: zml.moe.Options, + hidden_states: zml.Tensor, + routing: LocalRouting, + gate_up_: zml.nn.Linear, + down_: zml.nn.Linear, + opts: GlobalOptions, ) zml.Tensor { validateOptions(opts); + const expert_map_ = routing.expertMap(); - const gate_up_scales: ?zml.Tensor = if (gate_up.quantization) |q| q.scales else null; - const down_scales: ?zml.Tensor = if (down.quantization) |q| q.scales else null; - - const expert_partition = gate_up.weight.shape().partition(.expert); - - if (expert_partition.isSharded()) { - const global_num_experts = down.weight.dim(.expert); - const partial_output = zml.ops.manualComputation( - (struct { - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - weights_gate_up: zml.Tensor, - weights_down: zml.Tensor, - activation: zml.moe.Activation, - global_num_experts: i64, - gate_up_scales: ?zml.Tensor, - bias_gate_up: ?zml.Tensor, - down_scales: ?zml.Tensor, - bias_down: ?zml.Tensor, - - fn body(self: @This(), _: zml.Shape) zml.Tensor { - const local_num_experts = self.weights_gate_up.dim(.expert); - const partition_id = zml.ops.partitionId().convert(.i32); - const expert_start = partition_id.scale(local_num_experts).convert(.i32); - const global_expert_ids = zml.Tensor.arange(.{ .end = self.global_num_experts }, .i32).withTags(.{.expert}); - - const local_expert_mask = global_expert_ids.cmp(.GE, expert_start) - .logical(.AND, global_expert_ids.cmp(.LT, expert_start.addConstant(local_num_experts))); - const expert_map = local_expert_mask.select( - global_expert_ids.sub(expert_start), - zml.Tensor.scalar(-1, .i32), - ); - const local_output = fusedExpertsImpl( - self.input, - self.weights_gate_up, - self.weights_down, - self.topk_weights, - self.topk_ids, - .{ - .activation = self.activation, - .global_num_experts = self.global_num_experts, - .expert_map = expert_map, - .w1_scale = self.gate_up_scales, - .w2_scale = self.down_scales, - .w1_bias = self.bias_gate_up, - .w2_bias = self.bias_down, - }, - ); - return local_output.reshape(self.input.shape()).withTags(.{ .t, .d }); - } - }).body, - .{ - .input = input, - .topk_ids = topk_ids, - .topk_weights = topk_weights, - .weights_gate_up = gate_up.weight, - .weights_down = down.weight, - .activation = opts.activation, - .global_num_experts = global_num_experts, - .gate_up_scales = gate_up_scales, - .bias_gate_up = gate_up.bias, - .down_scales = down_scales, - .bias_down = down.bias, - }, - input.shape(), - ); - return zml.ops.allReduce(partial_output, zml.Tensor.add); - } - - return fusedExpertsImpl( - input, - gate_up.weight, - down.weight, - topk_weights, - topk_ids, - .{ - .activation = opts.activation, - .global_num_experts = gate_up.weight.dim(.expert), - .w1_scale = gate_up_scales, - .w2_scale = down_scales, - .w1_bias = gate_up.bias, - .w2_bias = down.bias, - }, - ); -} - -pub fn fusedExpertsImpl( - hidden_states: Tensor, - w1: Tensor, - w2: Tensor, - topk_weights: Tensor, - topk_ids: Tensor, - opts: Options, -) Tensor { const tokens = hidden_states.dim(.t); const hidden = hidden_states.reshape(.{ .token = tokens, .in = hidden_states.dim(.d) }); // TODO(Corentin): Better error message - const gate_up = canonicalizeGateUp(w1, hidden.dim(.in)) catch |e| stdx.debug.panic("Invalid gate_up weights for fusedExpertsImpl: {}", .{e}); - const down = canonicalizeDown(w2, hidden.dim(.in), gate_up.dim(.out)) catch |e| stdx.debug.panic("Invalid down weights for fusedExpertsImpl: {}", .{e}); - const weights = topk_weights.reshape(.{ .token = tokens, .topk = topk_weights.dim(.topk) }); - const ids = topk_ids.reshape(.{ .token = tokens, .topk = topk_ids.dim(.topk) }); + const gate_up = canonicalizeGateUp(gate_up_.weight, hidden.dim(.in)) catch |e| stdx.debug.panic("Invalid gate_up weights for fusedExperts: {}", .{e}); + const down = canonicalizeDown(down_.weight, hidden.dim(.in), gate_up.dim(.out)) catch |e| stdx.debug.panic("Invalid down weights for fusedExperts: {}", .{e}); + const weights = routing.weights.reshape(.{ .token = tokens, .topk = routing.weights.dim(.topk) }); + const ids = routing.ids.reshape(.{ .token = tokens, .topk = routing.ids.dim(.topk) }); - const num_experts = if (opts.global_num_experts != -1) opts.global_num_experts else gate_up.dim(.expert); - if (opts.expert_map) |expert_map| { + const num_experts = routing.global_num_experts; + if (expert_map_) |expert_map| { if (expert_map.dtype() != .i32) stdx.debug.panic("expert_map must be of dtype i32, got {}", .{expert_map.dtype()}); if (expert_map.rank() != 1 or expert_map.dim(.expert) != num_experts) stdx.debug.panic("expert_map must be a 1D tensor of length num_experts, got shape {}", .{expert_map.shape()}); } @@ -345,7 +240,7 @@ pub fn fusedExpertsImpl( const flat_ids_global = ids.transpose(.{ .topk, .token }).flatten().withTags(.{.route}); const flat_weights = weights.transpose(.{ .topk, .token }).flatten().withTags(.{.route}); - const flat_ids_local, const local_route_mask = if (opts.expert_map) |expert_map| blk: { + const flat_ids_local, const local_route_mask = if (expert_map_) |expert_map| blk: { const mapped = expert_map.gather(.{ .expert = flat_ids_global }, .{}).withTags(.{.route}); const route_mask = mapped.cmp(.GE, Tensor.scalar(0, .i32).broad(mapped.shape())); const safe_ids = route_mask.select(mapped, Tensor.zeroes(mapped.shape())); @@ -356,7 +251,7 @@ pub fn fusedExpertsImpl( }; const sorted_route = flat_ids_local.argsort(.route, .{ .descending = false }).rename(.{ .route = .perm }); - const hidden_sorted = if (opts.expert_map != null) + const hidden_sorted = if (expert_map_ != null) permuteSmallBatch(hidden, sorted_route, ids.dim(.topk)) else blk: { // If not sharded, we can save the sort + gather by just repeating the hidden states for each expert and gathering with the sorted route @@ -379,19 +274,19 @@ pub fn fusedExpertsImpl( .gather(.{ .token = aligned_hidden.positions.rename(.{ .token = .route }) }, .{}) .rename(.{ .route = .token }); - if (opts.w1_bias) |bias| { + if (gate_up_.bias) |bias| { const bias_per_token = bias.withTags(.{ .expert, .out }).gather(.{ .expert = expert_ids_sorted }, .{}); gate_up_out = gate_up_out.add(bias_per_token); } - const activated = zml.moe.applyActivation(gate_up_out, opts.activation, .concatenated); + const activated = opts.activation.applyActivation(gate_up_out, .concatenated); const aligned_activated = alignSortedRowsByGroup(activated, expert_ids_sorted, group_sizes, tile_m); var down_out = callGmmEp(aligned_activated.rows, down, aligned_activated.group_sizes, hidden.dtype(), .none) .gather(.{ .token = aligned_activated.positions.rename(.{ .token = .route }) }, .{}) .rename(.{ .route = .token }); - if (opts.w2_bias) |bias| { + if (down_.bias) |bias| { const bias_per_token = bias.withTags(.{ .expert, .out }).gather(.{ .expert = expert_ids_sorted }, .{}); down_out = down_out.add(bias_per_token); } diff --git a/zml/kernels/moe/stablehlo/stablehlo.zig b/zml/kernels/moe/stablehlo/stablehlo.zig index e33d2a14..ba723234 100644 --- a/zml/kernels/moe/stablehlo/stablehlo.zig +++ b/zml/kernels/moe/stablehlo/stablehlo.zig @@ -1,13 +1,18 @@ //! Deliberately simple MoE reference: no routed GEMMs, packing, or custom calls. //! Weights use [expert, output, input] order with concatenated gate/up columns. const std = @import("std"); + const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; const Shape = zml.Shape; +const GlobalOptions = @import("../api.zig").Options; +const Activation = @import("../api.zig").Activation; +const RoutingWeightPlacement = @import("../api.zig").RoutingWeightPlacement; +const LocalRouting = @import("../api.zig").LocalRouting; -pub fn activate(x: Tensor, activation: zml.moe.Activation) Tensor { +pub fn activate(x: Tensor, activation: Activation) Tensor { return switch (activation) { .silu => x.silu(), .gelu => x.gelu(), @@ -148,23 +153,28 @@ fn project(x: Tensor, linear: zml.nn.Linear, ids: Tensor, opts: Options, dtype: } pub const Options = struct { - activation: zml.moe.Activation, + activation: Activation, quantize_input: bool = false, /// Model fixed backend contracts independently of optional input quantization: /// Metal uses weight-only quantization; specialized MXFP4 GEMMs always use MXFP8 inputs. /// CuTe additionally rounds the intermediate activation to BF16 before quantizing. input_quantization: enum { automatic, none, mxfp8, mxfp8_bf16 } = .automatic, - routing_weight_placement: zml.moe.RoutingWeightPlacement = .after_down, + routing_weight_placement: RoutingWeightPlacement = .after_down, }; +/// Negative expert IDs denote nonlocal routes and contribute zero, including bias. pub fn reference(input: Tensor, topk_ids: Tensor, topk_weights: Tensor, gate_up: zml.nn.Linear, down: zml.nn.Linear, opts: Options) Tensor { + stdx.debug.assert(topk_ids.dim(.topk) > 0, "reference top-k must be positive", .{}); stdx.debug.assert(topk_ids.dtype() == .i32, "reference expects i32 expert ids", .{}); stdx.debug.assert(gate_up.weight.rank() == 3 and down.weight.rank() == 3, "reference expects rank-three expert weights", .{}); stdx.debug.assert(gate_up.weight.dim(.expert) == down.weight.dim(.expert), "reference expert counts must match", .{}); const tokens = input.dim(.t); const k = topk_ids.dim(.topk); - stdx.debug.assert(k > 0 and k <= gate_up.weight.dim(.expert), "reference top-k must fit the expert count", .{}); - const ids = topk_ids.reshape(.{ .route = tokens * k }); + // Nonlocal routes use expert zero for safe gathers; discard their full + // contribution after the down projection, including any projection bias. + const local_routes = topk_ids.cmp(.GE, Tensor.scalar(0, .i32)); + const safe_ids = topk_ids.maximum(Tensor.scalar(0, .i32)); + const ids = safe_ids.reshape(.{ .route = tokens * k }); const weights = topk_weights.reshape(.{ .route = tokens * k }).convert(.f32); const hidden = input.reshape(.{ .token = tokens, .in = input.dim(.d) }) .insertAxes(.in, .{.topk}).broad(Shape.init(.{ .token = tokens, .topk = k, .in = input.dim(.d) }, input.dtype())) @@ -176,13 +186,15 @@ pub fn reference(input: Tensor, topk_ids: Tensor, topk_weights: Tensor, gate_up: if (opts.routing_weight_placement == .before_down) activated = activated.mul(weights.broad(activated.shape())); var output = project(activated.rename(.{ .out = .in }), down, ids, opts, input.dtype()); if (opts.routing_weight_placement == .after_down) output = output.mul(weights.broad(output.shape())); + output = local_routes.reshape(.{ .route = tokens * k }).broad(output.shape()).select(output, Tensor.scalar(0, output.dtype())); return output.convert(input.dtype()).convert(.f32) .reshape(.{ .t = tokens, .topk = k, .d = output.dim(.out) }) .sum(.topk).squeeze(.topk).convert(input.dtype()); } -pub fn fusedExperts(input: Tensor, ids: Tensor, weights: Tensor, gate_up: zml.nn.Linear, down: zml.nn.Linear, opts: zml.moe.Options) Tensor { - return reference(input, ids, weights, gate_up, down, .{ +pub fn fusedExperts(input: Tensor, routing: LocalRouting, gate_up: zml.nn.Linear, down: zml.nn.Linear, opts: GlobalOptions) Tensor { + stdx.debug.assert(routing.ids.dim(.topk) > 0 and routing.ids.dim(.topk) <= routing.global_num_experts, "MoE top-k must fit the global expert count", .{}); + return reference(input, routing.localIds(), routing.weights, gate_up, down, .{ .activation = opts.activation, .quantize_input = opts.quantize_input, .routing_weight_placement = opts.routing_weight_placement, diff --git a/zml/kernels/moe/tcl/BUILD.bazel b/zml/kernels/moe/tcl/BUILD.bazel index 4057eea3..44b56e08 100644 --- a/zml/kernels/moe/tcl/BUILD.bazel +++ b/zml/kernels/moe/tcl/BUILD.bazel @@ -2,5 +2,8 @@ package(default_visibility = ["//zml:__pkg__"]) filegroup( name = "tcl", - srcs = ["tcl.zig"], + srcs = [ + "moe.zig", + "tcl.zig", + ], ) diff --git a/zml/kernels/moe/tcl/moe.zig b/zml/kernels/moe/tcl/moe.zig new file mode 100644 index 00000000..23fb2274 --- /dev/null +++ b/zml/kernels/moe/tcl/moe.zig @@ -0,0 +1,457 @@ +//! TCL kernel schedules for Furiosa RNGD fused experts. +const std = @import("std"); + +const builder = @import("dsl/tcl/builder"); + +const zml = @import("../../../zml.zig"); +const tcl = zml.kernel.tcl; +const ActivationKind = @import("../api.zig").ActivationKind; + +pub const Format = enum { bf16, fp8_block128 }; + +/// Dispatch in `tcl.zig` uses the first matching schedule. +/// +/// - `.small_batch` for FP8 up to 64 tokens before padding. Run each selected +/// expert on the whole batch to reuse its weights, with zero routing weight +/// for unrouted pairs. +/// - `.per_route` when `num_tokens * top_k <= local_num_experts`. Gather +/// weights for each route to avoid the overhead of grouping routes. +/// - `.blockwise` otherwise. Pack routed tokens into blocks by expert to +/// reuse weights within each block, then gather results back by route. +pub const Schedule = enum { per_route, blockwise, small_batch }; + +pub const Cfg = struct { + num_tokens: i64, + top_k: i64, + local_num_experts: i64, + hidden_size: i64, + expert_intermediate_size: i64, + weight_scale_dtype: tcl.DType = .bf16, + activation: ActivationKind = .swiglu, +}; + +const base_inputs: []const [:0]const u8 = &.{ "x", "ids", "w", "gate_up", "down" }; +const scale_inputs: []const [:0]const u8 = &.{ "gate_up_scale", "down_scale" }; + +pub fn Kernel(comptime format: Format, comptime schedule: Schedule) type { + if (schedule == .small_batch and format != .fp8_block128) { + @compileError("TCL MoE FP8 small-batch schedule requires block-128 FP8"); + } + return tcl.Kernel(Cfg, .{ + .name = "moe_" ++ @tagName(schedule) ++ switch (format) { + .bf16 => "", + .fp8_block128 => "_fp8", + }, + .inputs = switch (format) { + .bf16 => base_inputs, + .fp8_block128 => base_inputs ++ scale_inputs, + }, + .outputs = &.{"y"}, + .run = struct { + fn run(b: *tcl.Builder, cfg: Cfg) tcl.FinishError!void { + configure(b, schedule); + const a: Axes = .init(b, cfg); + const in = try args(b, a, cfg, format); + switch (schedule) { + .per_route => try perRoute(b, a, in, cfg.activation, format), + .blockwise => try blockwise(b, a, in, cfg, format), + .small_batch => try fp8SmallBatch(b, a, in, cfg), + } + } + }.run, + }); +} + +const Axes = struct { + T: tcl.Axis, + K: tcl.Axis, + E: tcl.Axis, + D: tcl.Axis, + I: tcl.Axis, + I2: tcl.Axis, + S2: tcl.Axis, + S1: tcl.Axis, + + fn init(b: *tcl.Builder, cfg: Cfg) Axes { + return .{ + .T = b.axis("T", cfg.num_tokens), + .K = b.axis("K", cfg.top_k), + .E = b.axis("E", cfg.local_num_experts), + .D = b.axis("D", cfg.hidden_size), + .I = b.axis("I", cfg.expert_intermediate_size), + .I2 = b.axis("I2", 2 * cfg.expert_intermediate_size), + .S2 = b.axis("S2", 2), + .S1 = b.axis("S1", 1), + }; + } +}; + +fn args(b: *tcl.Builder, a: Axes, cfg: Cfg, comptime format: Format) tcl.FinishError!Args { + const weight_dtype: tcl.DType = switch (format) { + .bf16 => .bf16, + .fp8_block128 => .f8_e4, + }; + const spec = .{ + .x = .{ .dtype = .bf16, .axes = &.{ a.T, a.D } }, + .ids = .{ .dtype = .i32, .axes = &.{ a.T, a.K } }, + .w = .{ .dtype = .f32, .axes = &.{ a.T, a.K } }, + .gate_up = .{ .dtype = weight_dtype, .axes = &.{ a.E, a.I2, a.D } }, + .down = .{ .dtype = weight_dtype, .axes = &.{ a.E, a.D, a.I } }, + }; + if (format == .bf16) { + const t = try b.declareArgs(spec); + return .{ .x = t.x, .ids = t.ids, .w = t.w, .gate_up = t.gate_up, .down = t.down }; + } + const IG = b.axis("IG", @divExact(cfg.expert_intermediate_size, 128)); + const DG = b.axis("DG", @divExact(cfg.hidden_size, 128)); + const I2G = b.axis("I2G", 2 * @divExact(cfg.expert_intermediate_size, 128)); + const t = try b.declareArgs(.{ + .x = spec.x, + .ids = spec.ids, + .w = spec.w, + .gate_up = spec.gate_up, + .down = spec.down, + .gate_up_scale = .{ .dtype = cfg.weight_scale_dtype, .axes = &.{ a.E, I2G, DG } }, + .down_scale = .{ .dtype = cfg.weight_scale_dtype, .axes = &.{ a.E, DG, IG } }, + }); + return .{ .x = t.x, .ids = t.ids, .w = t.w, .gate_up = t.gate_up, .down = t.down, .gate_up_scale = t.gate_up_scale, .down_scale = t.down_scale }; +} + +const Args = struct { + x: builder.Tensor, + ids: builder.Tensor, + w: builder.Tensor, + gate_up: builder.Tensor, + down: builder.Tensor, + gate_up_scale: ?builder.Tensor = null, + down_scale: ?builder.Tensor = null, +}; + +const GatheredScales = struct { gate_up: ?builder.Tensor = null, down: ?builder.Tensor = null }; + +fn gatherScales(b: *tcl.Builder, a: Axes, in: Args, ids: builder.Tensor, batch: []const tcl.Axis, comptime format: Format) tcl.FinishError!GatheredScales { + if (format == .bf16) return .{}; + const IG = b.axis("IG", @divExact(a.I.size, 128)); + const DG = b.axis("DG", @divExact(a.D.size, 128)); + const gate_up = try b.reshape(in.gate_up_scale.?, &.{ a.E, a.S2, IG, DG }, .{}); + var gu_axes: [8]tcl.Axis = undefined; + var dn_axes: [8]tcl.Axis = undefined; + @memcpy(gu_axes[0..batch.len], batch); + @memcpy(dn_axes[0..batch.len], batch); + @memcpy(gu_axes[batch.len..][0..3], &[_]tcl.Axis{ a.S2, IG, DG }); + @memcpy(dn_axes[batch.len..][0..2], &[_]tcl.Axis{ DG, IG }); + return .{ + .gate_up = try b.gather(gate_up, ids, a.E, gu_axes[0 .. batch.len + 3], .{}), + .down = try b.gather(in.down_scale.?, ids, a.E, dn_axes[0 .. batch.len + 2], .{}), + }; +} + +fn blockedAxes(axes: []const tcl.Axis, axis: tcl.Axis, group: tcl.Axis, lane: tcl.Axis, storage: *[16]tcl.Axis) []const tcl.Axis { + var n: usize = 0; + for (axes) |a| { + if (std.mem.eql(u8, a.name, axis.name)) { + storage[n] = group; + n += 1; + storage[n] = lane; + } else storage[n] = a; + n += 1; + } + return storage[0..n]; +} + +/// Block scales multiply each K-block partial before K-block reduction. +/// Only the requested experts are gathered; their bank remains FP8 in HBM. +fn project(b: *tcl.Builder, x: builder.Tensor, weight: builder.Tensor, scale: ?builder.Tensor, contracting: tcl.Axis, output: tcl.Axis, to: []const tcl.Axis, dtype: ?tcl.DType, comptime format: Format) tcl.FinishError!builder.Tensor { + if (format == .bf16) { + const op = b.tensorOperation(.{}); + return op.commit(op.contract(x, weight, to), .{ .dtype = dtype }); + } + const CG = b.axis(if (std.mem.eql(u8, contracting.name, "D")) "DG" else "IG", @divExact(contracting.size, 128)); + const CL = b.axis(if (std.mem.eql(u8, contracting.name, "D")) "DB" else "IB", 128); + const OG = b.axis(if (std.mem.eql(u8, output.name, "D")) "DG" else "IG", @divExact(output.size, 128)); + const OL = b.axis(if (std.mem.eql(u8, output.name, "D")) "DB" else "IB", 128); + var x_axes: [16]tcl.Axis = undefined; + var w_axes: [16]tcl.Axis = undefined; + var w2_axes: [16]tcl.Axis = undefined; + var out_axes: [16]tcl.Axis = undefined; + var partial_axes: [16]tcl.Axis = undefined; + const xb = try b.reshape(x, blockedAxes(x.axes, contracting, CG, CL, &x_axes), .{}); + const wb = try b.reshape(weight, blockedAxes(blockedAxes(weight.axes, output, OG, OL, &w_axes), contracting, CG, CL, &w2_axes), .{}); + const out = blockedAxes(to, output, OG, OL, &out_axes); + @memcpy(partial_axes[0..out.len], out); + partial_axes[out.len] = CG; + + // Furiosa's published FP8 recipe uses an additive epsilon, including zero blocks. + var op = b.tensorOperation(.{}); + const qscale = try op.commit(op.fetch(xb, .{}).unary(.absf).reduce(&.{CL}, .maxf).divf(448.0).addf(1e-12), .{ .dtype = .f32 }); + op = b.tensorOperation(.{}); + // Native F32 reciprocal accuracy is qualified separately (within one ULP). + const reciprocal = op.ve(.divf, .{ 1.0, op.fetch(qscale, .{}) }); + const quantized = try op.commit(reciprocal.mulf(op.fetch(xb, .{})), .{ .dtype = .f8_e4 }); + op = b.tensorOperation(.{}); + const scaled = op.contract(quantized, wb, partial_axes[0 .. out.len + 1]).mulf(op.fetch(qscale, .{})).mulf(op.fetch(scale.?, .{})).reduce(&.{CG}, .addf); + return b.reshape(try op.commit(scaled, .{ .dtype = dtype, .axes = out }), to, .{}); +} + +/// `silu(gate) * up` or `gelu_tanh(gate) * up` of `h = [batch..., S2, I]` +/// (gate first), times `weights` (`[batch...]`) unless null. +fn gated(b: *tcl.Builder, a: Axes, h: builder.Tensor, batch: []const tcl.Axis, weights: ?builder.Tensor, activation: ActivationKind, comptime format: Format) tcl.FinishError!builder.Tensor { + var sliced: [4]tcl.Axis = undefined; + var flat: [3]tcl.Axis = undefined; + @memcpy(sliced[0..batch.len], batch); + @memcpy(flat[0..batch.len], batch); + sliced[batch.len] = a.S1; + sliced[batch.len + 1] = a.I; + flat[batch.len] = a.I; + const s = sliced[0 .. batch.len + 2]; + const f = flat[0 .. batch.len + 1]; + const gate = try b.reshape(try b.slice(h, a.S2, 0, s, .{}), f, .{}); + const up = try b.reshape(try b.slice(h, a.S2, 1, s, .{}), f, .{}); + // The builder can infer the TCL `Interleaving` tactic for two inputs with + // matching axes, such as `gate` and `up`. That tactic allows exactly one + // binary instruction, so force `.elementwise` for this activation chain. + var op = b.tensorOperation(.{ .tactic = .elementwise }); + const g = op.fetch(gate, .{}); + const activated = if (format == .fp8_block128) blk: { + if (activation != .swiglu) std.debug.panic("TCL MoE block-128 FP8 requires SwiGLU, got {s}", .{@tagName(activation)}); + + // WORKAROUND (TCC 2026.5.0b0): native fusion elides an internal + // BF16 SiLU write, even through as_dram/as_logical. Make its RNE + // rounding observable as integer bits before multiplying by up. + // Remove once the native reproducer preserves the BF16 boundary. + const bits = try op.commit(g.sigmoid().mulf(g).unary(.reinterpret_as_int), .{ .dtype = .i32 }); + op = b.tensorOperation(.{ .tactic = .elementwise }); + const v = op.fetch(bits, .{}); + const rounded = v.binary(.shr_logical, 16).binary(.bit_andi, 1).addi(32767).addi(v).binary(.bit_andi, -65536).unary(.reinterpret_as_float); + const silu = try op.commit(rounded, .{ .dtype = .bf16 }); + op = b.tensorOperation(.{ .tactic = .elementwise }); + break :blk op.fetch(silu, .{}); + } else switch (activation) { + .swiglu => g.sigmoid().mulf(g), + .geglu_tanh => geluTanh(g), + else => std.debug.panic("TCL MoE does not support activation {s}", .{@tagName(activation)}), + }; + var v = activated.mulf(op.fetch(up, .{})); + if (weights) |w| v = v.mulf(op.fetch(w, .{})); + return op.commit(v, .{ .dtype = .bf16 }); +} + +fn geluTanh(g: builder.Value) builder.Value { + const cube = g.mulf(g).mulf(g); + const inner = g.addf(cube.mulf(0.044715)).mulf(@sqrt(2.0 / std.math.pi)); + return inner.tanh().addf(1.0).mulf(g).mulf(0.5); +} + +/// Furiosa's compiler configuration for MoE kernels (`qwen3_moe/config.py`): +/// Use IO-bound tactics for per-route and FP8 small-batch schedules and +/// compute-bound tactics for blockwise. +/// Keep the compiler's default lowering mode and omit the adaptive-einsum and +/// half-VRF settings removed in TCC 2026.5. +fn configure(b: *tcl.Builder, comptime schedule: Schedule) void { + const config = comptime blk: { + const common = .{ + .enable_einsum_fusion = true, + .padding_policy = .{ .Small = 1.1 }, + .instruction_mem_budget = 0xB0000, + .reshape_einsum_mode = .{ .Reshape = .{ .permute = true } }, + .tensor_unit_bridge_threshold_in_page = 12, + .allow_external_operators = false, + .enable_tactic_pruning = false, + .scheduler_beam_search = true, + .allow_reduce_by_ve_cluster_chip_reduce = false, + .allow_reduce_by_ve_cluster_chip_reduce_base_population = false, + .propagate_sparse_axis_from_op = "None", + }; + const tactics = switch (schedule) { + .blockwise => .{ + .tactic_hint = .{ .ForLlmModelComputeBound = 200 }, + .dma_preference = 1.2, + }, + .per_route, .small_batch => .{ + .tactic_hint = .{ .ForLlmModelIOBound = 200 }, + .num_transaction_simulation_per_pe = 1024, + .dma_preference = 0.8, + }, + }; + // The transaction-simulation setting is only emitted for IO-bound schedules. + const common_info = @typeInfo(@TypeOf(common)).@"struct"; + const tactics_info = @typeInfo(@TypeOf(tactics)).@"struct"; + const Config = @Struct( + .auto, + null, + common_info.field_names ++ tactics_info.field_names, + common_info.field_types ++ tactics_info.field_types, + common_info.field_attrs ++ tactics_info.field_attrs, + ); + var merged: Config = undefined; + for (common_info.field_names) |field| @field(merged, field) = @field(common, field); + for (tactics_info.field_names) |field| @field(merged, field) = @field(tactics, field); + break :blk merged; + }; + b.compilerConfig(config); +} + +/// Gather the selected experts for every token. +fn perRoute(b: *tcl.Builder, a: Axes, in: Args, activation: ActivationKind, comptime format: Format) tcl.FinishError!void { + const gate_up = try b.reshape(in.gate_up, &.{ a.E, a.S2, a.I, a.D }, .{}); + const gu = try b.gather(gate_up, in.ids, a.E, &.{ a.T, a.K, a.S2, a.I, a.D }, .{}); + const dn = try b.gather(in.down, in.ids, a.E, &.{ a.T, a.K, a.D, a.I }, .{}); + + const scales = try gatherScales(b, a, in, in.ids, &.{ a.T, a.K }, format); + const h = try project(b, in.x, gu, scales.gate_up, a.D, a.I, &.{ a.T, a.K, a.S2, a.I }, if (format == .fp8_block128) .bf16 else null, format); + const act = try gated(b, a, h, &.{ a.T, a.K }, if (format == .fp8_block128) null else in.w, activation, format); + if (format == .fp8_block128) { + const y = try project(b, act, dn, scales.down, a.I, a.D, &.{ a.T, a.K, a.D }, .f32, format); + const op = b.tensorOperation(.{}); + b.ret(&.{try op.commit(op.fetch(y, .{}).mulf(op.fetch(in.w, .{})).reduce(&.{a.K}, .addf), .{ .dtype = .bf16 })}); + } else { + // The BF16 recipe weights activations and contracts experts together. + const op = b.tensorOperation(.{}); + b.ret(&.{try op.commit(op.contract(act, dn, &.{ a.T, a.D }), .{ .dtype = .bf16 })}); + } +} + +/// Rows per expert block: the next power of two of the routes per expert. +fn blockRows(cfg: Cfg) i64 { + const per_expert = std.math.divCeil(i64, cfg.num_tokens * cfg.top_k, cfg.local_num_experts) catch unreachable; + return @intCast(std.math.ceilPowerOfTwo(u64, @intCast(@max(per_expert, 4))) catch unreachable); +} + +const blocks_per_chunk = 32; + +/// Group routes into expert-major blocks, then run both projections in a +/// device loop over chunks of blocks. +fn blockwise(b: *tcl.Builder, a: Axes, in: Args, cfg: Cfg, comptime format: Format) tcl.FinishError!void { + const bb = blockRows(cfg); + // One block per expert at least, then enough for every route. + const max_blocks = std.math.divCeil(i64, @max(cfg.num_tokens * cfg.top_k - cfg.local_num_experts, 0), bb) catch unreachable; + const chunks = std.math.divCeil(i64, max_blocks + cfg.local_num_experts, blocks_per_chunk) catch unreachable; + const BB = b.axis("BB", bb); + const BGi = b.axis("BGi", blocks_per_chunk); + const Go = b.axis("Go", chunks); + const BG = b.axis("BG", chunks * blocks_per_chunk); + const GB = b.axis("GB", chunks * blocks_per_chunk * bb); + const T, const K, const E, const D, const I = .{ a.T, a.K, a.E, a.D, a.I }; + + // Routing: route (t, k) goes to row `slot` of block `start[e] + rank`. + const experts = try b.arange(cfg.local_num_experts, .i32, &.{E}, .{}); + var op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); + const dist = try op.commit(op.fetch(in.ids, .{}).subi(op.fetch(experts, .{})), .{}); + op = b.tensorOperation(.{}); + const onehot = try op.commit(op.where(dist, .eq, 0, 1, 0), .{}); + op = b.tensorOperation(.{}); + const rank = try op.commit(op.fetch(onehot, .{}).cumsum(&.{ T, K }), .{}); + op = b.tensorOperation(.{}); + const log2_bb: i64 = std.math.log2_int(u64, @intCast(bb)); + const blocks = try op.commit(op.fetch(onehot, .{}).reduce(&.{ T, K }, .addi).addi(bb - 1).binary(.shr_arith, log2_bb), .{}); + op = b.tensorOperation(.{}); + const block_end = try op.commit(op.fetch(blocks, .{}).cumsum(&.{E}), .{}); + op = b.tensorOperation(.{}); + const start_rows = try op.commit(op.fetch(block_end, .{}).subi(op.fetch(blocks, .{})).muli(bb), .{}); + op = b.tensorOperation(.{}); + const slot = try op.commit(op.fetch(rank, .{}).addi(op.fetch(start_rows, .{})).muli(op.fetch(onehot, .{})).reduce(&.{E}, .addi).subi(1), .{}); + // The expert of block g: how many experts end at or before g. + const block_ids = try b.arange(chunks * blocks_per_chunk, .i32, &.{BG}, .{}); + op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); + const ends = try op.commit(op.fetch(block_end, .{}).subi(op.fetch(block_ids, .{})), .{}); + op = b.tensorOperation(.{}); + const ended = try op.commit(op.where(ends, .le, 0, 1, 0), .{}); + op = b.tensorOperation(.{}); + // Padding blocks past the last expert reuse the last one. + const block_expert = try op.commit(op.fetch(ended, .{}).reduce(&.{E}, .addi).binary(.mini, cfg.local_num_experts - 1), .{}); + const used = try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(block_end), blocks_per_chunk - 1), blocks_per_chunk); + + // Rows of unused slots are never read back. + const rows = try b.reshape(try b.scatter(in.x, slot, GB, &.{ GB, D }, .{}), &.{ Go, BGi, BB, D }, .{}); + const chunk_experts = try b.reshape(block_expert, &.{ Go, BGi }, .{}); + const gate_up = try b.reshape(in.gate_up, &.{ E, a.S2, I, D }, .{}); + + var loop = b.openFor(used, .{try b.scratchpad(if (format == .fp8_block128) .f32 else .bf16, &.{ Go, BGi, BB, D })}); + { + const chunk = try b.indexRead(rows, loop.iv); + const ids = try b.indexRead(chunk_experts, loop.iv); + const gu = try b.gather(gate_up, ids, E, &.{ BGi, a.S2, I, D }, .{}); + const dn = try b.gather(in.down, ids, E, &.{ BGi, D, I }, .{}); + const scales = try gatherScales(b, a, in, ids, &.{BGi}, format); + const h = try project(b, chunk, gu, scales.gate_up, D, I, &.{ BGi, BB, a.S2, I }, if (format == .fp8_block128) .bf16 else null, format); + const act = try gated(b, a, h, &.{ BGi, BB }, null, cfg.activation, format); + const y = try project(b, act, dn, scales.down, I, D, &.{ BGi, BB, D }, if (format == .fp8_block128) .f32 else .bf16, format); + try loop.yield(.{try b.indexWrite(loop.carried[0], loop.iv, y)}); + } + const out = loop.results[0]; + + const per_route = try b.gather(try b.reshape(out, &.{ GB, D }, .{}), slot, GB, &.{ T, K, D }, .{}); + op = b.tensorOperation(.{}); + b.ret(&.{try op.commit(op.fetch(per_route, .{ .typecast_to = .f32 }).mulf(op.fetch(in.w, .{})).reduce(&.{K}, .addf), .{ .dtype = .bf16 })}); +} + +/// Number of experts processed per device-loop iteration, based on the +/// batch's route count (`cfg.num_tokens * cfg.top_k`). +fn expertsPerChunk(cfg: Cfg) i64 { + const routes: u64 = @intCast(std.math.divCeil(i64, cfg.num_tokens * cfg.top_k, 16) catch unreachable); + return @intCast(std.math.clamp(std.math.ceilPowerOfTwo(u64, routes) catch unreachable, 2, 16)); +} + +/// Process the FP8 experts selected by the batch in groups sized by `expertsPerChunk`. +/// Load each selected expert once and run it for every token. Routing weights +/// scale the F32 down-projection results before summing the experts and steps. +fn fp8SmallBatch(b: *tcl.Builder, a: Axes, in: Args, cfg: Cfg) tcl.FinishError!void { + const chunk = expertsPerChunk(cfg); + const steps = std.math.divCeil(i64, @min(cfg.local_num_experts, cfg.num_tokens * cfg.top_k), chunk) catch unreachable; + const C = b.axis("C", chunk); + const Go = b.axis("Go", steps); + const J = b.axis("J", steps * chunk); + const T, const K, const E, const D, const I = .{ a.T, a.K, a.E, a.D, a.I }; + + // Each token's routing weight per expert, and the used experts. + const experts = try b.arange(cfg.local_num_experts, .i32, &.{E}, .{}); + var op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); + const dist = try op.commit(op.fetch(in.ids, .{}).subi(op.fetch(experts, .{})), .{}); + op = b.tensorOperation(.{}); + const gate = try op.commit(op.where(op.fetch(dist, .{}), .eq, 0, op.fetch(in.w, .{ .broadcast_to = &.{ T, K, E } }), 0.0).reduce(&.{K}, .addf), .{}); + op = b.tensorOperation(.{}); + const onehot = try op.commit(op.where(dist, .eq, 0, 1, 0), .{}); + op = b.tensorOperation(.{}); + const counts = try op.commit(op.fetch(onehot, .{}).reduce(&.{ T, K }, .addi), .{}); + op = b.tensorOperation(.{}); + const used = try op.commit(op.where(counts, .gt, 0, 1, 0), .{}); + op = b.tensorOperation(.{}); + const position = try op.commit(op.fetch(used, .{}).cumsum(&.{E}), .{}); + // A used expert's rank among them, -1 for the others. + op = b.tensorOperation(.{}); + const rank = try op.commit(op.fetch(position, .{}).muli(op.fetch(used, .{})).subi(1), .{}); + const trips = try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(position), chunk - 1), chunk); + + // Slot j holds the used expert of rank j and its tokens' weights; + // slots past the last hold expert 0 with no weight. + const ranks = try b.arange(steps * chunk, .i32, &.{J}, .{}); + op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); + const offset = try op.commit(op.fetch(rank, .{}).subi(op.fetch(ranks, .{})), .{}); + op = b.tensorOperation(.{}); + const slot_experts = try op.commit(op.where(op.fetch(offset, .{}), .eq, 0, op.fetch(experts, .{ .broadcast_to = &.{ E, J } }), 0).reduce(&.{E}, .addi), .{ .dtype = .i32 }); + // A reduction's condition is read plainly, so broadcast it first. + op = b.tensorOperation(.{}); + const token_offset = try op.commit(op.fetch(offset, .{ .broadcast_to = &.{ T, E, J } }), .{}); + op = b.tensorOperation(.{}); + const slot_gate = try op.commit(op.where(op.fetch(token_offset, .{}), .eq, 0, op.fetch(gate, .{ .broadcast_to = &.{ T, E, J } }), 0.0).reduce(&.{E}, .addf), .{ .dtype = .f32, .axes = &.{ J, T } }); + const step_experts = try b.reshape(slot_experts, &.{ Go, C }, .{}); + const step_gate = try b.reshape(slot_gate, &.{ Go, C, T }, .{}); + const gate_up = try b.reshape(in.gate_up, &.{ E, a.S2, I, D }, .{}); + + var loop = b.openFor(trips, .{try b.full(0, .f32, &.{ Go, T, D })}); + { + const ids = try b.indexRead(step_experts, loop.iv); + const weights = try b.indexRead(step_gate, loop.iv); + const gu = try b.gather(gate_up, ids, E, &.{ C, a.S2, I, D }, .{}); + const dn = try b.gather(in.down, ids, E, &.{ C, D, I }, .{}); + const scales = try gatherScales(b, a, in, ids, &.{C}, .fp8_block128); + const h = try project(b, in.x, gu, scales.gate_up, D, I, &.{ C, T, a.S2, I }, .bf16, .fp8_block128); + const act = try gated(b, a, h, &.{ C, T }, null, cfg.activation, .fp8_block128); + const per_expert = try project(b, act, dn, scales.down, I, D, &.{ C, T, D }, .f32, .fp8_block128); + op = b.tensorOperation(.{}); + const y = try op.commit(op.fetch(per_expert, .{}).mulf(op.fetch(weights, .{})).reduce(&.{C}, .addf), .{ .dtype = .f32 }); + try loop.yield(.{try b.indexWrite(loop.carried[0], loop.iv, y)}); + } + op = b.tensorOperation(.{}); + b.ret(&.{try op.commit(op.fetch(loop.results[0], .{}).reduce(&.{Go}, .addf), .{ .dtype = .bf16 })}); +} diff --git a/zml/kernels/moe/tcl/tcl.zig b/zml/kernels/moe/tcl/tcl.zig index 4b7bcfb3..f05727ce 100644 --- a/zml/kernels/moe/tcl/tcl.zig +++ b/zml/kernels/moe/tcl/tcl.zig @@ -1,270 +1,165 @@ //! Furiosa RNGD fused experts in TCL, after Furiosa's production kernels. -//! -//! Few tokens (at most `experts / top_k`): activation stationary, each token -//! gathers the weights of its experts and the down projection contracts over -//! the experts and the intermediate axis at once. -//! -//! More tokens: blockwise. The kernel routes on the device: routes are placed -//! in expert-major blocks of `BB` rows (one-hot, cumulative sums), then a -//! device loop over the used chunks of `BGi` blocks gathers each block's -//! expert weights and runs both projections; outputs are gathered back per -//! route and weighted. const std = @import("std"); -const builder = @import("dsl/tcl/builder"); - const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; const tcl = zml.kernel.tcl; +const kernels = @import("moe.zig"); +const api = @import("../api.zig"); -/// `input` is `{.t, .d}`, `topk_ids` and `topk_weights` `{.t, .topk}`, `gate_up` `{.expert, .dout = 2 * inter (gate then up), .d}` -/// and `down` `{.expert, .d, .dout = inter}`, all bf16 but the ids (i32). -/// SwiGLU, routing weights applied after the down projection. -pub fn fusedExperts(input: Tensor, topk_ids: Tensor, topk_weights: Tensor, gate_up_linear: zml.nn.Linear, down_linear: zml.nn.Linear, opts: zml.moe.Options) Tensor { - stdx.debug.assert(!opts.quantize_input and opts.routing_weight_placement == .after_down, "TCL MoE applies routing weights after the down projection, without FP8 input quantization", .{}); - stdx.debug.assert(opts.activation == .swiglu and std.meta.eql(opts.activation.swiglu, .{}), "TCL MoE supports SwiGLU without limit, scale or bias", .{}); - stdx.debug.assert(gate_up_linear.quantization == null and down_linear.quantization == null, "TCL MoE takes bf16 experts", .{}); +/// `input` is `{.t, .d}`, routing IDs and weights are `{.t, .topk}`, `gate_up` +/// is `{.expert, .dout = 2 * inter (gate then up), .d}`, and `down` is +/// `{.expert, .d, .dout = inter}`. Experts are BF16 or block-128 FP8 with +/// BF16/F32 scales. Supports plain SwiGLU and BF16 GEGLU with tanh GELU. +pub fn fusedExperts(input: Tensor, routing: api.LocalRouting, gate_up_linear: zml.nn.Linear, down_linear: zml.nn.Linear, opts: api.Options) Tensor { + const topk_ids = routing.localIds(); + stdx.debug.assert(!opts.quantize_input and opts.routing_weight_placement == .after_down, "TCL MoE requires quantize_input=false and routing_weight_placement=after_down", .{}); stdx.debug.assert(gate_up_linear.bias == null and down_linear.bias == null, "TCL MoE takes experts without bias", .{}); + stdx.debug.assert(input.dtype() == .bf16 and topk_ids.dtype() == .i32, "TCL MoE takes BF16 activations and i32 ids", .{}); + const format = expertFormat(gate_up_linear, down_linear); + const swiglu = opts.activation == .swiglu and std.meta.eql(opts.activation.swiglu, .{}); + stdx.debug.assert(swiglu or (format == .bf16 and opts.activation == .geglu_tanh), "TCL MoE supports plain SwiGLU or GEGLU-tanh with BF16 experts, got {s}", .{@tagName(opts.activation)}); + const gate_up = gate_up_linear.weight; const down = down_linear.weight; - stdx.debug.assert(input.dtype() == .bf16 and gate_up.dtype() == .bf16 and down.dtype() == .bf16 and topk_ids.dtype() == .i32, "TCL MoE takes bf16 activations and experts, i32 ids", .{}); - const t = input.dim(.t); - const k = topk_ids.dim(.topk); - const cfg: Cfg = .{ - .t = t, - .k = k, - .e = gate_up.dim(.expert), - .d = input.dim(.d), - .i = down.dim(.dout), + const cfg: kernels.Cfg = .{ + .num_tokens = input.dim(.t), + .top_k = topk_ids.dim(.topk), + .local_num_experts = gate_up.dim(.expert), + .hidden_size = input.dim(.d), + .expert_intermediate_size = down.dim(.dout), + .weight_scale_dtype = switch (format) { + .bf16 => .bf16, + .fp8_block128 => |s| tcl.from(s.gate_up.dtype()), + }, + .activation = std.meta.activeTag(opts.activation), }; - stdx.debug.assert(gate_up.dim(.dout) == 2 * cfg.i and down.dim(.d) == cfg.d and gate_up.dim(.d) == cfg.d, "TCL MoE: mismatched expert shapes", .{}); + stdx.debug.assert(gate_up.dim(.dout) == 2 * cfg.expert_intermediate_size and down.dim(.d) == cfg.hidden_size and gate_up.dim(.d) == cfg.hidden_size, "TCL MoE: mismatched expert shapes", .{}); - const x = input.reshape(.{ t, cfg.d }); - const ids = topk_ids.reshape(.{ t, k }); - const weights = topk_weights.convert(.f32).reshape(.{ t, k }); - const out_shape = x.shape(); - const y = if (cfg.t * cfg.k <= cfg.e) - ActivationStationary.call(.{ .x = x, .ids = ids, .w = weights, .gate_up = gate_up, .down = down }, .{ .y = out_shape }, .{ .cfg = cfg }).y - else - Blockwise.call(.{ .x = x, .ids = ids, .w = weights, .gate_up = gate_up, .down = down }, .{ .y = out_shape }, .{ .cfg = cfg }).y; - return y.reshape(input.shape()); + const weights = routing.weights.convert(.f32); + const local_weights = switch (format) { + .bf16 => weights, + .fp8_block128 => topk_ids.cmp(.GE, .scalar(0, .i32)).select(weights, .scalar(0, .f32)), + }; + return switch (format) { + .bf16 => selectScheduleAndRun(.bf16, cfg, .{ .x = input, .ids = topk_ids, .w = local_weights, .gate_up = gate_up, .down = down }), + .fp8_block128 => |s| selectScheduleAndRun(.fp8_block128, cfg, .{ + .x = input, + .ids = topk_ids, + .w = local_weights, + .gate_up = gate_up, + .down = down, + .gate_up_scale = s.gate_up, + .down_scale = s.down, + }), + }; } -pub const Cfg = struct { t: i64, k: i64, e: i64, d: i64, i: i64 }; +const Scales = struct { gate_up: Tensor, down: Tensor }; -const Axes = struct { - T: tcl.Axis, - K: tcl.Axis, - E: tcl.Axis, - D: tcl.Axis, - I: tcl.Axis, - I2: tcl.Axis, - S2: tcl.Axis, - S1: tcl.Axis, - - fn init(b: *tcl.Builder, cfg: Cfg) Axes { - return .{ - .T = b.axis("T", cfg.t), - .K = b.axis("K", cfg.k), - .E = b.axis("E", cfg.e), - .D = b.axis("D", cfg.d), - .I = b.axis("I", cfg.i), - .I2 = b.axis("I2", 2 * cfg.i), - .S2 = b.axis("S2", 2), - .S1 = b.axis("S1", 1), - }; - } +const Format = union(kernels.Format) { + bf16: void, + fp8_block128: Scales, }; -fn args(b: *tcl.Builder, a: Axes) tcl.FinishError!Args { - const t = try b.declareArgs(.{ - .x = .{ .dtype = .bf16, .axes = &.{ a.T, a.D } }, - .ids = .{ .dtype = .i32, .axes = &.{ a.T, a.K } }, - .w = .{ .dtype = .f32, .axes = &.{ a.T, a.K } }, - .gate_up = .{ .dtype = .bf16, .axes = &.{ a.E, a.I2, a.D } }, - .down = .{ .dtype = .bf16, .axes = &.{ a.E, a.D, a.I } }, - }); - return .{ .x = t.x, .ids = t.ids, .w = t.w, .gate_up = t.gate_up, .down = t.down }; +fn expertFormat(gate_up: zml.nn.Linear, down: zml.nn.Linear) Format { + const scheme = gate_up.quantizationScheme(); + stdx.debug.assert(scheme == down.quantizationScheme(), "TCL MoE projections must have matching quantization schemes", .{}); + return switch (scheme orelse { + stdx.debug.assert(gate_up.weight.dtype() == .bf16 and down.weight.dtype() == .bf16, "TCL MoE unquantized expert weights must be BF16", .{}); + return .bf16; + }) { + .fp8_block128 => blk: { + const scales: Scales = .{ .gate_up = fp8Scales(gate_up), .down = fp8Scales(down) }; + stdx.debug.assert(scales.gate_up.dtype() == scales.down.dtype(), "TCL MoE projection scales must have matching types", .{}); + break :blk .{ .fp8_block128 = scales }; + }, + else => |unsupported| stdx.debug.panic("TCL MoE does not support quantization scheme {s}", .{@tagName(unsupported)}), + }; } -const Args = struct { x: builder.Tensor, ids: builder.Tensor, w: builder.Tensor, gate_up: builder.Tensor, down: builder.Tensor }; - -/// `silu(gate) * up` of `h = [batch..., S2, I]` (gate first), times -/// `weights` (`[batch...]`) unless null. -fn gated(b: *tcl.Builder, a: Axes, h: builder.Tensor, batch: []const tcl.Axis, weights: ?builder.Tensor) tcl.FinishError!builder.Tensor { - var sliced: [4]tcl.Axis = undefined; - var flat: [3]tcl.Axis = undefined; - @memcpy(sliced[0..batch.len], batch); - @memcpy(flat[0..batch.len], batch); - sliced[batch.len] = a.S1; - sliced[batch.len + 1] = a.I; - flat[batch.len] = a.I; - const s = sliced[0 .. batch.len + 2]; - const f = flat[0 .. batch.len + 1]; - const gate = try b.reshape(try b.slice(h, a.S2, 0, s, .{}), f, .{}); - const up = try b.reshape(try b.slice(h, a.S2, 1, s, .{}), f, .{}); - // Interleaving requires one binary instruction; SwiGLU chains several. - const op = b.tensorOperation(.{ .tactic = .elementwise }); - const g = op.fetch(gate, .{}); - var v = g.sigmoid().mulf(g).mulf(op.fetch(up, .{})); - if (weights) |w| v = v.mulf(op.fetch(w, .{})); - return op.commit(v, .{ .dtype = .bf16 }); +fn fp8Scales(linear: zml.nn.Linear) Tensor { + const q = linear.quantization.?; + stdx.debug.assert(linear.weight.dtype() == .f8e4m3fn and !q.swizzled_scales and q.global_scale == null and q.input_scale == null, "TCL MoE takes plain E4M3FN block-128 weights", .{}); + stdx.debug.assert(q.scales.dtype() == .bf16 or q.scales.dtype() == .f32, "TCL MoE FP8 scales must be BF16 or F32", .{}); + const w = linear.weight; + const scale = q.scales; + stdx.debug.assert(@mod(w.dim(.d), 128) == 0 and @mod(w.dim(.dout), 128) == 0 and scale.dim(.expert) == w.dim(.expert) and scale.dim(.d) * 128 == w.dim(.d) and scale.dim(.dout) * 128 == w.dim(.dout), "TCL MoE FP8 scales must cover 128x128 weight blocks", .{}); + return scale; } -/// Furiosa's compiler configuration for MoE kernels (`qwen3_moe/config.py`): -/// IO-bound tactics for activation stationary, compute-bound for blockwise. -/// Keep the compiler's default lowering mode and omit the adaptive-einsum and -/// half-VRF settings removed in TCC 2026.5. -fn configure(b: *tcl.Builder, blockwise: bool) void { - if (blockwise) b.compilerConfig(.{ - .enable_einsum_fusion = true, - .padding_policy = .{ .Small = 1.1 }, - .instruction_mem_budget = 0xB0000, - .reshape_einsum_mode = .{ .Reshape = .{ .permute = true } }, - .tensor_unit_bridge_threshold_in_page = 12, - .allow_external_operators = false, - .enable_tactic_pruning = false, - .scheduler_beam_search = true, - .allow_reduce_by_ve_cluster_chip_reduce = false, - .allow_reduce_by_ve_cluster_chip_reduce_base_population = false, - .tactic_hint = .{ .ForLlmModelComputeBound = 200 }, - .dma_preference = 1.2, - .propagate_sparse_axis_from_op = "None", - }) else b.compilerConfig(.{ - .enable_einsum_fusion = true, - .padding_policy = .{ .Small = 1.1 }, - .instruction_mem_budget = 0xB0000, - .reshape_einsum_mode = .{ .Reshape = .{ .permute = true } }, - .tensor_unit_bridge_threshold_in_page = 12, - .allow_external_operators = false, - .enable_tactic_pruning = false, - .scheduler_beam_search = true, - .allow_reduce_by_ve_cluster_chip_reduce = false, - .allow_reduce_by_ve_cluster_chip_reduce_base_population = false, - .tactic_hint = .{ .ForLlmModelIOBound = 200 }, - .num_transaction_simulation_per_pe = 1024, - .dma_preference = 0.8, - .propagate_sparse_axis_from_op = "None", - }); +// Maximum total input tokens per MoE call before padding for the FP8 +// small-batch schedule. It loads each used expert's weights once and reuses +// them across the batch. +// Larger batches fall back to per-route or blockwise dispatch. +const fp8_small_batch_max_tokens = 64; + +fn selectScheduleAndRun(comptime format: kernels.Format, cfg: kernels.Cfg, inputs: kernels.Kernel(format, .per_route).Inputs) Tensor { + if (format == .fp8_block128 and cfg.num_tokens <= fp8_small_batch_max_tokens) { + return run(format, .small_batch, cfg, inputs); + } + if (cfg.num_tokens * cfg.top_k <= cfg.local_num_experts) { + return run(format, .per_route, cfg, inputs); + } + return run(format, .blockwise, cfg, inputs); } -pub const ActivationStationary = tcl.Kernel(Cfg, .{ - .name = "moe_activation_stationary", - .inputs = &.{ "x", "ids", "w", "gate_up", "down" }, - .outputs = &.{"y"}, - .run = struct { - fn run(b: *tcl.Builder, cfg: Cfg) tcl.FinishError!void { - configure(b, false); - const a: Axes = .init(b, cfg); - const in = try args(b, a); - const gate_up = try b.reshape(in.gate_up, &.{ a.E, a.S2, a.I, a.D }, .{}); - const gu = try b.gather(gate_up, in.ids, a.E, &.{ a.T, a.K, a.S2, a.I, a.D }, .{}); - const dn = try b.gather(in.down, in.ids, a.E, &.{ a.T, a.K, a.D, a.I }, .{}); +fn run(comptime format: kernels.Format, comptime schedule: kernels.Schedule, cfg: kernels.Cfg, inputs: kernels.Kernel(format, .per_route).Inputs) Tensor { + const call_cfg, const prepared = if (schedule == .small_batch) + prepareFp8SmallBatch(cfg, inputs) + else + .{ cfg, prepareRoutedInputs(cfg, inputs) }; - var op = b.tensorOperation(.{}); - const h = try op.commit(op.contract(in.x, gu, &.{ a.T, a.K, a.S2, a.I }), .{}); - // Routing weights scale the activations, so the down projection - // also sums over the experts. - const act = try gated(b, a, h, &.{ a.T, a.K }, in.w); - op = b.tensorOperation(.{}); - b.ret(&.{try op.commit(op.contract(act, dn, &.{ a.T, a.D }), .{ .dtype = .bf16 })}); - } - }.run, -}); + const shape = prepared.x.shape(); + var y = kernels.Kernel(format, schedule).call(prepared, .{ .y = shape }, .{ .cfg = call_cfg }).y; -/// Rows per expert block: the next power of two of the routes per expert. -pub fn blockRows(cfg: Cfg) i64 { - const per_expert = std.math.divCeil(i64, cfg.t * cfg.k, cfg.e) catch unreachable; - return @intCast(std.math.ceilPowerOfTwo(u64, @intCast(@max(per_expert, 4))) catch unreachable); + if (schedule == .small_batch) y = y.slice(0, .{ .end = cfg.num_tokens }); + return y.reshape(inputs.x.shape()); } -const blocks_per_chunk = 32; +fn prepareRoutedInputs(cfg: kernels.Cfg, inputs: anytype) @TypeOf(inputs) { + var prepared = inputs; + prepared.x = inputs.x.reshape(.{ cfg.num_tokens, cfg.hidden_size }); + const ids = inputs.ids.reshape(.{ cfg.num_tokens, cfg.top_k }); + const w = inputs.w.reshape(.{ cfg.num_tokens, cfg.top_k }); + // These kernels gather every route, and blockwise also scatters each route + // into an expert slot. Give nonlocal routes a valid slot and zero weight. + const valid = ids.cmp(.GE, Tensor.scalar(0, .i32)).logical(.AND, ids.cmp(.LT, Tensor.scalar(cfg.local_num_experts, .i32))); + prepared.ids = valid.select(ids, Tensor.scalar(0, .i32)); + prepared.w = valid.select(w, Tensor.scalar(0, .f32)); + return prepared; +} -pub const Blockwise = tcl.Kernel(Cfg, .{ - .name = "moe_blockwise", - .inputs = &.{ "x", "ids", "w", "gate_up", "down" }, - .outputs = &.{"y"}, - .run = struct { - fn run(b: *tcl.Builder, cfg: Cfg) tcl.FinishError!void { - configure(b, true); - const a: Axes = .init(b, cfg); - const in = try args(b, a); - const bb = blockRows(cfg); - // One block per expert at least, then enough for every route. - const max_blocks = std.math.divCeil(i64, @max(cfg.t * cfg.k - cfg.e, 0), bb) catch unreachable; - const chunks = std.math.divCeil(i64, max_blocks + cfg.e, blocks_per_chunk) catch unreachable; - const BB = b.axis("BB", bb); - const BGi = b.axis("BGi", blocks_per_chunk); - const Go = b.axis("Go", chunks); - const BG = b.axis("BG", chunks * blocks_per_chunk); - const GB = b.axis("GB", chunks * blocks_per_chunk * bb); - const T, const K, const E, const D, const I = .{ a.T, a.K, a.E, a.D, a.I }; - - // Routing: route (t, k) goes to row `slot` of block `start[e] + rank`. - const experts = try b.arange(cfg.e, .i32, &.{E}, .{}); - var op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); - const dist = try op.commit(op.fetch(in.ids, .{}).subi(op.fetch(experts, .{})), .{}); - op = b.tensorOperation(.{}); - const onehot = try op.commit(op.where(dist, .eq, 0, 1, 0), .{}); - op = b.tensorOperation(.{}); - const rank = try op.commit(op.fetch(onehot, .{}).cumsum(&.{ T, K }), .{}); - op = b.tensorOperation(.{}); - const log2_bb: i64 = std.math.log2_int(u64, @intCast(bb)); - const blocks = try op.commit(op.fetch(onehot, .{}).reduce(&.{ T, K }, .addi).addi(bb - 1).binary(.shr_arith, log2_bb), .{}); - op = b.tensorOperation(.{}); - const block_end = try op.commit(op.fetch(blocks, .{}).cumsum(&.{E}), .{}); - op = b.tensorOperation(.{}); - const start_rows = try op.commit(op.fetch(block_end, .{}).subi(op.fetch(blocks, .{})).muli(bb), .{}); - op = b.tensorOperation(.{}); - const slot = try op.commit(op.fetch(rank, .{}).addi(op.fetch(start_rows, .{})).muli(op.fetch(onehot, .{})).reduce(&.{E}, .addi).subi(1), .{}); - // The expert of block g: how many experts end at or before g. - const block_ids = try b.arange(chunks * blocks_per_chunk, .i32, &.{BG}, .{}); - op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); - const ends = try op.commit(op.fetch(block_end, .{}).subi(op.fetch(block_ids, .{})), .{}); - op = b.tensorOperation(.{}); - const ended = try op.commit(op.where(ends, .le, 0, 1, 0), .{}); - op = b.tensorOperation(.{}); - // Padding blocks past the last expert reuse the last one. - const block_expert = try op.commit(op.fetch(ended, .{}).reduce(&.{E}, .addi).binary(.mini, cfg.e - 1), .{}); - const used = try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(block_end), blocks_per_chunk - 1), blocks_per_chunk); - - // Rows of unused slots are never read back. - const rows = try b.reshape(try b.scatter(in.x, slot, GB, &.{ GB, D }, .{}), &.{ Go, BGi, BB, D }, .{}); - const chunk_experts = try b.reshape(block_expert, &.{ Go, BGi }, .{}); - const gate_up = try b.reshape(in.gate_up, &.{ E, a.S2, I, D }, .{}); - - var loop = b.openFor(used, .{try b.scratchpad(.bf16, &.{ Go, BGi, BB, D })}); - { - const chunk = try b.indexRead(rows, loop.iv); - const ids = try b.indexRead(chunk_experts, loop.iv); - const gu = try b.gather(gate_up, ids, E, &.{ BGi, a.S2, I, D }, .{}); - const dn = try b.gather(in.down, ids, E, &.{ BGi, D, I }, .{}); - op = b.tensorOperation(.{}); - const h = try op.commit(op.contract(chunk, gu, &.{ BGi, BB, a.S2, I }), .{}); - const act = try gated(b, a, h, &.{ BGi, BB }, null); - op = b.tensorOperation(.{}); - const y = try op.commit(op.contract(act, dn, &.{ BGi, BB, D }), .{ .dtype = .bf16 }); - try loop.yield(.{try b.indexWrite(loop.carried[0], loop.iv, y)}); - } - const out = loop.results[0]; - - const per_route = try b.gather(try b.reshape(out, &.{ GB, D }, .{}), slot, GB, &.{ T, K, D }, .{}); - op = b.tensorOperation(.{}); - b.ret(&.{try op.commit(op.fetch(per_route, .{ .typecast_to = .f32 }).mulf(op.fetch(in.w, .{})).reduce(&.{K}, .addf), .{ .dtype = .bf16 })}); - } - }.run, -}); +fn prepareFp8SmallBatch(cfg: kernels.Cfg, inputs: kernels.Kernel(.fp8_block128, .small_batch).Inputs) struct { kernels.Cfg, @TypeOf(inputs) } { + // WORKAROUND (TCC 2026.5.0b0): a token count other than a power of two + // panics its adder tree ("Fail to cover valid range", 5 tokens padded + // to 8): pad with unrouted tokens. + var padded = cfg; + padded.num_tokens = @intCast(std.math.ceilPowerOfTwo(u64, @intCast(cfg.num_tokens)) catch unreachable); + const pad = padded.num_tokens - cfg.num_tokens; + var prepared = inputs; + prepared.x = inputs.x.pad(0, .{ .t = Tensor.Pad{ .high = pad } }).reshape(.{ padded.num_tokens, cfg.hidden_size }); + prepared.ids = inputs.ids.pad(-1, .{ .t = Tensor.Pad{ .high = pad } }).reshape(.{ padded.num_tokens, cfg.top_k }); + prepared.w = inputs.w.pad(0, .{ .t = Tensor.Pad{ .high = pad } }).reshape(.{ padded.num_tokens, cfg.top_k }); + return .{ padded, prepared }; +} test "tcl fused experts run on furiosa" { + try testFusedExperts(.{ .swiglu = .{} }); + try testFusedExperts(.geglu_tanh); +} + +fn testFusedExperts(activation: zml.moe.Activation) !void { const platform = zml.testing.env(); if (platform.target != .furiosa) return error.SkipZigTest; const allocator = std.testing.allocator; const io = std.testing.io; const bf16 = zml.floats.BFloat16; - // 2 * 4 <= 8 experts is activation stationary, 16 tokens is blockwise. + // With two routes per token and eight experts, four tokens use per-route + // dispatch and sixteen use blockwise dispatch. for ([_]i64{ 4, 16 }) |t| { const k = 2; const e = 8; @@ -276,15 +171,15 @@ test "tcl fused experts run on furiosa" { const gu: Tensor = .init(.{ .expert = e, .dout = 2 * i, .d = d }, .bf16); const dn: Tensor = .init(.{ .expert = e, .d = d, .dout = i }, .bf16); const Mod = struct { - pub fn forward(x_: Tensor, ids_: Tensor, w_: Tensor, gu_: Tensor, dn_: Tensor) Tensor { - return fusedExperts(x_, ids_, w_, .init(gu_, null, .dout), .init(dn_, null, .dout), .{ - .activation = .{ .swiglu = .{} }, + pub fn forward(x_: Tensor, ids_: Tensor, w_: Tensor, gu_: Tensor, dn_: Tensor, act: zml.moe.Activation) Tensor { + return zml.moe.forwardMoe(x_, ids_, w_, .init(gu_, null, .dout), .init(dn_, null, .dout), .tcl, zml.Compiler.current().platform.replicated_mesh, .expert(.replicated), .{ + .activation = act, .quantize_input = false, .routing_weight_placement = .after_down, }); } }; - var exe = try zml.module.compile(allocator, io, Mod.forward, .{ x, ids, w, gu, dn }, platform, .{}); + var exe = try platform.compileFn(allocator, io, Mod.forward, .{ x, ids, w, gu, dn, activation }, .{}); defer exe.deinit(); var prng: std.Random.DefaultPrng = .init(@intCast(t)); @@ -336,7 +231,12 @@ test "tcl fused experts run on furiosa" { g += xv * hgu[(ex * 2 * i + n) * d + m].toF32(); u += xv * hgu[(ex * 2 * i + i + n) * d + m].toF32(); } - act[n] = g / (1 + @exp(-g)) * u; + const gate = switch (activation) { + .swiglu => g / (1 + @exp(-g)), + .geglu_tanh => 0.5 * g * (1 + std.math.tanh(@sqrt(2.0 / std.math.pi) * (g + 0.044715 * g * g * g))), + else => unreachable, + }; + act[n] = gate * u; } for (0..d) |m| { var y: f32 = 0; diff --git a/zml/kernels/moe/tests.zig b/zml/kernels/moe/tests.zig index 2bdebdfe..c74406c3 100644 --- a/zml/kernels/moe/tests.zig +++ b/zml/kernels/moe/tests.zig @@ -5,12 +5,15 @@ const std = @import("std"); const zml = @import("../../zml.zig"); const Tensor = zml.Tensor; const Shape = zml.Shape; +const Activation = @import("api.zig").Activation; +const ExpertsLayout = @import("api.zig").ExpertsLayout; +const GlobalOptions = @import("api.zig").Options; +const Placement = @import("api.zig").RoutingWeightPlacement; const reference = @import("stablehlo/stablehlo.zig"); -const Placement = @import("moe.zig").RoutingWeightPlacement; const Case = struct { backend: zml.moe.Backend, - activation: zml.moe.Activation = .{ .swiglu = .{} }, + activation: Activation = .{ .swiglu = .{} }, scheme: ?zml.Quantization.Scheme = null, activation_dtype: zml.DataType = .bf16, quantize_input: bool = false, @@ -22,15 +25,22 @@ const Case = struct { width: i64 = 128, intermediate: i64 = 128, experts: i64 = 8, + expert_mesh: *const zml.Mesh, + + fn inputShape(c: Case) Shape { + const shape = Shape.init(.{ .t = c.tokens, .d = c.width }, c.activation_dtype); + const spec: zml.Sharding.Partitioning.Logical = if (c.expert_mesh.resolveLogicalAxis("data") != null) .parse(.{ .t = .data }) else .replicated; + return shape.withPartitioning(c.expert_mesh, spec); + } pub fn compile(c: Case, allocator: std.mem.Allocator, io: std.Io, platform: *const zml.Platform) !zml.Exe { std.debug.assert(c.backend.isAvailable(platform)); - const x: Tensor = .init(.{ .t = c.tokens, .d = c.width }, c.activation_dtype); - return platform.compileFn(allocator, io, forward, .{ x, c }, .{}); + const x: Tensor = .fromShape(c.inputShape()); + return platform.compileFn(allocator, io, forward, .{ x, c }, .{ .meshes = &.{c.expert_mesh} }); } pub fn check(c: Case, allocator: std.mem.Allocator, io: std.Io, platform: *const zml.Platform, exe: *const zml.Exe) !void { - const x: Tensor = .init(.{ .t = c.tokens, .d = c.width }, c.activation_dtype); + const x: Tensor = .fromShape(c.inputShape()); var host = try zml.Slice.alloc(allocator, x.shape()); defer host.free(allocator); @@ -52,6 +62,7 @@ const Case = struct { var output: zml.Bufferized(Outputs) = undefined; runner.run(io, .{input}, .{&output}, .{ .wait = true }); defer zml.Buffer.deinitAll(Outputs, &output); + try std.testing.expectEqualDeep(c.inputShape()._sharding, output.actual.shape()._sharding); // BF16 intermediate rounding and different GEMM reduction orders. Every // element must pass; near-zero values use the absolute bound. @@ -168,20 +179,16 @@ fn makeExpertFixture(c: Case) ExpertFixture { return fixture; } -/// Interleave the gate/up halves of the output axis: [E, 2, N/2, K] -> [E, N/2, 2, K]. -fn interleave(tensor: Tensor) Tensor { - const tags = tensor.shape().tags(); - return tensor.withTags(.{ .expert, .out, .in }) - .splitAxis(.out, .{ .projection = 2, .mid = @divExact(tensor.dim(1), 2) }) - .transpose(.{ .expert, .mid, .projection, .in }) - .merge(.{ .out = .{ .mid, .projection } }).withTags(tags); +fn interleave(tensor: Tensor, rows: i64) Tensor { + return tensor.reshape(.{ .expert = tensor.dim(0), .projection = 2, .row_block = @divExact(tensor.dim(1), 2 * rows), .row = rows, .column = tensor.dim(2) }) + .transpose(.{ .expert, .row_block, .projection, .row, .column }).reshape(tensor.shape()); } /// Blackwell block scales: [E, N/128, 4, 32, K/4, 4] -> /// [E, N/128, K/4, 32, 4, 4], retaining the declared [E, N, K] shape. fn swizzleScales(scales: Tensor) Tensor { - return scales.reshape(.{ scales.dim(0), @divExact(scales.dim(1), 128), 4, 32, @divExact(scales.dim(2), 4), 4 }) - .transpose(.{ 0, 1, 4, 3, 2, 5 }).reshape(scales.shape()); + return scales.reshape(.{ .expert = scales.dim(0), .row_block = @divExact(scales.dim(1), 128), .row_quad = 4, .row_lane = 32, .scale_word = @divExact(scales.dim(2), 4), .scale_byte = 4 }) + .transpose(.{ .expert, .row_block, .scale_word, .row_lane, .row_quad, .scale_byte }).reshape(scales.shape()); } fn packFp4Bytes(weight: Tensor) Tensor { @@ -189,10 +196,11 @@ fn packFp4Bytes(weight: Tensor) Tensor { return weight.reshape(weight.shape().setDim(2, @divExact(weight.dim(2), 2)).append(.{ .nibble = 2 })).bitCast(.u8); } -fn prepareGateUp(tensor: Tensor, layout: zml.moe.ExpertsLayout) Tensor { +fn prepareGateUp(tensor: Tensor, layout: ExpertsLayout) Tensor { return switch (layout.gate_up) { .concatenated => tensor, - .interleaved => interleave(tensor), + .row_interleaved_1 => interleave(tensor, 1), + .row_interleaved_8 => interleave(tensor, 8), }; } @@ -210,17 +218,29 @@ fn prepareForBackend(canonical: ExpertFixture, backend: zml.moe.Backend) ExpertF } for ([_]*zml.nn.Linear{ &prepared.gate_up, &prepared.down }) |projection| { - if (layout.packing == .swizzled_scales or layout.packing == .flashinfer_nvfp4) { - const q = &projection.quantization.?; - q.scales = swizzleScales(q.scales); - q.swizzled_scales = true; + switch (layout.packing) { + .swizzled_scales, .flashinfer_nvfp4 => { + const q = &projection.quantization.?; + q.scales = swizzleScales(q.scales); + q.swizzled_scales = true; + }, + .mxfp4_interleaved8 => { + const q = &projection.quantization.?; + const sf = q.scales; + q.scales = sf.bitCast(.u8) + .reshape(.{ .expert = sf.dim(0), .row_block = @divExact(sf.dim(1), 128), .row_quad = 4, .row_lane = 32, .scale_word = @divExact(sf.dim(2), 4), .scale_byte = 4 }) + .transpose(.{ .expert, .scale_word, .row_block, .row_lane, .row_quad, .scale_byte }) + .reshape(.{ .expert = sf.dim(0), .scale_word = @divExact(sf.dim(2), 4), .row = sf.dim(1), .scale_byte = 4 }).bitCast(.i32) + .withTags(sf.shape().transpose(.{ 0, 2, 1 })); + }, + .plain => {}, } switch (backend) { .cute => { projection.weight = packFp4Bytes(projection.weight); const q = &projection.quantization.?; - q.scales = q.scales.bitCast(.u8); if (scheme == .nvfp4) { + q.scales = q.scales.bitCast(.u8); // CuTe takes the input quantization multiplier and a GEMM // output multiplier that folds in its inverse. const input_scale = q.input_scale.?.asMultiplier(); @@ -263,14 +283,26 @@ fn forward(input: Tensor, c: Case) Outputs { const ids = route.mul(Tensor.scalar(5, .i32)).addConstant(2).remainder(Tensor.scalar(@min(c.experts, 6), .i32)); const weights = route.remainder(Tensor.scalar(3, .i32)).convert(.f32).addConstant(1).scale(0.25); - const options: zml.moe.Options = .{ + const options: GlobalOptions = .{ .activation = c.activation, .quantize_input = c.quantize_input, .routing_weight_placement = c.placement, }; - const actual = zml.moe.forwardMoe(input, ids, weights, prepared.gate_up, prepared.down, c.backend, options); - const expected = reference.reference(input, ids, weights, canonical.gate_up, canonical.down, .{ + const data_spec: zml.Sharding.Partitioning.Logical = if (c.expert_mesh.resolveLogicalAxis("data") != null) .parse(.{ .t = .data }) else .replicated; + const expert_spec: zml.Sharding.Partitioning.Logical = if (c.expert_mesh.resolveLogicalAxis("experts") != null) .parse(.{ .expert = .experts }) else .replicated; + const actual = zml.moe.forwardMoe(input, ids, weights, prepared.gate_up, prepared.down, c.backend, c.expert_mesh, .{ + .in_specs = .{ + .input = data_spec, + .ids = data_spec, + .weights = data_spec, + .gate_up = expert_spec, + .down = expert_spec, + }, + .out_specs = data_spec, + }, options); + // Keep the oracle input replicated so token sharding does not change its GEMM rounding. + const expected = reference.reference(input.replicate(), ids, weights, canonical.gate_up, canonical.down, .{ .activation = c.activation, .quantize_input = c.quantize_input, .routing_weight_placement = c.placement, @@ -285,7 +317,7 @@ fn forward(input: Tensor, c: Case) Outputs { return .{ .actual = actual, .expected = expected }; } -const gated_activations = [_]zml.moe.Activation{ +const gated_activations = [_]Activation{ .{ .swiglu = .{} }, .{ .swiglu = .{ .limit = 0.25 } }, .{ .swiglu = .{ .bias = 1 } }, @@ -295,8 +327,8 @@ const gated_activations = [_]zml.moe.Activation{ .geglu, .geglu_tanh, }; -const all_activations = gated_activations ++ [_]zml.moe.Activation{ .relu, .silu, .gelu }; -const clipped_swiglu = [_]zml.moe.Activation{ +const all_activations = gated_activations ++ [_]Activation{ .relu, .silu, .gelu }; +const clipped_swiglu = [_]Activation{ .{ .swiglu = .{ .limit = 0.25 } }, .{ .swiglu = .{ .limit = 7 } }, }; @@ -334,7 +366,7 @@ const Format = struct { const Matrix = struct { formats: []const Format = &.{.{}}, - activations: []const zml.moe.Activation, + activations: []const Activation, quantize_inputs: []const bool = &.{false}, placements: []const Placement = &.{.after_down}, biases: []const bool = &.{false}, @@ -346,7 +378,7 @@ const Matrix = struct { } /// Shapes vary fastest, matching the nested loops in the backend tests. - pub fn caseAt(self: Matrix, backend: zml.moe.Backend, index: usize) Case { + pub fn caseAt(self: Matrix, backend: zml.moe.Backend, expert_mesh: *const zml.Mesh, index: usize) Case { std.debug.assert(index < self.count()); var remaining = index; const shape = self.shapes[remaining % self.shapes.len]; @@ -374,6 +406,7 @@ const Matrix = struct { .width = shape.width, .intermediate = shape.intermediate, .experts = shape.experts, + .expert_mesh = expert_mesh, }; } }; @@ -414,7 +447,7 @@ const Pipeline = struct { var index: usize = 0; for (0..matrix.count()) |local_index| { - const c = matrix.caseAt(backend, local_index); + const c = matrix.caseAt(backend, self.platform.replicated_mesh, local_index); if (available) { case_queue.putOne(io, .{ .index = index, .case = c }) catch |err| return switch (err) { error.Closed => {}, @@ -591,6 +624,7 @@ test "Triton MoE prefill launch buckets fit shared memory" { for ([_]i64{ 8, 32, 96, 128, 256, 512, 1024, 2048, 3072, 4096 }) |tokens| { const case: Case = .{ .backend = .triton, + .expert_mesh = platform.replicated_mesh, .tokens = tokens, .width = 256, .intermediate = 128, diff --git a/zml/kernels/moe/triton/moe.zig b/zml/kernels/moe/triton/moe.zig index 3eb457da..6e4006fa 100644 --- a/zml/kernels/moe/triton/moe.zig +++ b/zml/kernels/moe/triton/moe.zig @@ -12,6 +12,7 @@ const tri = zml.kernel.triton; const Builder = tri.Builder; const Value = tri.Value; const DType = tri.DType; +const RoutingWeightPlacement = @import("../api.zig").RoutingWeightPlacement; const log = std.log.scoped(.@"zml/moe/triton_kernels"); @@ -1026,7 +1027,7 @@ pub const Combine = struct { topk: i64, expert_offset: i64, experts: i64, - routing_weight_placement: zml.moe.RoutingWeightPlacement = .before_down, + routing_weight_placement: RoutingWeightPlacement = .before_down, bn: i64, }; pub const Kernel = tri.Kernel(Cfg, .{ .name = "mxfp4_triton_combine", .inputs = &.{ "d", "ids", "perm", "rw" }, .outputs = &.{"y"}, .run = run }); @@ -1092,7 +1093,7 @@ const GemmConfig = struct { // Swiglu clamping. Set to 0 to disable. swiglu_limit: f32, scale_stride: i64, - routing_weight_placement: zml.moe.RoutingWeightPlacement = .before_down, + routing_weight_placement: RoutingWeightPlacement = .before_down, bn: i64, bk: i64, }; diff --git a/zml/kernels/moe/triton/triton.zig b/zml/kernels/moe/triton/triton.zig index 908cd163..8ba799a5 100644 --- a/zml/kernels/moe/triton/triton.zig +++ b/zml/kernels/moe/triton/triton.zig @@ -6,10 +6,14 @@ const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; const Shape = zml.Shape; const DataType = zml.DataType; -const kernels = @import("moe.zig"); const toDType = zml.kernel.triton.from; -const ProjectionLayout = zml.moe.ProjectionLayout; -const RoutingWeightPlacement = zml.moe.RoutingWeightPlacement; +const ExpertsLayout = @import("../api.zig").ExpertsLayout; +const ProjectionLayout = @import("../api.zig").ProjectionLayout; +const RoutingWeightPlacement = @import("../api.zig").RoutingWeightPlacement; +const Activation = @import("../api.zig").Activation; +const Options = @import("../api.zig").Options; +const LocalRouting = @import("../api.zig").LocalRouting; +const kernels = @import("moe.zig"); const log = std.log.scoped(.@"zml/moe/triton"); @@ -35,109 +39,38 @@ test { pub fn fusedExperts( input: zml.Tensor, - ids: zml.Tensor, - weights: zml.Tensor, + routing: LocalRouting, gate_up: zml.nn.Linear, down: zml.nn.Linear, - options: zml.moe.Options, + options: Options, ) zml.Tensor { // HACK(Corentin): only rely on the scheme of gate_up for now. if (gate_up.quantization) |q| switch (q.scheme) { .mxfp4 => if (zml.Compiler.current().platform.target == .cuda) { - return fusedExpertsMxfp4(input, ids, weights, gate_up, down, options); + return fusedExpertsMxfp4(input, routing, gate_up, down, options); }, else => {}, }; - return fusedExpertsBase(input, ids, weights, gate_up, down, options); + const layout: ProjectionLayout = if (gate_up.quantizationScheme() == .mxfp4) .row_interleaved_1 else .concatenated; + return fusedExpertsBase(input, routing, gate_up, down, layout, options); } pub fn fusedExpertsBase( - input: zml.Tensor, - topk_ids: zml.Tensor, - topk_weights: zml.Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - opts: zml.moe.Options, -) zml.Tensor { - const gate_up_scheme: ?zml.Quantization.Scheme = if (gate_up.quantization) |q| q.scheme else null; - const layout: zml.moe.ExpertsLayout = if (gate_up_scheme == .mxfp4) - .{ .gate_up = .interleaved, .packing = .plain } - else - .{ .gate_up = .concatenated, .packing = .plain }; - - const args: FusedExpertsArgs = .{ - .hidden_states = input, - .gate_up = gate_up, - .down = down, - .topk_weights = topk_weights, - .topk_ids = topk_ids, - .activation = opts.activation, - .quantize_input = opts.quantize_input, - .gate_up_layout = layout.gate_up, - .routing_weight_placement = opts.routing_weight_placement, - }; - const expert_partition = gate_up.weight.shape().partition(.expert); - - if (!expert_partition.isSharded()) { - return fusedExpertsImpl(args); - } - - return zml.ops.manualComputation( - (struct { - args: FusedExpertsArgs, - global_num_experts: i64, - - fn call(self: @This(), _: zml.Shape) zml.Tensor { - const local_args = self.args; - const local_num_experts = local_args.gate_up.weight.dim(.expert); - const partition_id = zml.ops.partitionId().convert(.i32); - const expert_start = partition_id.scale(local_num_experts).convert(.i32); - const global_expert_ids = zml.Tensor.arange(.{ .end = self.global_num_experts }, .i32).withTags(.{.expert}); - - // Map global expert ids to local ids, or -1 for experts outside this partition. - const local_expert_mask = global_expert_ids.cmp(.GE, expert_start) - .logical(.AND, global_expert_ids.cmp(.LT, expert_start.addConstant(local_num_experts))); - var mapped_args = local_args; - mapped_args.expert_map = local_expert_mask.select( - global_expert_ids.sub(expert_start), - zml.Tensor.scalar(-1, .i32), - ); - - const local_output = fusedExpertsImpl(mapped_args); - const local_reshaped = local_output.reshape(local_args.hidden_states.shape()).withTags(.{ .t, .d }); - return zml.ops.allReduce(local_reshaped, zml.Tensor.add); - } - }).call, - .{ .args = args, .global_num_experts = gate_up.weight.dim(.expert) }, - input.shape(), - ); -} - -pub const FusedExpertsArgs = struct { hidden_states: Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - topk_weights: Tensor, - topk_ids: Tensor, - activation: zml.moe.Activation, - expert_map: ?Tensor = null, - /// Use FP8 activations for FP8 weights; false keeps BF16 activations. - quantize_input: bool, - gate_up_layout: ProjectionLayout, - routing_weight_placement: RoutingWeightPlacement, -}; - -pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { - const hidden_states = opts.hidden_states; - const topk_weights = opts.topk_weights; - const topk_ids = opts.topk_ids; + routing_: LocalRouting, + gate_up_: zml.nn.Linear, + down_: zml.nn.Linear, + layout: ProjectionLayout, + opts: Options, +) Tensor { + const expert_map_ = routing_.expertMap(); const num_tokens = hidden_states.dim(.t); var launch_config = launchConfigForTokens(num_tokens); - const gate_up_scheme = opts.gate_up.quantizationScheme(); - const down_scheme = opts.down.quantizationScheme(); + const gate_up_scheme = gate_up_.quantizationScheme(); + const down_scheme = down_.quantizationScheme(); if (gate_up_scheme) |scheme| switch (scheme) { .nvfp4 => stdx.debug.panic("triton backend doesn't support gate_up nvfp4 quantization scheme", .{}), .fp8_block128 => launch_config.block_size_k = 128, @@ -154,10 +87,10 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { }; const hidden = hidden_states.reshape(.{ .token = num_tokens, .in = hidden_states.dim(.d) }).withTags(.{ .token, .in }); - const gate_up = opts.gate_up.weight.withTags(.{ .expert, .out, .in }); - const down = opts.down.weight.withTags(.{ .expert, .out, .mid }); - const routing_weights = topk_weights.reshape(.{ .token = num_tokens, .in = topk_weights.dim(.topk) }).withTags(.{ .token, .topk }); - const ids = topk_ids.reshape(.{ .token = num_tokens, .in = topk_ids.dim(.topk) }).withTags(.{ .token, .topk }); + const gate_up = gate_up_.weight.withTags(.{ .expert, .out, .in }); + const down = down_.weight.withTags(.{ .expert, .out, .mid }); + const routing_weights = routing_.weights.reshape(.{ .token = num_tokens, .in = routing_.weights.dim(.topk) }).withTags(.{ .token, .topk }); + const ids = routing_.ids.reshape(.{ .token = num_tokens, .in = routing_.ids.dim(.topk) }).withTags(.{ .token, .topk }); stdx.debug.assert(hidden.dtype() == .bf16, "triton backend expected BF16 hidden states, got {}", .{hidden.dtype()}); if (gate_up_scheme == .mxfp4) { @@ -185,13 +118,13 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { stdx.debug.assert(ids.dim(.token) == hidden.dim(.token) and routing_weights.dim(.token) == hidden.dim(.token), "triton backend expected routing ids and weights to match hidden token count, got {}, {} and {}", .{ ids.dim(.token), routing_weights.dim(.token), hidden.dim(.token) }); stdx.debug.assert(ids.dim(.topk) == routing_weights.dim(.topk), "triton backend expected routing ids and weights to have matching top-k dimensions, got {} and {}", .{ ids.dim(.topk), routing_weights.dim(.topk) }); stdx.debug.assert(gate_up.dim(.expert) == down.dim(.expert), "triton backend expected gate/up and down expert counts to match, got {} and {}", .{ gate_up.dim(.expert), down.dim(.expert) }); - if (opts.expert_map) |expert_map| + if (expert_map_) |expert_map| stdx.debug.assert(expert_map.dtype() == .i32 and expert_map.rank() == 1, "triton backend expected a rank-1 i32 expert map, got rank {} and dtype {}", .{ expert_map.rank(), expert_map.dtype() }); - const num_experts = if (opts.expert_map) |expert_map| expert_map.dim(.expert) else gate_up.dim(.expert); + const num_experts = if (expert_map_) |expert_map| expert_map.dim(.expert) else gate_up.dim(.expert); const routing = prepareRouting(ids, num_experts, @intCast(launch_config.block_size_m)); - const expert_ids = if (opts.expert_map) |expert_map| + const expert_ids = if (expert_map_) |expert_map| expert_map.gather(.{ .expert = routing.expert_ids }, .{}).withTags(.{.g}) else routing.expert_ids; @@ -201,9 +134,9 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { const gate_up_out = gemm(.{ .input = hidden_quant, .weight = gate_up, - .bias = opts.gate_up.bias, + .bias = gate_up_.bias, .input_scale = input_scale, - .weight_scale = opts.gate_up.quantizationScales(), + .weight_scale = gate_up_.quantizationScales(), .routing = routing, .expert_ids = expert_ids, .launch_config = launch_config, @@ -212,7 +145,7 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { .output_shape = Shape.init(.{ .token = routing.num_assignments, .out = gate_up.dim(.out) }, .bf16), }); - var activated = zml.moe.applyActivation(gate_up_out, opts.activation, opts.gate_up_layout); + var activated = opts.activation.applyActivation(gate_up_out, layout); if (opts.routing_weight_placement == .before_down) { const weights = routing_weights.reshape(.{ .token = routing.num_assignments }).convert(.f32); activated = activated.mul(weights.broad(activated.shape())); @@ -222,9 +155,9 @@ pub fn fusedExpertsImpl(opts: FusedExpertsArgs) Tensor { const down_out = gemm(.{ .input = activated_quant, .weight = down, - .bias = opts.down.bias, + .bias = down_.bias, .input_scale = down_input_scale, - .weight_scale = opts.down.quantizationScales(), + .weight_scale = down_.quantizationScales(), .routing_weights = if (opts.routing_weight_placement == .after_down) routing_weights else null, .routing = routing, .expert_ids = expert_ids, @@ -561,14 +494,14 @@ test "SwiGLU uses FP32 math for concatenated and interleaved BF16 inputs" { const io = std.testing.io; const platform = zml.testing.env(); const Local = struct { - fn forward(x: Tensor, layout: ProjectionLayout, activation: zml.moe.Activation) Tensor { - return zml.moe.applyActivation(x, activation, layout); + fn forward(x: Tensor, layout: ProjectionLayout, activation: Activation) Tensor { + return activation.applyActivation(x, layout); } }; const x: Tensor = .init(.{ .token = 1, .out = 6 }, .bf16); const gates = [_]f32{ 0.75, -1.25, 3.5 }; const ups = [_]f32{ 0.875, -2.25, 4.5 }; - for ([_]ProjectionLayout{ .concatenated, .interleaved }) |layout| { + for ([_]ProjectionLayout{ .concatenated, .row_interleaved_1 }) |layout| { var values: [6]zml.floats.BFloat16 = undefined; for (gates, ups, 0..) |gate, up, i| { values[if (layout == .concatenated) i else 2 * i] = .fromF32(gate); @@ -577,7 +510,7 @@ test "SwiGLU uses FP32 math for concatenated and interleaved BF16 inputs" { var input = try zml.Buffer.fromBytes(io, platform, x.shape(), std.mem.asBytes(&values)); defer input.deinit(); for ([_]?f32{ null, 2 }) |threshold| { - const activation: zml.moe.Activation = .{ .swiglu = .{ .limit = threshold } }; + const activation: Activation = .{ .swiglu = .{ .limit = threshold } }; var exe = try platform.compileFn(allocator, io, Local.forward, .{ x, layout, activation }, .{}); defer exe.deinit(); try zml.testing.expectEqualShapes(Shape.init(.{ .token = 1, .out = 3 }, .f32), exe.output_shapes[0]); @@ -733,7 +666,8 @@ test "fused experts support BF16 and MXFP4 layouts, bias, and routing weights" { const columns = Tensor.arange(.{ .end = 256 }, .i32).withTags(.{.dout}); const is_gate = switch (layout) { .concatenated => columns.cmp(.LT, Tensor.scalar(128, .i32)), - .interleaved => columns.remainder(Tensor.scalar(2, .i32)).cmp(.EQ, Tensor.scalar(0, .i32)), + .row_interleaved_1 => columns.remainder(Tensor.scalar(2, .i32)).cmp(.EQ, Tensor.scalar(0, .i32)), + .row_interleaved_8 => columns.divByConst(8).remainder(Tensor.scalar(2, .i32)).cmp(.EQ, Tensor.scalar(0, .i32)), }; const dtype: DataType = if (fp4) .u8 else .bf16; // E2M1 codes 1, 2, 3 encode 0.5, 1, 1.5; each byte holds two values. @@ -765,13 +699,12 @@ test "fused experts support BF16 and MXFP4 layouts, bias, and routing weights" { } const route = Tensor.arange(.{ .end = 2 }, .i32).reshape(.{ .t = 1, .topk = 2 }) .broad(Shape.init(.{ .t = x.dim(.t), .topk = 2 }, .i32)); - return fusedExpertsImpl(.{ - .hidden_states = x, - .gate_up = gate_up, - .down = down, - .topk_ids = route.addConstant(2), - .topk_weights = route.convert(.f32).addConstant(1).scale(0.25), - .gate_up_layout = layout, + return fusedExpertsBase(x, .{ + .ids = route.addConstant(2), + .weights = route.convert(.f32).addConstant(1).scale(0.25), + .global_num_experts = 8, + .local_num_experts = 8, + }, gate_up, down, layout, .{ .routing_weight_placement = placement, .quantize_input = false, .activation = .{ .swiglu = .{} }, @@ -786,7 +719,7 @@ test "fused experts support BF16 and MXFP4 layouts, bias, and routing weights" { @memset(host.items(zml.floats.BFloat16), .fromF32(1.0 / 128.0)); var input = try zml.Buffer.fromSlice(io, platform, host); defer input.deinit(); - for ([_]ProjectionLayout{ .concatenated, .interleaved }) |layout| { + for ([_]ProjectionLayout{ .concatenated, .row_interleaved_1 }) |layout| { for ([_]RoutingWeightPlacement{ .before_down, .after_down }) |placement| { var exe = try platform.compileFn(allocator, io, Local.forward, .{ x, layout, placement, storage_dtype }, .{}); defer exe.deinit(); @@ -866,58 +799,13 @@ pub fn isAvailable(platform: *const zml.Platform) bool { return cc.major == 10; } -pub fn validateOptions(opts: zml.moe.Options) void { +pub fn validateOptions(opts: Options) void { stdx.debug.assert(opts.activation == .swiglu, "triton backend only accepts swiglu activation, got {}", .{opts.activation}); stdx.debug.assert(opts.activation.swiglu.limit != null, "triton backend requires swiglu limit to be set", .{}); stdx.debug.assert(opts.activation.swiglu.bias == null, "triton backend requires swiglu bias to be null", .{}); stdx.debug.assert(opts.activation.swiglu.scale == null, "triton backend requires swiglu scale to be null", .{}); } -/// Row-major MXFP4 E2M1 weights and linear E8M0 block32 scales. -/// BF16 activations are quantized on the GPU to FP8 with per-32 E8M0 scales. -pub fn fusedExpertsMxfp4( - input: zml.Tensor, - ids: zml.Tensor, - weights: zml.Tensor, - gate_up: zml.nn.Linear, - down: zml.nn.Linear, - options: zml.moe.Options, -) zml.Tensor { - validateOptions(options); - stdx.debug.assert(input.dtype() == .bf16, "triton backend only supports bf16 inputs, got {}", .{input.dtype()}); - stdx.debug.assert(gate_up.bias == null and down.bias == null, "triton backend expects gate_up bias and down bias to be null", .{}); - - const gq = gate_up.quantization orelse @panic("triton backend requires gate_up quantization to be set"); - const dq = down.quantization orelse @panic("triton backend requires down quantization to be set"); - stdx.debug.assert(gq.scheme == .mxfp4 and dq.scheme == .mxfp4, "triton expects gate_up and down quantization scheme to be mxfp4, got {} and {}", .{ gq.scheme, dq.scheme }); - - // Weight storage for mxfp4 in HF is expressed as u8 or i8 - stdx.debug.assert(gate_up.weight.dtype() == .u8 or gate_up.weight.dtype() == .i8, "triton expects gate_up weight dtype to be u8 or i8, got {}", .{gate_up.weight.dtype()}); - stdx.debug.assert(down.weight.dtype() == .u8 or down.weight.dtype() == .i8, "triton expects down weight dtype to be u8 or i8, got {}", .{gate_up.weight.dtype()}); - - const expert_parallelism = gate_up.weight.shape().partition(.expert).isSharded(); - - const context: Context = .{ - .input = input, - .ids = ids, - .weights = weights, - .w1 = gate_up.weight.bitCast(.u8), - .s1 = gq.scales, - .w2 = down.weight.bitCast(.u8), - .s2 = dq.scales, - .global_experts = gate_up.weight.dim(.expert), - .topk = ids.dim(.topk), - .limit = options.activation.swiglu.limit.?, - .routing_weight_placement = options.routing_weight_placement, - .expert_parallel = expert_parallelism, - }; - - return if (expert_parallelism) - zml.ops.manualComputation(Context.body, context, input.shape()) - else - context.body(input.shape()); -} - pub const Config = struct { tokens: i64, hidden: i64, @@ -952,210 +840,233 @@ pub const Config = struct { } }; -pub fn validateInputs( +fn validateInputs( c: Config, input: zml.Tensor, ids: zml.Tensor, - weights: zml.Tensor, - w1: zml.Tensor, - s1: zml.Tensor, - w2: zml.Tensor, - s2: zml.Tensor, -) !void { - if (s1.rank() != 3 or s2.rank() != 3) return error.InvalidInputShape; - if (s1.dim(2) < @divTrunc(c.hidden, 32) or - s2.dim(2) < @divTrunc(c.intermediate, 32)) return error.InvalidInputShape; + expert_weights: zml.Tensor, + gate_up: zml.Tensor, + gate_up_scales: zml.Tensor, + down: zml.Tensor, + down_scales: zml.Tensor, +) void { + stdx.debug.assert(gate_up_scales.rank() == 3 and down_scales.rank() == 3, "Invalid MXFP4 tensor layout, expected gate_up_scales and down_scales to be rank 3, got {f} and {f}", .{ gate_up_scales, down_scales }); + const shapes = .{ - .x = zml.Shape.init(.{ c.tokens, c.hidden }, .bf16), + .input = zml.Shape.init(.{ c.tokens, c.hidden }, .bf16), .ids = zml.Shape.init(.{ c.tokens, c.topk }, .i32), - .scales = zml.Shape.init(.{ c.tokens, c.topk }, .f32), - .w1 = zml.Shape.init(.{ c.experts, 2 * c.intermediate, @divTrunc(c.hidden, 2) }, .u8), - .s1 = zml.Shape.init(.{ c.experts, 2 * c.intermediate, s1.dim(2) }, .f8e8m0), - .w2 = zml.Shape.init(.{ c.experts, c.hidden, @divTrunc(c.intermediate, 2) }, .u8), - .s2 = zml.Shape.init(.{ c.experts, c.hidden, s2.dim(2) }, .f8e8m0), + .expert_weights = zml.Shape.init(.{ c.tokens, c.topk }, .f32), + .gate_up = zml.Shape.init(.{ c.experts, 2 * c.intermediate, @divExact(c.hidden, 2) }, .i8), + .gate_up_scales = zml.Shape.init(.{ c.experts, 2 * c.intermediate, gate_up_scales.dim(2) }, .f8e8m0), + .down = zml.Shape.init(.{ c.experts, c.hidden, @divExact(c.intermediate, 2) }, .i8), + .down_scales = zml.Shape.init(.{ c.experts, c.hidden, down_scales.dim(2) }, .f8e8m0), }; - if (!input.shape().eql(shapes.x)) return error.InvalidInputShape; - if (!ids.shape().eql(shapes.ids)) return error.InvalidInputShape; - if (!weights.shape().eql(shapes.scales)) return error.InvalidInputShape; - if (!w1.shape().eql(shapes.w1)) return error.InvalidInputShape; - if (!s1.shape().eql(shapes.s1)) return error.InvalidInputShape; - if (!w2.shape().eql(shapes.w2)) return error.InvalidInputShape; - if (!s2.shape().eql(shapes.s2)) return error.InvalidInputShape; + stdx.debug.assert(input.shape().eql(shapes.input) and + ids.shape().eql(shapes.ids) and + expert_weights.shape().eql(shapes.expert_weights) and + // Note: we accept both i8 and u8 for gate_up and down + // TODO: we should instead use f4e2m1 + (gate_up.shape().eql(shapes.gate_up) or gate_up.shape().eql(shapes.gate_up.withDtype(.u8))) and + gate_up_scales.shape().eql(shapes.gate_up_scales) and + (down.shape().eql(shapes.down) or down.shape().eql(shapes.down.withDtype(.u8))) and + down_scales.shape().eql(shapes.down_scales), + \\Invalid MXFP4 tensor layout: + \\ - expected input={f}, got {f} + \\ - expected ids={f}, got {f} + \\ - expected expert_weights={f}, got {f} + \\ - expected gate_up={f}, got {f} + \\ - expected gate_up_scales={f}, got {f} + \\ - expected down={f}, got {f} + \\ - expected down_scales={f}, got {f} + , .{ + shapes.input, input, // + shapes.ids, ids, // + shapes.expert_weights, expert_weights, // + shapes.gate_up, gate_up, // + shapes.gate_up_scales, gate_up_scales, // + shapes.down, down, // + shapes.down_scales, down_scales, // + }); + + stdx.debug.assert( + gate_up_scales.dim(2) >= @divTrunc(c.hidden, 32) and + down_scales.dim(2) >= @divTrunc(c.intermediate, 32), + "Invalid MXFP4 tensor layout, not enough scales, expected a ratio of at most 32 for config {}, got {f} and {f}", + .{ c, gate_up_scales, down_scales }, + ); } -const Context = struct { +/// Row-major MXFP4 E2M1 weights and linear E8M0 block32 scales. +/// BF16 activations are quantized on the GPU to FP8 with per-32 E8M0 scales. +pub fn fusedExpertsMxfp4( input: zml.Tensor, - ids: zml.Tensor, - weights: zml.Tensor, - w1: zml.Tensor, - s1: zml.Tensor, - w2: zml.Tensor, - s2: zml.Tensor, - global_experts: i64, - topk: i64, - limit: f32, - routing_weight_placement: RoutingWeightPlacement, - expert_parallel: bool, + routing: LocalRouting, + gate_up: zml.nn.Linear, + down: zml.nn.Linear, + options: Options, +) zml.Tensor { + validateOptions(options); + stdx.debug.assert(input.dtype() == .bf16, "triton backend only supports bf16 inputs, got {}", .{input.dtype()}); + stdx.debug.assert(gate_up.bias == null and down.bias == null, "triton backend expects gate_up bias and down bias to be null", .{}); - fn body(self: Context, _: zml.Shape) zml.Tensor { - const experts = self.w1.dim(.expert); - const hidden = self.w2.dim(1); - const intermediate = self.w2.dim(2) * 2; - var ids = self.ids.convert(.i32).reshape(.{ .token = .auto, .topk = self.topk }); - const tokens = ids.dim(.token); - if (self.expert_parallel) { - const partition_id = zml.ops.partitionId().convert(.i32); - ids = ids.sub(partition_id.scale(experts)); - } - const x = self.input.reshape(.{ tokens, hidden }); - const scales = self.weights.convert(.f32).reshape(.{ tokens, self.topk }); + const gq = gate_up.quantization orelse @panic("triton backend requires gate_up quantization to be set"); + const dq = down.quantization orelse @panic("triton backend requires down quantization to be set"); + stdx.debug.assert(gq.scheme == .mxfp4 and dq.scheme == .mxfp4, "triton expects gate_up and down quantization scheme to be mxfp4, got {} and {}", .{ gq.scheme, dq.scheme }); - const cfg: Config = .{ - .tokens = tokens, - .hidden = hidden, - .intermediate = intermediate, - .experts = experts, - .global_experts = self.global_experts, - .topk = self.topk, - .swiglu_limit = self.limit, - .routing_weight_placement = self.routing_weight_placement, - }; + // Weight storage for mxfp4 in HF is expressed as u8 or i8 + stdx.debug.assert(gate_up.weight.dtype() == .u8 or gate_up.weight.dtype() == .i8, "triton expects gate_up weight dtype to be u8 or i8, got {}", .{gate_up.weight.dtype()}); + stdx.debug.assert(down.weight.dtype() == .u8 or down.weight.dtype() == .i8, "triton expects down weight dtype to be u8 or i8, got {}", .{gate_up.weight.dtype()}); - cfg.validate() catch @panic("Invalid MXFP4 MoE configuration"); - validateInputs(cfg, x, ids, scales, self.w1, self.s1, self.w2, self.s2) catch @panic("Invalid MXFP4 tensor layout"); + const topk = routing.ids.dim(.topk); + const experts = gate_up.weight.dim(.expert); + const hidden = down.weight.dim(1); + const intermediate = down.weight.dim(2) * 2; + const ids = routing.localIds().convert(.i32).reshape(.{ .token = .auto, .topk = topk }); + const tokens = ids.dim(.token); - const bn = 16; - const bk = 512; + const x = input.reshape(.{ tokens, hidden }); + const scales = routing.weights.convert(.f32).reshape(.{ tokens, topk }); - const tile_count = kernels.tiles(cfg.tokens, cfg.topk, cfg.experts, bn); + const cfg: Config = .{ + .tokens = tokens, + .hidden = hidden, + .intermediate = intermediate, + .experts = experts, + .global_experts = routing.global_num_experts, + .topk = topk, + .swiglu_limit = options.activation.swiglu.limit.?, + .routing_weight_placement = options.routing_weight_placement, + }; - // Quantize activations - const q = kernels.Quant.Kernel.call( - .{ .x = x }, + cfg.validate() catch @panic("Invalid MXFP4 MoE configuration"); + validateInputs(cfg, x, ids, scales, gate_up.weight, gq.scales, down.weight, dq.scales); + + const bn = 16; + const bk = 512; + + const tile_count = kernels.tiles(cfg.tokens, cfg.topk, cfg.experts, bn); + + // Quantize activations + const q = kernels.Quant.Kernel.call( + .{ .x = x }, + .{ + .q = .init(.{ cfg.tokens, cfg.hidden }, .f8e4m3fn), + .s = .init(.{ cfg.tokens, @divTrunc(cfg.hidden, 32) }, .u8), + }, + .{ + .cfg = .{ .tokens = cfg.tokens, .hidden = cfg.hidden }, + .grid = .{ @intCast(@divTrunc(cfg.tokens + kernels.quantRows(cfg.tokens) - 1, kernels.quantRows(cfg.tokens))), @intCast(@divTrunc(cfg.hidden, 128)), 1 }, + .num_warps = 4, + .num_stages = 1, + }, + ); + + const dummy = zml.Tensor.scalar(0, .i32); + var map = dummy; + var perm = dummy; + var sched = dummy; + + if (cfg.tokens != 1) { + const r = kernels.Route.Kernel.call( + .{ .ids = ids }, + .{ .counts = .init(.{cfg.experts}, .i32), .pos = .init(.{cfg.tokens * cfg.topk}, .i32) }, + .{ .cfg = .{ + .tokens = cfg.tokens, + .topk = cfg.topk, + .expert_offset = cfg.expert_offset, + }, .grid = .{ @intCast(cfg.experts), 1, 1 }, .num_warps = 4, .num_stages = 1 }, + ); + + const s = kernels.Schedule.Kernel.call( + .{ .ids = ids, .counts = r.counts, .pos = r.pos }, + .{ .map = .init(.{tile_count * bn}, .i32), .perm = .init(.{cfg.tokens * cfg.topk}, .i32), .sched = .init(.{2 * tile_count + 1}, .i32) }, .{ - .q = .init(.{ cfg.tokens, cfg.hidden }, .f8e4m3fn), - .s = .init(.{ cfg.tokens, @divTrunc(cfg.hidden, 32) }, .u8), - }, - .{ - .cfg = .{ .tokens = cfg.tokens, .hidden = cfg.hidden }, - .grid = .{ @intCast(@divTrunc(cfg.tokens + kernels.quantRows(cfg.tokens) - 1, kernels.quantRows(cfg.tokens))), @intCast(@divTrunc(cfg.hidden, 128)), 1 }, + .cfg = .{ + .tokens = cfg.tokens, + .topk = cfg.topk, + .expert_offset = cfg.expert_offset, + .experts = cfg.experts, + .capacity = tile_count, + .group = bn, + }, + .grid = .{ @intCast(@divTrunc(cfg.tokens * cfg.topk + 127, 128)), 1, 1 }, .num_warps = 4, .num_stages = 1, }, ); - const dummy = zml.Tensor.scalar(0, .i32); - var map = dummy; - var perm = dummy; - var sched = dummy; - - if (cfg.tokens != 1) { - const r = kernels.Route.Kernel.call( - .{ .ids = ids }, - .{ .counts = .init(.{cfg.experts}, .i32), .pos = .init(.{cfg.tokens * cfg.topk}, .i32) }, - .{ .cfg = .{ - .tokens = cfg.tokens, - .topk = cfg.topk, - .expert_offset = cfg.expert_offset, - }, .grid = .{ @intCast(cfg.experts), 1, 1 }, .num_warps = 4, .num_stages = 1 }, - ); - - const s = kernels.Schedule.Kernel.call( - .{ .ids = ids, .counts = r.counts, .pos = r.pos }, - .{ .map = .init(.{tile_count * bn}, .i32), .perm = .init(.{cfg.tokens * cfg.topk}, .i32), .sched = .init(.{2 * tile_count + 1}, .i32) }, - .{ - .cfg = .{ - .tokens = cfg.tokens, - .topk = cfg.topk, - .expert_offset = cfg.expert_offset, - .experts = cfg.experts, - .capacity = tile_count, - .group = bn, - }, - .grid = .{ @intCast(@divTrunc(cfg.tokens * cfg.topk + 127, 128)), 1, 1 }, - .num_warps = 4, - .num_stages = 1, - }, - ); - - map = s.map; - perm = s.perm; - sched = s.sched; - } - - // Each narrow tile produces 64 activations: two complete scale groups. - const up_columns: i64 = if (cfg.tokens <= 8) 64 else 128; - const m = kernels.GateUp.Kernel.call( - .{ .q = q.q, .s = q.s, .w = self.w1, .ws = self.s1.bitCast(.u8), .ids = ids, .rw = scales, .map = map, .sched = sched }, - .{ .mid = .init(.{ tile_count * bn, cfg.intermediate }, .f8e4m3fn), .ms = .init(.{ tile_count * bn, @divTrunc(cfg.intermediate, 32) }, .u8) }, - .{ - .cfg = .{ - .tokens = cfg.tokens, - .hidden = cfg.hidden, - .intermediate = cfg.intermediate, - .experts = cfg.experts, - .topk = cfg.topk, - .expert_offset = cfg.expert_offset, - .swiglu_limit = cfg.swiglu_limit, - .scale_stride = self.s1.dim(2), - .routing_weight_placement = cfg.routing_weight_placement, - .bn = bn, - .bk = bk, - }, - .grid = .{ @intCast(@divTrunc(cfg.intermediate, up_columns)), @intCast(tile_count), 1 }, - .num_warps = if (cfg.tokens <= 8) 4 else 8, - .num_stages = 2, - .global_scratch_memory_size = @intCast(128 * @divTrunc(cfg.intermediate, up_columns) * tile_count), - }, - ); - - const d = kernels.Down.Kernel.call( - .{ .mid = m.mid, .ms = m.ms, .w = self.w2, .ws = self.s2.bitCast(.u8), .ids = ids, .sched = sched }, - .{ .d = .init(.{ tile_count * bn, cfg.hidden }, .bf16) }, - .{ - .cfg = .{ - .tokens = cfg.tokens, - .hidden = cfg.hidden, - .intermediate = cfg.intermediate, - .experts = cfg.experts, - .topk = cfg.topk, - .expert_offset = cfg.expert_offset, - .swiglu_limit = cfg.swiglu_limit, - .scale_stride = self.s2.dim(2), - .routing_weight_placement = cfg.routing_weight_placement, - .bn = bn, - .bk = bk, - }, - .grid = .{ @intCast(@divTrunc(cfg.hidden, 128)), @intCast(tile_count), 1 }, - .num_warps = if (cfg.tokens <= 8) 4 else 8, - .num_stages = 2, - .global_scratch_memory_size = @intCast(128 * @divTrunc(cfg.hidden, 128) * tile_count), - }, - ); - - const result = kernels.Combine.Kernel.call( - .{ .d = d.d, .ids = ids, .perm = perm, .rw = scales }, - .{ .y = x.shape() }, - .{ - .cfg = .{ - .tokens = cfg.tokens, - .hidden = cfg.hidden, - .topk = cfg.topk, - .expert_offset = cfg.expert_offset, - .experts = cfg.experts, - .routing_weight_placement = cfg.routing_weight_placement, - .bn = bn, - }, - .grid = .{ @intCast(cfg.tokens), @intCast(@divTrunc(cfg.hidden + kernels.combineBlock(cfg.tokens, cfg.hidden) - 1, kernels.combineBlock(cfg.tokens, cfg.hidden))), 1 }, - .num_warps = 4, - .num_stages = 1, - }, - ).y.reshape(self.input.shape().dims()).withTags(self.input.shape()); - - return if (self.expert_parallel) - zml.ops.allReduce(result, zml.Tensor.add) - else - result; + map = s.map; + perm = s.perm; + sched = s.sched; } -}; + + // Each narrow tile produces 64 activations: two complete scale groups. + const up_columns: i64 = if (cfg.tokens <= 8) 64 else 128; + const m = kernels.GateUp.Kernel.call( + .{ .q = q.q, .s = q.s, .w = gate_up.weight.bitCast(.u8), .ws = gq.scales.bitCast(.u8), .ids = ids, .rw = scales, .map = map, .sched = sched }, + .{ .mid = .init(.{ tile_count * bn, cfg.intermediate }, .f8e4m3fn), .ms = .init(.{ tile_count * bn, @divTrunc(cfg.intermediate, 32) }, .u8) }, + .{ + .cfg = .{ + .tokens = cfg.tokens, + .hidden = cfg.hidden, + .intermediate = cfg.intermediate, + .experts = cfg.experts, + .topk = cfg.topk, + .expert_offset = cfg.expert_offset, + .swiglu_limit = cfg.swiglu_limit, + .scale_stride = gq.scales.dim(2), + .routing_weight_placement = cfg.routing_weight_placement, + .bn = bn, + .bk = bk, + }, + .grid = .{ @intCast(@divTrunc(cfg.intermediate, up_columns)), @intCast(tile_count), 1 }, + .num_warps = if (cfg.tokens <= 8) 4 else 8, + .num_stages = 2, + .global_scratch_memory_size = @intCast(128 * @divTrunc(cfg.intermediate, up_columns) * tile_count), + }, + ); + + const d = kernels.Down.Kernel.call( + .{ .mid = m.mid, .ms = m.ms, .w = down.weight.bitCast(.u8), .ws = dq.scales.bitCast(.u8), .ids = ids, .sched = sched }, + .{ .d = .init(.{ tile_count * bn, cfg.hidden }, .bf16) }, + .{ + .cfg = .{ + .tokens = cfg.tokens, + .hidden = cfg.hidden, + .intermediate = cfg.intermediate, + .experts = cfg.experts, + .topk = cfg.topk, + .expert_offset = cfg.expert_offset, + .swiglu_limit = cfg.swiglu_limit, + .scale_stride = dq.scales.dim(2), + .routing_weight_placement = cfg.routing_weight_placement, + .bn = bn, + .bk = bk, + }, + .grid = .{ @intCast(@divTrunc(cfg.hidden, 128)), @intCast(tile_count), 1 }, + .num_warps = if (cfg.tokens <= 8) 4 else 8, + .num_stages = 2, + .global_scratch_memory_size = @intCast(128 * @divTrunc(cfg.hidden, 128) * tile_count), + }, + ); + + return kernels.Combine.Kernel.call( + .{ .d = d.d, .ids = ids, .perm = perm, .rw = scales }, + .{ .y = x.shape() }, + .{ + .cfg = .{ + .tokens = cfg.tokens, + .hidden = cfg.hidden, + .topk = cfg.topk, + .expert_offset = cfg.expert_offset, + .experts = cfg.experts, + .routing_weight_placement = cfg.routing_weight_placement, + .bn = bn, + }, + .grid = .{ @intCast(cfg.tokens), @intCast(@divTrunc(cfg.hidden + kernels.combineBlock(cfg.tokens, cfg.hidden) - 1, kernels.combineBlock(cfg.tokens, cfg.hidden))), 1 }, + .num_warps = 4, + .num_stages = 1, + }, + ).y.reshape(input.shape().dims()).withTags(input.shape()); +} diff --git a/zml/kernels/moe/weights.zig b/zml/kernels/moe/weights.zig new file mode 100644 index 00000000..0e288856 --- /dev/null +++ b/zml/kernels/moe/weights.zig @@ -0,0 +1,10 @@ +const nn = @import("../../nn.zig"); +const Tensor = @import("../../tensor.zig").Tensor; + +pub fn unpackedWeight(linear: nn.Linear) Tensor { + const quantization = linear.quantization orelse return linear.weight; + return if (nn.isPackedFp4(quantization.scheme, linear.weight.dtype())) + nn.unpackFp4(linear.weight, linear.tag, linear.tag) + else + linear.weight; +} diff --git a/zml/kernels/paged_attention/cute/benchmark.zig b/zml/kernels/paged_attention/cute/benchmark.zig index 9fd6a355..587af69f 100644 --- a/zml/kernels/paged_attention/cute/benchmark.zig +++ b/zml/kernels/paged_attention/cute/benchmark.zig @@ -54,7 +54,7 @@ fn runBenchmark(a: std.mem.Allocator, io: std.Io, platform: *const zml.Platform, inputs.sinks = inputs.sinks.withPartitioning(mesh, .{ ._0 = .model }); const compile_start: std.Io.Timestamp = .now(io, .awake); - const exe = try platform.compileFn(a, io, fixtures.forward, .{ inputs, global_case }, .{ + const exe = try platform.compileFn(a, io, forward, .{ inputs, global_case, mesh }, .{ .program_name = "paged_attention_harness", .meshes = &.{mesh}, }); @@ -115,6 +115,31 @@ fn runBenchmark(a: std.mem.Allocator, io: std.Io, platform: *const zml.Platform, }); } +fn forward(inputs: Inputs, c: cases.Case, mesh: *const zml.Mesh) zml.Tensor { + const Call = struct { + inputs: Inputs, + case: cases.Case, + + fn local(self: @This(), _: zml.Shape) zml.Tensor { + return fixtures.forward(self.inputs, self.case); + } + }; + const heads: zml.Sharding.Partitioning.Logical = .parse(.{ .hkv = .model }); + return zml.ops.manualComputation(Call.local, .{ .inputs = inputs, .case = c }, inputs.q.shape(), mesh, .{ + .in_specs = .{ .inputs = .{ + .q = heads, + .k = heads, + .v = heads, + .table = .replicated, + .lengths = .replicated, + .starts = .replicated, + .sinks = .parse(.{ ._0 = .model }), + .indices = .replicated, + } }, + .out_specs = heads, + }); +} + fn elapsedUs(start: std.Io.Timestamp, io: std.Io) f64 { return @as(f64, @floatFromInt(start.untilNow(io, .awake).toNanoseconds())) / 1000; } diff --git a/zml/kernels/paged_attention/cute/cute.zig b/zml/kernels/paged_attention/cute/cute.zig index e572cce1..ef78cc49 100644 --- a/zml/kernels/paged_attention/cute/cute.zig +++ b/zml/kernels/paged_attention/cute/cute.zig @@ -1,10 +1,10 @@ const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); -const paged = @import("../paged_attention.zig"); +pub const AttentionOptions = @import("../options.zig").AttentionOptions; +const kernel = @import("attention.zig"); const sm10x = @import("sm10x/sm10x.zig"); const sm120 = @import("sm120/sm120.zig"); -const kernel = @import("attention.zig"); pub const Options = struct { batch_size: usize, @@ -64,7 +64,7 @@ pub fn isAvailable(platform: *const zml.Platform) bool { return cc.major > 8 or (cc.major == 8 and cc.minor >= 9); } -pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, opts: paged.AttentionOptions) zml.Tensor { +pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, opts: AttentionOptions) zml.Tensor { stdx.debug.assert(isAvailable(zml.Compiler.current().platform), "CuTe paged attention requires SM89 or newer", .{}); stdx.debug.assert(q.dtype() == k.dtype() and q.dtype() == v.dtype() and (q.dtype() == .bf16 or q.dtype() == .f16), "CuTe paged attention requires matching BF16 or FP16 Q/K/V", .{}); stdx.debug.assert(q.dim(.hd) == 32 or q.dim(.hd) == 64 or q.dim(.hd) == 128 or q.dim(.hd) == 256 or q.dim(.hd) == 512, "CuTe paged attention supports head dimensions 32, 64, 128, 256 and 512", .{}); @@ -72,57 +72,44 @@ pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: z return sm10x.pagedAttention(parameters, q, k, v, opts); if (sm120.isAvailable(zml.Compiler.current().platform) and opts.mask != .indices) return sm120.pagedAttention(parameters, q, k, v, opts); - return zml.ops.manualComputation(Context.body, Context{ .q = q, .k = k, .v = v, .parameters = parameters, .opts = opts }, q.shape()); + const q_ = q.transpose(.{ .b, .hkv, .hg, .hd }); + const k_ = k.transpose(.{ .page, .k_chunk, .hkv, .hd }); + const v_ = v.transpose(.{ .page, .k_chunk, .hkv, .hd }); + stdx.debug.assert(q_.dim(.hkv) == k_.dim(.hkv) and k_.shape().eql(v_.shape()) and k_.dim(.hd) == q_.dim(.hd) and q_.dim(.hg) > 0, "CuTe paged attention requires aligned local Q/KV heads and matching caches", .{}); + const indices = switch (opts.mask) { + .indices => |indices| indices.transpose(.{ .b, .topk }), + else => zml.Tensor.zeroes(.init(.{ q_.dim(.b), 1 }, .i32)), + }; + const cfg: kernel.Config = .{ + .head_dim = q_.dim(.hd), + .tokens = q_.dim(.b), + .batch = parameters.block_table.dim(.b), + .kv_heads = k_.dim(.hkv), + .group = q_.dim(.hg), + .pages = k_.dim(.page), + .pages_per_seq = parameters.block_table.dim(.p), + .page_size = k_.dim(.k_chunk), + .dtype = zml.kernel.cute.from(q_.dtype()), + .causal = opts.mask.isCausal(), + .window = opts.mask.slidingWindow() orelse 0, + .scale = opts.scale orelse (1 / @sqrt(@as(f32, @floatFromInt(q_.dim(.hd))))), + .sink = opts.sink != null, + .indexed = opts.mask == .indices, + .topk = indices.dim(-1), + }; + const sink = if (opts.sink) |s| s.convert(.f32).reshape(.{cfg.kv_heads * cfg.group}) else zml.Tensor.zeroes(.init(.{cfg.kv_heads * cfg.group}, .f32)); + return kernel.Kernel.call(.{ + .q = q_, + .k = k_, + .v = v_, + .table = parameters.block_table, + .lengths = parameters.seq_lens, + .starts = parameters.query_start_len, + .sink = sink, + .indices = indices, + }, .{ .out = q_.shape() }, .{ + .cfg = cfg, + .grid = .{ @intCast(cfg.tokens), @intCast(cfg.kv_heads * cfg.group), 1 }, + .block = .{ 32, 1, 1 }, + }).out.transpose(q.shape()); } - -const Context = struct { - q: zml.Tensor, - k: zml.Tensor, - v: zml.Tensor, - parameters: Parameters, - opts: paged.AttentionOptions, - - fn body(self: Context, output_shape: zml.Shape) zml.Tensor { - const q = self.q.transpose(.{ .b, .hkv, .hg, .hd }); - const k = self.k.transpose(.{ .page, .k_chunk, .hkv, .hd }); - const v = self.v.transpose(.{ .page, .k_chunk, .hkv, .hd }); - stdx.debug.assert(q.dim(.hkv) == k.dim(.hkv) and k.shape().eql(v.shape()) and k.dim(.hd) == q.dim(.hd) and q.dim(.hg) > 0, "CuTe paged attention requires aligned local Q/KV heads and matching caches", .{}); - const p = self.parameters; - const indices = switch (self.opts.mask) { - .indices => |indices| indices.transpose(.{ .b, .topk }), - else => zml.Tensor.zeroes(.init(.{ q.dim(.b), 1 }, .i32)), - }; - const cfg: kernel.Config = .{ - .head_dim = q.dim(.hd), - .tokens = q.dim(.b), - .batch = p.block_table.dim(.b), - .kv_heads = k.dim(.hkv), - .group = q.dim(.hg), - .pages = k.dim(.page), - .pages_per_seq = p.block_table.dim(.p), - .page_size = k.dim(.k_chunk), - .dtype = zml.kernel.cute.from(q.dtype()), - .causal = self.opts.mask.isCausal(), - .window = self.opts.mask.slidingWindow() orelse 0, - .scale = self.opts.scale orelse (1 / @sqrt(@as(f32, @floatFromInt(q.dim(.hd))))), - .sink = self.opts.sink != null, - .indexed = self.opts.mask == .indices, - .topk = indices.dim(-1), - }; - const sink = if (self.opts.sink) |s| s.convert(.f32).reshape(.{cfg.kv_heads * cfg.group}) else zml.Tensor.zeroes(.init(.{cfg.kv_heads * cfg.group}, .f32)); - return kernel.Kernel.call(.{ - .q = q, - .k = k, - .v = v, - .table = p.block_table, - .lengths = p.seq_lens, - .starts = p.query_start_len, - .sink = sink, - .indices = indices, - }, .{ .out = q.shape() }, .{ - .cfg = cfg, - .grid = .{ @intCast(cfg.tokens), @intCast(cfg.kv_heads * cfg.group), 1 }, - .block = .{ 32, 1, 1 }, - }).out.transpose(output_shape); - } -}; diff --git a/zml/kernels/paged_attention/cute/fixtures.zig b/zml/kernels/paged_attention/cute/fixtures.zig index 8c289b04..c0815d1d 100644 --- a/zml/kernels/paged_attention/cute/fixtures.zig +++ b/zml/kernels/paged_attention/cute/fixtures.zig @@ -1,5 +1,7 @@ const std = @import("std"); + const zml = @import("zml"); + const cases = @import("cases.zig"); pub const Fixture = struct { @@ -122,7 +124,7 @@ pub fn forward(i: Inputs, c: cases.Case) zml.Tensor { }, }, }; - return zml.paged_attention.pagedAttention(parameters, i.q, i.k, i.v, .{ .split = .{ .k = i.k, .v = i.v } }, i.options(c)); + return zml.paged_attention.cute.pagedAttention(parameters.cute, i.q, i.k, i.v, i.options(c)); } pub fn upload(io: std.Io, platform: *const zml.Platform, inputs: Inputs, f: Fixture) !zml.Bufferized(Inputs) { diff --git a/zml/kernels/paged_attention/cute/sm10x/attention.zig b/zml/kernels/paged_attention/cute/sm10x/attention.zig index 89761bdd..12d6273c 100644 --- a/zml/kernels/paged_attention/cute/sm10x/attention.zig +++ b/zml/kernels/paged_attention/cute/sm10x/attention.zig @@ -317,7 +317,7 @@ fn produce(t: Tile) void { const id = tiles.iv.mul(c.features() * 2).add(stage); var reuse = b.openIf(id.ge(c.buffers())); { - b.mbarrierWait(t.released(id), t.phase(id).bitXor(1)); + b.mbarrierWait(t.released(id), t.phase(id).bitXor(1), .{ .scope = .cta }); reuse.close(); } const source = if (stage < c.features()) t.k else t.v; @@ -339,7 +339,7 @@ fn multiply(t: Tile, tile: V, comptime qk: bool) void { { for (0..@as(usize, @intCast(c.features()))) |part| { const id = tile.mul(c.features() * 2).add(@as(i64, @intCast(part)) + if (qk) @as(i64, 0) else c.features()); - b.mbarrierWait(t.ready(id), t.phase(id)); + b.mbarrierWait(t.ready(id), t.phase(id), .{ .scope = .cta }); b.tcgen05Fence(.after); const kv = descriptor(b, t.shared.add(id.rem(c.buffers()).mul(c.kvBytes())), if (qk) 16 else c.keys() * 128); const qp = descriptor(b, t.shared.add(if (qk) c.qBase() else c.pBase()), 16); @@ -353,22 +353,22 @@ fn multiply(t: Tile, tile: V, comptime qk: bool) void { const accumulate = if (qk) b.constant(.i32, @intFromBool(part > 0 or step > 0)).ne(0) else tile.gt(0).bitOr(b.constant(.i32, @intFromBool(step > 0)).ne(0)); b.tcgen05MmaF16(t.tmem.add(column), left, right, b.constant(.i32, @as(i32, @bitCast(instruction(c, qk)))), accumulate); } - b.tcgen05Commit(t.released(id)); + b.tcgen05Commit(t.released(id), .{}); } - b.tcgen05Commit(t.shared.add(c.completion())); + b.tcgen05Commit(t.shared.add(c.completion()), .{}); leader.close(); } - b.mbarrierWait(t.shared.add(c.completion()), b.constant(.i32, if (qk) 0 else 1)); + b.mbarrierWait(t.shared.add(c.completion()), b.constant(.i32, if (qk) 0 else 1), .{ .scope = .cta }); b.tcgen05Fence(.after); // A fast warp must not wrap completion parity before every consumer has // observed the preceding PV completion. - if (!qk) b.barrierSync(1, 128); + if (!qk) b.barrierCtaSyncAligned(1, 128); } fn publishProbabilities(t: Tile) void { t.b.fenceProxyAsyncShared(); t.b.tcgen05Fence(.before); - t.b.barrierSync(1, 128); + t.b.barrierCtaSyncAligned(1, 128); t.b.tcgen05Fence(.after); } @@ -393,7 +393,7 @@ fn reduceHeads(t: Tile, comptime heads: usize, values: [heads]V, comptime maximu leader.close(); } } - b.barrierSync(1, 128); + b.barrierCtaSyncAligned(1, 128); var result: [heads]V = undefined; for (0..heads) |h| { var value = b.loadWords(1, t.shared.add(t.c.reductions() + @as(i64, @intCast(h * 4))), .shared, .{})[0].bitCast(.f32); @@ -403,7 +403,7 @@ fn reduceHeads(t: Tile, comptime heads: usize, values: [heads]V, comptime maximu } result[h] = value; } - b.barrierSync(1, 128); + b.barrierCtaSyncAligned(1, 128); return result; } @@ -514,7 +514,7 @@ fn prefill(t: Tile) void { preserve.yieldElse(@as([64]V, @splat(zero))); } @memcpy(&saved, &preserve.results); - b.barrierSync(1, 128); + b.barrierCtaSyncAligned(1, 128); } multiply(t, tiles.iv, true); var scores: [64]V = undefined; diff --git a/zml/kernels/paged_attention/cute/sm10x/sm10x.zig b/zml/kernels/paged_attention/cute/sm10x/sm10x.zig index 664720ee..afc4298e 100644 --- a/zml/kernels/paged_attention/cute/sm10x/sm10x.zig +++ b/zml/kernels/paged_attention/cute/sm10x/sm10x.zig @@ -2,7 +2,6 @@ const std = @import("std"); const stdx = @import("stdx"); const zml = @import("../../../../zml.zig"); -const paged = @import("../../paged_attention.zig"); const cute = @import("../cute.zig"); const kernel = @import("attention.zig"); @@ -19,61 +18,48 @@ pub fn isAvailable(platform: *const zml.Platform) bool { return cc.major == 10; } -pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, opts: paged.AttentionOptions) zml.Tensor { - return zml.ops.manualComputation(Context.body, Context{ .q = q, .k = k, .v = v, .parameters = parameters, .opts = opts }, q.shape()); +pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, opts: cute.AttentionOptions) zml.Tensor { + const q_ = q.transpose(.{ .b, .hkv, .hg, .hd }); + const k_ = k.transpose(.{ .page, .k_chunk, .hkv, .hd }); + const v_ = v.transpose(.{ .page, .k_chunk, .hkv, .hd }); + stdx.debug.assert(q_.dim(.hkv) == k_.dim(.hkv) and k_.shape().eql(v_.shape()) and k_.dim(.hd) == q_.dim(.hd) and q_.dim(.hg) > 0, "CuTe paged attention requires aligned local Q/KV heads and matching caches", .{}); + const decode = !parameters.options_.is_prefill; + const cc = zml.platform.cuda.computeCapability(zml.Compiler.current().platform).?; + var c: kernel.Config = .{ + .architecture = if (cc.minor >= 3) .sm103 else .sm100, + .head_dim = q_.dim(.hd), + .tokens = q_.dim(.b), + .batch = parameters.block_table.dim(.b), + .kv_heads = k_.dim(.hkv), + .group = q_.dim(.hg), + .pages_per_seq = parameters.block_table.dim(.p), + .page_size = k_.dim(.k_chunk), + .max_query = if (decode) 1 else @min(q_.dim(.b), @as(i64, @intCast(parameters.options_.max_seqlen_q))), + .dtype = zml.kernel.cute.from(q_.dtype()), + .decode = decode, + .causal = opts.mask.isCausal(), + .window = opts.mask.slidingWindow() orelse 0, + .scale = opts.scale orelse 1 / @sqrt(@as(f32, @floatFromInt(q_.dim(.hd)))), + .sink = opts.sink != null, + }; + const query_tiles = if (decode) c.batch * c.queryTiles() else c.batch + @divTrunc(c.tokens * c.group + c.rows() - 1, c.rows()); + const work = query_tiles * c.kv_heads; + const bound = if (c.window > 0 and c.causal) @min(c.window + @divTrunc(c.rows() + c.group - 1, c.group), c.pages_per_seq * c.page_size) else c.pages_per_seq * c.page_size; + c.splits = @max(1, @min(32, @min(@divTrunc(@as(i64, @intCast(cute.multiprocessorCount(zml.Compiler.current().platform))) * 2 + work - 1, work), @divTrunc(bound + 2 * c.keys() - 1, 2 * c.keys())))); + // An unused scalar stays a constant buffer; broadcasting a dummy for + // every head would add a helper kernel to each attention invocation. + const sink = if (opts.sink) |s| s.convert(.f32).reshape(.{c.kv_heads * c.group}) else zml.Tensor.scalar(0, .f32).reshape(.{1}); + const out = kernel.Attention.call(.{ .q = q_, .k = k_, .v = v_, .table = parameters.block_table, .lengths = parameters.seq_lens, .starts = parameters.query_start_len, .sink = sink }, .{ + .out = if (c.splits > 1) zml.Shape.init(.{ c.splits, c.tokens, c.kv_heads * c.group, c.head_dim + 2 }, .f32) else q_.shape(), + }, .{ .cfg = c, .grid = .{ @intCast(c.batch * c.queryTiles()), @intCast(c.kv_heads), @intCast(c.splits) }, .block = kernel.block }).out; + const result = if (c.splits > 1) kernel.Merge.call(.{ .partials = out, .lengths = parameters.seq_lens, .starts = parameters.query_start_len, .sink = sink }, .{ .out = q_.shape() }, .{ + .cfg = c, + .grid = .{ @intCast(c.tokens), @intCast(c.kv_heads * c.group), 1 }, + .block = .{ 128, 1, 1 }, + }).out else out; + return result.transpose(q.shape()); } -const Context = struct { - q: zml.Tensor, - k: zml.Tensor, - v: zml.Tensor, - parameters: Parameters, - opts: paged.AttentionOptions, - - fn body(self: Context, output_shape: zml.Shape) zml.Tensor { - const q = self.q.transpose(.{ .b, .hkv, .hg, .hd }); - const k = self.k.transpose(.{ .page, .k_chunk, .hkv, .hd }); - const v = self.v.transpose(.{ .page, .k_chunk, .hkv, .hd }); - stdx.debug.assert(q.dim(.hkv) == k.dim(.hkv) and k.shape().eql(v.shape()) and k.dim(.hd) == q.dim(.hd) and q.dim(.hg) > 0, "CuTe paged attention requires aligned local Q/KV heads and matching caches", .{}); - const p = self.parameters; - const decode = !p.options_.is_prefill; - const cc = zml.platform.cuda.computeCapability(zml.Compiler.current().platform).?; - var c: kernel.Config = .{ - .architecture = if (cc.minor >= 3) .sm103 else .sm100, - .head_dim = q.dim(.hd), - .tokens = q.dim(.b), - .batch = p.block_table.dim(.b), - .kv_heads = k.dim(.hkv), - .group = q.dim(.hg), - .pages_per_seq = p.block_table.dim(.p), - .page_size = k.dim(.k_chunk), - .max_query = if (decode) 1 else @min(q.dim(.b), @as(i64, @intCast(p.options_.max_seqlen_q))), - .dtype = zml.kernel.cute.from(q.dtype()), - .decode = decode, - .causal = self.opts.mask.isCausal(), - .window = self.opts.mask.slidingWindow() orelse 0, - .scale = self.opts.scale orelse 1 / @sqrt(@as(f32, @floatFromInt(q.dim(.hd)))), - .sink = self.opts.sink != null, - }; - const query_tiles = if (decode) c.batch * c.queryTiles() else c.batch + @divTrunc(c.tokens * c.group + c.rows() - 1, c.rows()); - const work = query_tiles * c.kv_heads; - const bound = if (c.window > 0 and c.causal) @min(c.window + @divTrunc(c.rows() + c.group - 1, c.group), c.pages_per_seq * c.page_size) else c.pages_per_seq * c.page_size; - c.splits = @max(1, @min(32, @min(@divTrunc(@as(i64, @intCast(cute.multiprocessorCount(zml.Compiler.current().platform))) * 2 + work - 1, work), @divTrunc(bound + 2 * c.keys() - 1, 2 * c.keys())))); - // An unused scalar stays a constant buffer; broadcasting a dummy for - // every head would add a helper kernel to each attention invocation. - const sink = if (self.opts.sink) |s| s.convert(.f32).reshape(.{c.kv_heads * c.group}) else zml.Tensor.scalar(0, .f32).reshape(.{1}); - const out = kernel.Attention.call(.{ .q = q, .k = k, .v = v, .table = p.block_table, .lengths = p.seq_lens, .starts = p.query_start_len, .sink = sink }, .{ - .out = if (c.splits > 1) zml.Shape.init(.{ c.splits, c.tokens, c.kv_heads * c.group, c.head_dim + 2 }, .f32) else q.shape(), - }, .{ .cfg = c, .grid = .{ @intCast(c.batch * c.queryTiles()), @intCast(c.kv_heads), @intCast(c.splits) }, .block = kernel.block }).out; - const result = if (c.splits > 1) kernel.Merge.call(.{ .partials = out, .lengths = p.seq_lens, .starts = p.query_start_len, .sink = sink }, .{ .out = q.shape() }, .{ - .cfg = c, - .grid = .{ @intCast(c.tokens), @intCast(c.kv_heads * c.group), 1 }, - .block = .{ 128, 1, 1 }, - }).out else out; - return result.transpose(output_shape); - } -}; - const ArchitectureCheck = struct { fn forward(q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, c: kernel.Config) zml.Tensor { const table_values: [9]i32 = .{ 4, 2, 8, 1, 6, 0, 7, 3, 5 }; diff --git a/zml/kernels/paged_attention/cute/sm120/sm120.zig b/zml/kernels/paged_attention/cute/sm120/sm120.zig index 6dfeb7e2..6645abf1 100644 --- a/zml/kernels/paged_attention/cute/sm120/sm120.zig +++ b/zml/kernels/paged_attention/cute/sm120/sm120.zig @@ -1,7 +1,6 @@ const stdx = @import("stdx"); const zml = @import("../../../../zml.zig"); -const paged = @import("../../paged_attention.zig"); const cute = @import("../cute.zig"); const kernel = @import("attention.zig"); @@ -18,66 +17,53 @@ pub fn isAvailable(platform: *const zml.Platform) bool { return cc.major == 12; } -pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, opts: paged.AttentionOptions) zml.Tensor { - return zml.ops.manualComputation(Context.body, Context{ .q = q, .k = k, .v = v, .parameters = parameters, .opts = opts }, q.shape()); +pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, opts: cute.AttentionOptions) zml.Tensor { + const q_ = q.transpose(.{ .b, .hkv, .hg, .hd }); + const k_ = k.transpose(.{ .page, .k_chunk, .hkv, .hd }); + const v_ = v.transpose(.{ .page, .k_chunk, .hkv, .hd }); + stdx.debug.assert(q_.dim(.hkv) == k_.dim(.hkv) and k_.shape().eql(v_.shape()) and k_.dim(.hd) == q_.dim(.hd) and q_.dim(.hg) > 0, "CuTe paged attention requires aligned local Q/KV heads and matching caches", .{}); + const decode = !parameters.options_.is_prefill; + var c: kernel.Config = .{ + .head_dim = q_.dim(.hd), + .tokens = q_.dim(.b), + .batch = parameters.block_table.dim(.b), + .kv_heads = k_.dim(.hkv), + .group = q_.dim(.hg), + .pages_per_seq = parameters.block_table.dim(.p), + .page_size = k_.dim(.k_chunk), + .max_query = if (decode) 1 else @min(q_.dim(.b), @as(i64, @intCast(parameters.options_.max_seqlen_q))), + .dtype = zml.kernel.cute.from(q_.dtype()), + .warps_per_query = if (decode) 4 else 2, + .kv_tile = if (q_.dim(.hd) == 512) 32 else 64, + .causal = opts.mask.isCausal(), + .window = opts.mask.slidingWindow() orelse 0, + .scale = opts.scale orelse 1 / @sqrt(@as(f32, @floatFromInt(q_.dim(.hd)))), + .sink = opts.sink != null, + }; + // Mixed batches also contain single-token rows. Partition their KV + // range so one long decode cannot hold up all the prefill CTAs. + const query_tiles = if (decode) c.batch * c.queryTiles() else c.batch + @divTrunc(c.tokens * c.group + c.rows() - 1, c.rows()); + const work = query_tiles * c.kv_heads; + const bound = if (c.window > 0) @min(c.window, c.pages_per_seq * c.page_size) else c.pages_per_seq * c.page_size; + c.splits = @max(1, @min(32, @min(@divTrunc(@as(i64, @intCast(cute.multiprocessorCount(zml.Compiler.current().platform))) * 2 + work - 1, work), @divTrunc(bound + 127, 128)))); + const sink = if (opts.sink) |s| s.convert(.f32).reshape(.{c.kv_heads * c.group}) else zml.Tensor.zeroes(.init(.{c.kv_heads * c.group}, .f32)); + const out = kernel.Kernel.call(.{ + .q = q_, + .k = k_, + .v = v_, + .table = parameters.block_table, + .lengths = parameters.seq_lens, + .starts = parameters.query_start_len, + .sink = sink, + }, .{ .out = if (c.splits > 1) zml.Shape.init(.{ c.splits, c.tokens, c.kv_heads * c.group, c.head_dim + 2 }, .f32) else q_.shape() }, .{ + .cfg = c, + .grid = .{ @intCast(c.batch * c.queryTiles()), @intCast(c.kv_heads), @intCast(c.splits) }, + .block = .{ 128, 1, 1 }, + }).out; + const result = if (c.splits > 1) kernel.Merge.call(.{ .partials = out, .starts = parameters.query_start_len, .sink = sink }, .{ .out = q_.shape() }, .{ + .cfg = c, + .grid = .{ @intCast(c.tokens), @intCast(c.kv_heads * c.group), 1 }, + .block = .{ 128, 1, 1 }, + }).out else out; + return result.transpose(q.shape()); } - -const Context = struct { - q: zml.Tensor, - k: zml.Tensor, - v: zml.Tensor, - parameters: Parameters, - opts: paged.AttentionOptions, - - fn body(self: Context, output_shape: zml.Shape) zml.Tensor { - const q = self.q.transpose(.{ .b, .hkv, .hg, .hd }); - const k = self.k.transpose(.{ .page, .k_chunk, .hkv, .hd }); - const v = self.v.transpose(.{ .page, .k_chunk, .hkv, .hd }); - stdx.debug.assert(q.dim(.hkv) == k.dim(.hkv) and k.shape().eql(v.shape()) and k.dim(.hd) == q.dim(.hd) and q.dim(.hg) > 0, "CuTe paged attention requires aligned local Q/KV heads and matching caches", .{}); - const p = self.parameters; - const decode = !p.options_.is_prefill; - var c: kernel.Config = .{ - .head_dim = q.dim(.hd), - .tokens = q.dim(.b), - .batch = p.block_table.dim(.b), - .kv_heads = k.dim(.hkv), - .group = q.dim(.hg), - .pages_per_seq = p.block_table.dim(.p), - .page_size = k.dim(.k_chunk), - .max_query = if (decode) 1 else @min(q.dim(.b), @as(i64, @intCast(p.options_.max_seqlen_q))), - .dtype = zml.kernel.cute.from(q.dtype()), - .warps_per_query = if (decode) 4 else 2, - .kv_tile = if (q.dim(.hd) == 512) 32 else 64, - .causal = self.opts.mask.isCausal(), - .window = self.opts.mask.slidingWindow() orelse 0, - .scale = self.opts.scale orelse 1 / @sqrt(@as(f32, @floatFromInt(q.dim(.hd)))), - .sink = self.opts.sink != null, - }; - // Mixed batches also contain single-token rows. Partition their KV - // range so one long decode cannot hold up all the prefill CTAs. - const query_tiles = if (decode) c.batch * c.queryTiles() else c.batch + @divTrunc(c.tokens * c.group + c.rows() - 1, c.rows()); - const work = query_tiles * c.kv_heads; - const bound = if (c.window > 0) @min(c.window, c.pages_per_seq * c.page_size) else c.pages_per_seq * c.page_size; - c.splits = @max(1, @min(32, @min(@divTrunc(@as(i64, @intCast(cute.multiprocessorCount(zml.Compiler.current().platform))) * 2 + work - 1, work), @divTrunc(bound + 127, 128)))); - const sink = if (self.opts.sink) |s| s.convert(.f32).reshape(.{c.kv_heads * c.group}) else zml.Tensor.zeroes(.init(.{c.kv_heads * c.group}, .f32)); - const out = kernel.Kernel.call(.{ - .q = q, - .k = k, - .v = v, - .table = p.block_table, - .lengths = p.seq_lens, - .starts = p.query_start_len, - .sink = sink, - }, .{ .out = if (c.splits > 1) zml.Shape.init(.{ c.splits, c.tokens, c.kv_heads * c.group, c.head_dim + 2 }, .f32) else q.shape() }, .{ - .cfg = c, - .grid = .{ @intCast(c.batch * c.queryTiles()), @intCast(c.kv_heads), @intCast(c.splits) }, - .block = .{ 128, 1, 1 }, - }).out; - const result = if (c.splits > 1) kernel.Merge.call(.{ .partials = out, .starts = p.query_start_len, .sink = sink }, .{ .out = q.shape() }, .{ - .cfg = c, - .grid = .{ @intCast(c.tokens), @intCast(c.kv_heads * c.group), 1 }, - .block = .{ 128, 1, 1 }, - }).out else out; - return result.transpose(output_shape); - } -}; diff --git a/zml/kernels/paged_attention/metal/README.md b/zml/kernels/paged_attention/metal/README.md index 5624de0e..8ccc6fd9 100644 --- a/zml/kernels/paged_attention/metal/README.md +++ b/zml/kernels/paged_attention/metal/README.md @@ -10,11 +10,13 @@ transposes queries and caches into the kernels' layouts, launches the the output. Queries and caches are `bf16` or `f16`; other dtypes panic. Causal `bf16` -decode with head dimension 128, 256 or 512 runs `fa_vec_paged`, which is `bf16` -only, split over four threadgroups and merged by `fa_vec_paged_reduce` at head -dimension 512 past 4096 KV positions. Everything else, `f16` decode included, -runs `paged_attention_tiled` (head dimension 32, 64, 96, 128, 256 or 512, -pages of 8, 16 or 32 positions), then `zero_padding_rows`. +decode with head dimension 32, 64, 128, 256 or 512 runs `fa_vec_paged`, which +is `bf16` only: a threadgroup holds up to four query heads of a KV head, times +splits of the KV positions, at most 256 threads. Rows split over the positions +are merged by `fa_vec_paged_reduce`; the split count follows the GPU core count +and the number of rows. Everything else, `f16` decode included, runs +`paged_attention_tiled` (head dimension 32, 64, 96, 128, 256 or 512, pages of +8, 16 or 32 positions), then `zero_padding_rows`. Causal and non-causal standard masks, left sliding windows, and a custom scale are passed to the kernels. Explicit key indices are rejected, and diff --git a/zml/kernels/paged_attention/metal/metal.zig b/zml/kernels/paged_attention/metal/metal.zig index e2d68398..2cec68f5 100644 --- a/zml/kernels/paged_attention/metal/metal.zig +++ b/zml/kernels/paged_attention/metal/metal.zig @@ -3,8 +3,7 @@ const std = @import("std"); const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); - -const AttentionOptions = @import("../paged_attention.zig").AttentionOptions; +const AttentionOptions = @import("../options.zig").AttentionOptions; const paged_attn = @import("paged_attn.zig"); const log = std.log.scoped(.@"zml/attention/metal"); @@ -86,8 +85,8 @@ const DispatchOptions = struct { scale: f32, }; -/// Launches fa_vec_paged for causal bf16 decode, split over 4 threadgroups at -/// head_dim 512 past 4096 KV positions, and paged_attention_tiled otherwise. +/// Launches fa_vec_paged for causal bf16 decode and paged_attention_tiled +/// otherwise. /// `q` and the result are `[tokens, heads, head_dim]`, the caches /// `[pages, block_size, kv_heads, head_dim]`, `block_table` /// `[seqs, max_blocks]`, `seq_lens` `[seqs]` and `query_start_len` @@ -106,7 +105,7 @@ fn dispatch(q: zml.Tensor, k_cache: zml.Tensor, v_cache: zml.Tensor, block_table const max_blocks: i32 = @intCast(block_table.dim(1)); const vec_decode = opts.is_causal and total_q_tokens == num_seqs and - (head_dim == 128 or head_dim == 256 or head_dim == 512) and + (head_dim == 32 or head_dim == 64 or head_dim == 128 or head_dim == 256 or head_dim == 512) and dtype == .bf16; if (vec_decode) { @@ -129,35 +128,30 @@ fn dispatch(q: zml.Tensor, k_cache: zml.Tensor, v_cache: zml.Tensor, block_table }; const kv_capacity = max_blocks * block_size; - if (head_dim == 512 and kv_capacity > 4096) { - const nwg = 4; - const nrows = total_q_tokens * num_heads; + const span = if (opts.sliding_window < 0) kv_capacity else @min(kv_capacity, opts.sliding_window + 32); + const nrows = total_q_tokens * num_heads; + const launch = vecLaunch(nrows, @divExact(num_heads, num_kv_heads), span, coreCount()); - const partial = paged_attn.FaVecPaged.Kernel.call(inputs, .{ .dst = zml.Shape.init(.{nrows * (head_dim * nwg + 2 * nwg)}, .f32) }, .{ - .cfg = .{ .head_dim = head_dim, .nsg = 8, .block_size = block_size, .position_stride = num_kv_heads * head_dim, .nwg = nwg }, - .grid = .{ nwg, @intCast(num_heads), @intCast(total_q_tokens) }, - .threadgroup = .{ 32, 8, 1 }, - .threadgroup_memory = paged_attn.FaVecPaged.threadgroupMemory(head_dim, 8), - .constants = .{ .args = .of(args) }, - }).dst; + const cfg: paged_attn.FaVecPaged.Cfg = .{ + .head_dim = head_dim, + .block_size = block_size, + .position_stride = num_kv_heads * head_dim, + .heads = launch.heads, + .splits = launch.splits, + }; + const dst = if (launch.splits == 1) q.shape() else zml.Shape.init(.{nrows * (head_dim * launch.splits + 2 * launch.splits)}, .f32); + const out = paged_attn.FaVecPaged.Kernel.call(inputs, .{ .dst = dst }, .{ + .cfg = cfg, + .grid = .{ @intCast(@divExact(launch.splits, launch.splits_per_threadgroup)), @intCast(@divExact(num_heads, launch.heads)), @intCast(total_q_tokens) }, + .threadgroup = .{ 32, @intCast(launch.heads * launch.splits_per_threadgroup), 1 }, + .constants = .{ .args = .of(args) }, + }).dst; + if (launch.splits == 1) return out; - return paged_attn.FaVecPagedReduce.Kernel.call(.{ .htmp = partial }, .{ .dst = q.shape() }, .{ - .cfg = .{ .nwg = nwg, .head_dim = head_dim }, - .grid = .{ @intCast(nrows), 1, 1 }, - .threadgroup = .{ 32, nwg, 1 }, - .constants = .{ .args = .of(@as(i32, nrows)) }, - }).dst; - } - - var idx: usize = if (kv_capacity <= 1024) 0 else if (kv_capacity <= 2048) 1 else 2; - if (head_dim == 512 and idx > 1) idx = 1; - const nsg = ([_]i32{ 4, 8, 16 })[idx]; - - return paged_attn.FaVecPaged.Kernel.call(inputs, .{ .dst = q.shape() }, .{ - .cfg = .{ .head_dim = head_dim, .nsg = nsg, .block_size = block_size, .position_stride = num_kv_heads * head_dim, .nwg = 1 }, - .grid = .{ 1, @intCast(num_heads), @intCast(total_q_tokens) }, - .threadgroup = .{ 32, @intCast(nsg), 1 }, - .threadgroup_memory = paged_attn.FaVecPaged.threadgroupMemory(head_dim, nsg), + return paged_attn.FaVecPagedReduce.Kernel.call(.{ .htmp = out, .seq_lens = seq_lens, .cu_seqlens_q = query_start_len }, .{ .dst = q.shape() }, .{ + .cfg = .{ .splits = launch.splits, .head_dim = head_dim }, + .grid = .{ @intCast(nrows), @intCast(@divExact(head_dim, 4 * paged_attn.FaVecPagedReduce.simdgroups)), 1 }, + .threadgroup = .{ 32, paged_attn.FaVecPagedReduce.simdgroups, 1 }, .constants = .{ .args = .of(args) }, }).dst; } @@ -211,3 +205,36 @@ fn dispatch(q: zml.Tensor, k_cache: zml.Tensor, v_cache: zml.Tensor, block_table .constants = .{ .num_seqs = .of(num_seqs), .row_stride = .of(num_heads * head_dim) }, }).out; } + +const VecLaunch = struct { + heads: i32, + splits: i32, + splits_per_threadgroup: i32, +}; + +/// At most 256 threads, which every Apple GPU pipeline accepts. Splits past a +/// row's length exit early, so `splits` follows the core and row counts. +fn vecLaunch(rows: i32, group: i32, span: i32, cores: i32) VecLaunch { + var heads: i32 = @min(group, 4); + while (@rem(group, heads) != 0) heads -= 1; + const want: i32 = if (rows <= 2 * cores) 128 else 64; + const chunks = @divFloor(span + 31, 32); + var splits: i32 = 1; + while (splits * 2 <= @min(want, chunks)) splits *= 2; + return .{ .heads = heads, .splits = splits, .splits_per_threadgroup = @min(splits, @divFloor(8, heads)) }; +} + +fn coreCount() i32 { + const platform = zml.Compiler.current().platform; + const attribute = platform.devices[0].pjrt_desc.attribute(platform.pjrt_api, "core_count") orelse + std.debug.panic("Metal device has no core_count attribute", .{}); + return @intCast(attribute.int64); +} + +test vecLaunch { + try std.testing.expectEqual(VecLaunch{ .heads = 4, .splits = 128, .splits_per_threadgroup = 2 }, vecLaunch(16, 4, 128000, 40)); + try std.testing.expectEqual(VecLaunch{ .heads = 2, .splits = 64, .splits_per_threadgroup = 4 }, vecLaunch(256, 2, 128000, 40)); + try std.testing.expectEqual(VecLaunch{ .heads = 4, .splits = 64, .splits_per_threadgroup = 2 }, vecLaunch(512, 4, 128000, 40)); + try std.testing.expectEqual(VecLaunch{ .heads = 3, .splits = 8, .splits_per_threadgroup = 2 }, vecLaunch(48, 6, 288, 40)); + try std.testing.expectEqual(VecLaunch{ .heads = 1, .splits = 1, .splits_per_threadgroup = 1 }, vecLaunch(8, 1, 20, 40)); +} diff --git a/zml/kernels/paged_attention/metal/paged_attn.zig b/zml/kernels/paged_attention/metal/paged_attn.zig index 20429565..04f4c1ac 100644 --- a/zml/kernels/paged_attention/metal/paged_attn.zig +++ b/zml/kernels/paged_attention/metal/paged_attn.zig @@ -3,9 +3,8 @@ const std = @import("std"); -const kernel = @import("../../../kernel.zig"); +const metal = @import("../../../kernel/kernel.zig").metal; -const metal = kernel.metal; const ArgSpec = metal.ArgSpec; const Builder = metal.Builder; const Type = metal.Type; @@ -18,25 +17,18 @@ const masked_score = -flt_max / 2.0; const max_half = 65504.0; const simd_width = 32; -fn pad2(x: i32, n: i32) i32 { - return (x + n - 1) & ~(n - 1); -} - -/// fa_vec_paged: decode attention over paged bf16 caches, one threadgroup -/// of `nsg` simdgroups per (position split, query head, query row). +/// fa_vec_paged: bf16 decode over paged caches. Simdgroup s takes query head +/// s % heads of the threadgroup's KV head group and position split s / heads. pub const FaVecPaged = struct { pub const Cfg = struct { - /// The key and value head dimension, DK = DV. + /// The key and value head dimension. head_dim: i32, - /// Simdgroups splitting the KV positions. - nsg: i32, /// The page size. block_size: i32, /// The in-page position stride, n_kv * head_dim elements. position_stride: i32, - /// Position-split threadgroups. Above 1 the kernel writes f32 partials - /// for FaVecPagedReduce. - nwg: i32, + heads: i32, + splits: i32, }; pub const Args: metal.Struct = .{ .name = "fa_vec_paged_args", .fields = &.{ @@ -66,17 +58,10 @@ pub const FaVecPaged = struct { .run = run, }); - /// The threadgroup memory of a launch with `nsg` simdgroups. - pub fn threadgroupMemory(head_dim: i32, nsg: i32) u64 { - const padded = pad2(head_dim, 128); - const bytes = (padded + 4 * 32 + 2 * padded) * nsg * 2; - return @intCast(@divFloor(bytes + 15, 16) * 16); - } - fn run(b: *Builder, cfg: Cfg) metal.FinishError!void { switch (cfg.head_dim) { - 128, 256, 512 => {}, - else => std.debug.panic("fa_vec_paged: head_dim {d} unsupported (128, 256 or 512)", .{cfg.head_dim}), + 32, 64, 128, 256, 512 => {}, + else => std.debug.panic("fa_vec_paged: head_dim {d} unsupported (32, 64, 128, 256 or 512)", .{cfg.head_dim}), } const char_in: ArgSpec = .{ .buffer = .{ .type = .char, .const_ = true } }; const int_in: ArgSpec = .{ .buffer = .{ .type = .int, .const_ = true } }; @@ -89,158 +74,90 @@ pub const FaVecPaged = struct { .seq_lens = int_in, .cu_seqlens_q = int_in, .dst = .{ .buffer = .{ .type = .char } }, - .shmem_f16 = .{ .threadgroup = .half }, .tgpig = .{ .threadgroup_position_in_grid = .uint3 }, .tgpg = .{ .threadgroups_per_grid = .uint3 }, .tiisg = .{ .thread_index_in_simdgroup = .ushort }, .sgitg = .{ .simdgroup_index_in_threadgroup = .ushort }, + .nsg = .{ .simdgroups_per_threadgroup = .ushort }, }); impl(b, cfg, a); } - /// The attention body. fn impl(b: *Builder, cfg: Cfg, a: anytype) void { - const DK = cfg.head_dim; - const DV = cfg.head_dim; - const NE = 1; - const C = 32; - const NSG = cfg.nsg; + const D = cfg.head_dim; + const D4 = @divExact(D, 4); + // A K/V row spans NL lanes of NI float4s; a simdgroup reads R rows per step. + const NL = @min(D4, simd_width); + const R = @divExact(simd_width, NL); + const NI = @divExact(D4, NL); + const C = simd_width; const BS = cfg.block_size; const PS = cfg.position_stride; - const NWG = cfg.nwg; + const HG = cfg.heads; + const NWG = cfg.splits; + const full: metal.LoopOpts = .{ .unroll = .full }; + const num_heads = a.args.field("num_heads"); + const group = num_heads.div(a.args.field("n_kv")); + const groups = group.div(HG); const r = a.tgpig.at(2).to(.int); - const iq2 = a.tgpig.at(1).to(.ushort); - const iwg = a.tgpig.at(0).to(.short); - - const DK4 = @divExact(DK, 4); - const DV4 = @divExact(DV, 4); - - const nrows_static = a.tgpg.at(2).to(.long).mul(a.args.field("num_heads").to(.uint)); + const tg_head = a.tgpig.at(1).to(.int); + const ikv = tg_head.div(groups); + const iq2 = ikv.mul(group).add(tg_head.rem(groups).mul(HG)).add(a.sgitg.rem(HG).to(.int)); + const iwg = a.tgpig.at(0).to(.int).mul(a.nsg.div(HG).to(.int)).add(a.sgitg.div(HG).to(.int)); + const rid = r.to(.long).mul(num_heads.to(.uint)).add(iq2); + const nrows = a.tgpg.at(2).to(.long).mul(num_heads.to(.uint)); + // Padding rows: FaVecPagedReduce zeroes them when there are splits. var padding = b.openIf(r.ge(a.cu_seqlens_q.load(a.args.field("num_seqs")))); { - var first = b.openIf(a.sgitg.eq(0)); - { - const rid = r.to(.long).mul(a.args.field("num_heads").to(.uint)).add(iq2); - if (NWG == 1) { - const dst4 = a.dst.to(.bfloat4); - var loop = b.openFor(a.tiisg.to(.short), DV4, simd_width); - { - const zero = b.constant(.float4, 0).to(.bfloat4); - dst4.store(rid.mul(DV4).add(loop.iv), zero); - loop.close(); - } - } else { - const dst4 = a.dst.to(.float4); - const dst1 = a.dst.to(.float).add(nrows_static.mul(DV).mul(NWG)); - var loop = b.openFor(a.tiisg.to(.short), DV4, simd_width); - { - dst4.store(rid.mul(DV4).mul(NWG).add(b.lift(NWG).mul(loop.iv)).add(iwg), b.constant(.float4, 0)); - loop.close(); - } - var lane0 = b.openIf(a.tiisg.eq(0)); - { - dst1.store(rid.mul(2 * NWG).add(b.lift(2).mul(iwg)).add(0), 0.0); - dst1.store(rid.mul(2 * NWG).add(b.lift(2).mul(iwg)).add(1), -flt_max / 2.0); - lane0.close(); - } + if (NWG == 1) { + const dst4 = a.dst.to(.bfloat4).add(rid.mul(D4)); + var loop = b.openFor(a.tiisg.to(.short), D4, simd_width); + { + dst4.store(loop.iv, b.constant(.float4, 0).to(.bfloat4)); + loop.close(); } - first.close(); } b.return_(); padding.close(); } - const first_seq = b.local(.int, 0); - const last_seq = b.local(.int, a.args.field("num_seqs").sub(1)); - var search = b.openLoop(.{}); - { - search.body(first_seq.lt(last_seq)); - } - { - const mid = first_seq.add(last_seq).add(1).div(2); - var half = b.openIfElse(a.cu_seqlens_q.load(mid).le(r), .{}); + const seq = sequenceOf(b, a, r); + const kv_len = seq.kv_len; + const kv_lo = seq.kv_lo; + const ic0_lo = kv_lo.div(C); + const live = if (NWG > 1) liveSplits(b, seq, NWG) else b.lift(@as(i32, 1)); + + if (NWG > 1) { + var idle = b.openIf(iwg.ge(live)); { - first_seq.assign(mid); - half.elseBranch(); + b.return_(); + idle.close(); } - { - last_seq.assign(mid.sub(1)); - half.close(); - } - search.close(); } - const seq_idx = first_seq.rvalue(); - const q_pos = r.sub(a.cu_seqlens_q.load(seq_idx)); - const query_len = a.cu_seqlens_q.load(seq_idx.add(1)).sub(a.cu_seqlens_q.load(seq_idx)); - const kv_total = a.seq_lens.load(seq_idx); - const kv_len = kv_total.sub(query_len).add(q_pos).add(1); - const kv_lo = b.ternary(a.args.field("sliding_window").lt(0), 0, b.max(0, kv_len.sub(a.args.field("sliding_window")))); - const pt = a.block_tables.add(seq_idx.to(.ulong).mul(a.args.field("max_blocks").to(.uint))); + const pt = a.block_tables.add(seq.idx.to(.ulong).mul(a.args.field("max_blocks").to(.uint))); + const k_cache = a.k_cache.add(ikv.to(.ulong).mul(D).mul(2)); + const v_cache = a.v_cache.add(ikv.to(.ulong).mul(D).mul(2)); - const PK = pad2(DK, 128); - const PK4 = @divExact(PK, 4); - const PV = pad2(DV, 128); - const PV4 = @divExact(PV, 4); - const NW = simd_width; - const NL = @divExact(NW, NE); - const SH = 4 * C; - - const shmem = a.shmem_f16; - const sq4 = shmem.add(0 * PK).to(.half4); - const ss = shmem.add(a.sgitg.mul(SH)).add(NSG * PK).to(.float); - const ss4 = shmem.add(a.sgitg.mul(SH)).add(NSG * PK).to(.float4); - const sm = shmem.add(a.sgitg.mul(SH)).add(2 * C).add(NSG * PK); - const so4 = shmem.add(b.lift(2).mul(a.sgitg).mul(PV)).add(NSG * PK).add(NSG * SH).to(.float4); - // This lane's output, until the simdgroups merge theirs. - const lane_so4 = so4.add(a.tiisg); - - const q = a.q.add(r.to(.ulong).mul(a.args.field("num_heads").to(.uint)).add(iq2).mul(DK).mul(2)); - const ikv2 = iq2.div(a.args.field("num_heads").div(a.args.field("n_kv"))).to(.short); - const k_cache = a.k_cache.add(ikv2.to(.ulong).mul(DK).mul(2)); - const v_cache = a.v_cache.add(ikv2.to(.ulong).mul(DK).mul(2)); - - const q4 = q.to(.bfloat4); - var load_q = b.openFor(a.tiisg.to(.short), PK4, NW); + const lane = if (R == 1) a.tiisg else a.tiisg.rem(NL); + const sub = a.tiisg.div(NL).to(.int); + const q4 = a.q.add(rid.mul(D).mul(2)).to(.bfloat4).add(lane); + const qr = b.array(.float4, @intCast(NI)); + const o = b.array(.float4, @intCast(NI)); + var init = b.openForOpts(b.constant(.short, 0), NI, 1, full); { - const i = load_q.iv; - var value = b.openIfElse(i.lt(DK4), .{.half4}); - { - value.yieldThen(.{q4.load(i).to(.half4)}); - } - { - value.yieldElse(.{b.constant(.half4, 0)}); - } - sq4.store(i, value.results[0]); - load_q.close(); - } - var zero_o = b.openFor(b.constant(.short, 0), @divExact(DV4, NL), 1); - { - lane_so4.store(zero_o.iv.mul(NL), b.constant(.float4, 0)); - zero_o.close(); - } - var zero_s = b.openFor(a.tiisg.to(.short), SH / 4, NW); - { - ss4.store(zero_s.iv, b.constant(.float4, 0)); - zero_s.close(); + const ii = init.iv; + qr.store(ii, q4.load(ii.mul(NL)).to(.float4)); + o.store(ii, b.constant(.float4, 0)); + init.close(); } - b.threadgroupBarrier(.mem_threadgroup); - - // When DV4/NL is a multiple of NW every lane updates the output, so the - // `ty == 0` guard is dropped at comptime. - const all_lanes = @rem(@divExact(DV4, NL), NW) == 0; - const full: metal.LoopOpts = .{ .unroll = .full }; const S = b.local(.float, 0.0); const M = b.local(.float, -flt_max / 2.0); - const tx = a.tiisg.rem(NL).to(.short); - const ty = a.tiisg.div(NL).to(.short); - - const ic0_lo = kv_lo.div(C); - const ic0 = b.local(.int, ic0_lo.add(iwg.mul(NSG)).add(a.sgitg)); + const ic0 = b.local(.int, ic0_lo.add(iwg)); var blocks = b.openLoop(.{}); { const ic = ic0.mul(C); @@ -250,175 +167,103 @@ pub const FaVecPaged = struct { past.close(); } - const in_kv = ic.add(a.tiisg).lt(kv_len).bitAnd(ic.add(a.tiisg).ge(kv_lo)); - sm.store(a.tiisg, b.ternary(in_kv, b.constant(.half, 0), b.constant(.half, -max_half))); - - const pq4 = sq4.add(tx); - const mqk = b.zeroedArray(.float, C / NE); - var qk = b.openForOpts(b.constant(.short, 0), C / NE, 1, full); + const s = b.local(.float, 0.0); + var qk = b.openForOpts(b.constant(.short, 0), @divExact(C, R), 1, full); { const cc = qk.iv; - const p = b.min(ic.add(cc), kv_len.sub(1)); - const pk4 = pagedRow(b, k_cache, pt, p, BS, PS).to(.bfloat4).add(ty.mul(PS).div(4).add(tx)); - var dots = b.openForOpts(b.constant(.short, 0), @divExact(DK4, NL), 1, full); + const p = b.min(if (R == 1) ic.add(cc) else ic.add(cc.to(.int).mul(R)).add(sub), kv_len.sub(1)); + const pk4 = pagedRow(b, k_cache, pt, p, BS, PS).to(.bfloat4).add(lane); + const d = b.local(.float, 0.0); + var dots = b.openForOpts(b.constant(.short, 0), NI, 1, full); { const ii = dots.iv; - const dot = b.dot(pk4.load(ii.mul(NL)).to(.float4), pq4.load(ii.mul(NL)).to(.float4)); - mqk.store(cc, mqk.load(cc).add(dot)); + d.assign(d.add(b.dot(pk4.load(ii.mul(NL)).to(.float4), qr.load(ii)))); dots.close(); } - mqk.store(cc, b.simdSum(mqk.load(cc))); + if (R == 1) { + s.assign(b.ternary(a.tiisg.to(.short).eq(cc), b.simdSum(d.rvalue()), s.rvalue())); + } else { + var x = d.rvalue(); + var mask: i32 = 1; + while (mask < NL) : (mask *= 2) x = x.add(b.simdShuffleXor(x, b.constant(.ushort, mask))); + const t = b.simdShuffle(x, a.tiisg.rem(R).mul(NL)); + s.assign(b.ternary(a.tiisg.div(R).to(.short).eq(cc), t, s.rvalue())); + } qk.close(); } - const score = b.fma(mqk.load(tx), a.args.field("scale"), sm.load(b.lift(NE).mul(tx).add(ty)).to(.float)); - ss.store(b.lift(NE).mul(tx).add(ty), score); - - b.simdgroupBarrier(.mem_threadgroup); + const pos = ic.add(a.tiisg.to(.int)); + const in_kv = pos.lt(kv_len).bitAnd(pos.ge(kv_lo)); + const score = b.ternary(in_kv, s.mul(a.args.field("scale")), b.constant(.float, -max_half)); const m = M.rvalue(); - const s = ss.load(a.tiisg); - - M.assign(b.simdMax(b.max(M.rvalue(), s))); - + M.assign(b.simdMax(b.max(M.rvalue(), score))); const ms = b.exp(m.sub(M)); - const vs = b.exp(s.sub(M)); - + const vs = b.exp(score.sub(M)); S.assign(S.mul(ms).add(b.simdSum(vs))); - ss.store(a.tiisg, vs); - - var rescale_lanes = if (all_lanes) null else b.openIf(ty.eq(0)); - var rescale = b.openForOpts(b.constant(.short, 0), @divExact(DV4, NL), 1, full); + var rescale = b.openForOpts(b.constant(.short, 0), NI, 1, full); { - const at = rescale.iv.mul(NL); - lane_so4.store(at, lane_so4.load(at).mul(ms)); + o.store(rescale.iv, o.load(rescale.iv).mul(ms)); rescale.close(); } - if (rescale_lanes) |*l| l.close(); - b.simdgroupBarrier(.mem_threadgroup); - - const lo = b.array(.float4, @intCast(@divExact(DV4, NL))); - var zero_lo = b.openForOpts(b.constant(.short, 0), @divExact(DV4, NL), 1, full); - { - lo.store(zero_lo.iv, b.constant(.float4, 0)); - zero_lo.close(); - } - - const sst = ss.add(ty); - - var pv = b.openForOpts(b.constant(.short, 0), C / NE, 1, full); + var pv = b.openForOpts(b.constant(.short, 0), @divExact(C, R), 1, full); { const cc = pv.iv; - const p = b.min(ic.add(cc), kv_len.sub(1)); - const pv4 = pagedRow(b, v_cache, pt, p, BS, PS).to(.bfloat4).add(ty.mul(PS).div(4).add(tx)); - var dims = b.openForOpts(b.constant(.short, 0), @divExact(DV4, NL), 1, full); + const row = cc.to(.int).mul(R).add(sub); + const p = b.min(if (R == 1) ic.add(cc) else ic.add(row), kv_len.sub(1)); + const pv4 = pagedRow(b, v_cache, pt, p, BS, PS).to(.bfloat4).add(lane); + const w = b.vec(.float4, .{if (R == 1) b.simdBroadcast(vs, cc.to(.ushort)) else b.simdShuffle(vs, row)}); + var dims = b.openForOpts(b.constant(.short, 0), NI, 1, full); { const ii = dims.iv; - const weighted = pv4.load(ii.mul(NL)).to(.float4).mul(b.vec(.float4, .{sst.load(cc.mul(NE))})); - lo.store(ii, lo.load(ii).add(weighted)); + o.store(ii, o.load(ii).add(pv4.load(ii.mul(NL)).to(.float4).mul(w))); dims.close(); } pv.close(); } - var add_lanes = if (all_lanes) null else b.openIf(ty.eq(0)); - var add_o = b.openForOpts(b.constant(.short, 0), @divExact(DV4, NL), 1, full); - { - const ii = add_o.iv; - const partial = lo.load(ii); - lane_so4.store(ii.mul(NL), lane_so4.load(ii.mul(NL)).add(partial)); - add_o.close(); - } - if (add_lanes) |*l| l.close(); blocks.next(); - ic0.assign(ic0.add(NWG * NSG)); + ic0.assign(ic0.add(live)); blocks.close(); } - var stats = b.openIf(a.tiisg.eq(0)); - { - ss.store(0, S); - ss.store(1, M); - stats.close(); - } - - b.threadgroupBarrier(.mem_threadgroup); - - const rr = b.local(.short, @divFloor(NSG, 2)); - var halving = b.openLoop(.{}); - { - halving.body(rr.gt(0)); - } - { - var group = b.openIf(a.sgitg.lt(rr)); + var mask: i32 = NL; + while (mask < simd_width) : (mask *= 2) { + var sum = b.openForOpts(b.constant(.short, 0), NI, 1, full); { - const S0 = ss.load(0); - const S1 = ss.load(rr.mul(@divExact(SH, 2)).add(0)); - - const M0 = ss.load(1); - const M1 = ss.load(rr.mul(@divExact(SH, 2)).add(1)); - - const M_ = b.max(M0, M1); - - const ms0 = b.exp(M0.sub(M_)); - const ms1 = b.exp(M1.sub(M_)); - - const S_ = S0.mul(ms0).add(S1.mul(ms1)); - - var lane0 = b.openIf(a.tiisg.eq(0)); - { - ss.store(0, S_); - ss.store(1, M_); - lane0.close(); - } - - var merge = b.openFor(a.tiisg.to(.short), DV4, NW); - { - const i = merge.iv; - so4.store(i, so4.load(i).mul(ms0).add(so4.load(i.add(rr.mul(PV4))).mul(ms1))); - merge.close(); - } - group.close(); + o.store(sum.iv, o.load(sum.iv).add(b.simdShuffleXor(o.load(sum.iv), b.constant(.ushort, mask)))); + sum.close(); } - b.threadgroupBarrier(.mem_threadgroup); - rr.assign(rr.shr(1)); - halving.close(); } - - var first = b.openIf(a.sgitg.eq(0)); - { - const rid = r.to(.long).mul(a.args.field("num_heads").to(.uint)).add(iq2); - if (NWG == 1) { - const inv_sum = b.ternary(ss.load(0).eq(0.0), 0.0, b.lift(1.0).div(ss.load(0))); - const dst4 = a.dst.to(.bfloat4); - var loop = b.openFor(a.tiisg.to(.short), DV4, NW); - { - const out = so4.load(loop.iv).mul(inv_sum).to(.bfloat4); - dst4.store(rid.mul(DV4).add(loop.iv), out); - loop.close(); - } - } else { - const dst4 = a.dst.to(.float4); - const dst1 = a.dst.to(.float).add(nrows_static.mul(DV).mul(NWG)); - var loop = b.openFor(a.tiisg.to(.short), DV4, NW); - { - const part = so4.load(loop.iv); - dst4.store(rid.mul(DV4).mul(NWG).add(b.lift(NWG).mul(loop.iv)).add(iwg), part); - loop.close(); - } - var lane0 = b.openIf(a.tiisg.eq(0)); - { - const sum = ss.load(0); - dst1.store(rid.mul(2 * NWG).add(b.lift(2).mul(iwg)).add(0), sum); - const max = ss.load(1); - dst1.store(rid.mul(2 * NWG).add(b.lift(2).mul(iwg)).add(1), max); - lane0.close(); - } + var first = if (R > 1) b.openIf(sub.eq(0)) else null; + if (NWG == 1) { + const inv_sum = b.ternary(S.eq(0.0), 0.0, b.lift(1.0).div(S)); + const dst4 = a.dst.to(.bfloat4).add(rid.mul(D4)).add(lane); + var out = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + dst4.store(out.iv.mul(NL), o.load(out.iv).mul(inv_sum).to(.bfloat4)); + out.close(); + } + } else { + // [rows][D4][splits] float4 partials, then [rows][splits] (S, M). + const dst4 = a.dst.to(.float4).add(rid.mul(D4).mul(NWG)); + var out = b.openForOpts(b.constant(.short, 0), NI, 1, full); + { + const i = lane.to(.int).add(out.iv.to(.int).mul(NL)); + dst4.store(i.mul(NWG).add(iwg), o.load(out.iv)); + out.close(); + } + var lane0 = b.openIf(a.tiisg.eq(0)); + { + const dst1 = a.dst.to(.float).add(nrows.mul(D).mul(NWG)).add(rid.mul(2 * NWG)); + dst1.store(b.lift(2).mul(iwg), S); + dst1.store(b.lift(2).mul(iwg).add(1), M); + lane0.close(); } - first.close(); } + if (first) |*scope| scope.close(); } /// The address of position `p`'s bf16 row in `cache`, through the page @@ -430,85 +275,144 @@ pub const FaVecPaged = struct { } }; -/// fa_vec_paged_reduce: merges the `nwg` partials of a split-K fa_vec_paged, -/// one threadgroup per row and head. +/// The sequence of query row `r` and its causal KV range [kv_lo, kv_len). +fn sequenceOf(b: *Builder, a: anytype, r: Value) struct { idx: Value, kv_len: Value, kv_lo: Value } { + const first_seq = b.local(.int, 0); + const last_seq = b.local(.int, a.args.field("num_seqs").sub(1)); + var search = b.openLoop(.{}); + { + search.body(first_seq.lt(last_seq)); + } + { + const mid = first_seq.add(last_seq).add(1).div(2); + var half = b.openIfElse(a.cu_seqlens_q.load(mid).le(r), .{}); + { + first_seq.assign(mid); + half.elseBranch(); + } + { + last_seq.assign(mid.sub(1)); + half.close(); + } + search.close(); + } + const idx = first_seq.rvalue(); + const q_pos = r.sub(a.cu_seqlens_q.load(idx)); + const query_len = a.cu_seqlens_q.load(idx.add(1)).sub(a.cu_seqlens_q.load(idx)); + const kv_len = a.seq_lens.load(idx).sub(query_len).add(q_pos).add(1); + const kv_lo = b.ternary(a.args.field("sliding_window").lt(0), 0, b.max(0, kv_len.sub(a.args.field("sliding_window")))); + return .{ .idx = idx, .kv_len = kv_len, .kv_lo = kv_lo }; +} + +fn liveSplits(b: *Builder, seq: anytype, splits: i32) Value { + const chunks = seq.kv_len.add(simd_width - 1).div(simd_width).sub(seq.kv_lo.div(simd_width)); + return b.min(splits, chunks); +} + +/// fa_vec_paged_reduce: merges the fa_vec_paged split partials. pub const FaVecPagedReduce = struct { pub const Cfg = struct { - /// The partials to merge, FaVecPaged's `nwg`. - nwg: i32, + splits: i32, /// The head dimension. head_dim: i32, }; - pub const Args: metal.Struct = .{ .name = "fa_vec_paged_reduce_args", .fields = &.{ - .{ .name = "nrows", .type = .int }, - } }; + pub const simdgroups = 8; pub const Kernel = metal.Kernel(Cfg, .{ .name = "fa_vec_paged_reduce", - .inputs = &.{"htmp"}, + .inputs = &.{ "htmp", "seq_lens", "cu_seqlens_q" }, .outputs = &.{"dst"}, .constants = &.{"args"}, .run = run, }); fn run(b: *Builder, cfg: Cfg) metal.FinishError!void { + const int_in: ArgSpec = .{ .buffer = .{ .type = .int, .const_ = true } }; const a = try b.declareArgs(.{ - .args = .{ .constant_struct = Args }, + .args = .{ .constant_struct = FaVecPaged.Args }, .htmp = .{ .buffer = .{ .type = .char, .const_ = true } }, + .seq_lens = int_in, + .cu_seqlens_q = int_in, .dst = .{ .buffer = .{ .type = .char } }, - .tgpig = .{ .threadgroup_position_in_grid = .uint }, + .tgpig = .{ .threadgroup_position_in_grid = .uint2 }, + .tgpg = .{ .threadgroups_per_grid = .uint2 }, .tiisg = .{ .thread_index_in_simdgroup = .ushort }, .sgitg = .{ .simdgroup_index_in_threadgroup = .ushort }, }); - const nwg = cfg.nwg; + const NWG = cfg.splits; + if (NWG > 4 * simd_width) std.debug.panic("fa_vec_paged_reduce: {d} splits, at most {d}", .{ NWG, 4 * simd_width }); + const NP = @divFloor(NWG + simd_width - 1, simd_width); const DV = cfg.head_dim; - const rid = a.tgpig.to(.ulong); - const iwg = a.tiisg.to(.short); - const ss = a.htmp.to(.float).add(a.args.field("nrows").to(.ulong).mul(DV).mul(nwg)); + const DV4 = @divExact(DV, 4); + const full: metal.LoopOpts = .{ .unroll = .full }; - var max = b.openIfElse(iwg.lt(nwg), .{.float}); - { - max.yieldThen(.{ss.load(rid.mul(2 * nwg).add(b.lift(2).mul(iwg)).add(1))}); - } - { - max.yieldElse(.{-flt_max / 2.0}); - } - const M = max.results[0]; - var sum = b.openIfElse(iwg.lt(nwg), .{.float}); - { - sum.yieldThen(.{ss.load(rid.mul(2 * nwg).add(b.lift(2).mul(iwg)).add(0))}); - } - { - sum.yieldElse(.{0.0}); - } - - const m = b.simdMax(M); - const ms = b.exp(M.sub(m)); - const S = b.simdSum(sum.results[0].mul(ms)); - const inv_sum = b.ternary(S.eq(0.0), 0.0, b.lift(1.0).div(S)); - - const DV4 = @divTrunc(DV, 4); - const htmp4 = a.htmp.to(.float4).add(rid.mul(DV4).mul(nwg)); + const rid = a.tgpig.at(0).to(.ulong); + const r = a.tgpig.at(0).div(a.args.field("num_heads").to(.uint)).to(.int); + const i = a.tgpig.at(1).to(.int).mul(simdgroups).add(a.sgitg.to(.int)); const dst4 = a.dst.to(.bfloat4).add(rid.mul(DV4)); - var loop = b.openFor(a.sgitg.to(.int), DV4, nwg); + + const live = b.local(.int, 0); + var real = b.openIf(r.lt(a.cu_seqlens_q.load(a.args.field("num_seqs")))); { - const i = loop.iv; - var part_ = b.openIfElse(iwg.lt(nwg), .{.float4}); + live.assign(liveSplits(b, sequenceOf(b, a, r), NWG)); + real.close(); + } + var empty = b.openIf(live.le(0)); + { + var lane0 = b.openIf(a.tiisg.eq(0)); { - part_.yieldThen(.{htmp4.load(i.mul(nwg).add(iwg)).mul(ms)}); - } - { - part_.yieldElse(.{b.constant(.float4, 0)}); - } - const part = part_.results[0]; - const v = b.simdSum(part); - var lane0 = b.openIf(iwg.eq(0)); - { - dst4.store(i, v.mul(inv_sum).to(.bfloat4)); + dst4.store(i, b.constant(.float4, 0).to(.bfloat4)); lane0.close(); } - loop.close(); + b.return_(); + empty.close(); + } + + const ss = a.htmp.to(.float).add(a.tgpg.at(0).to(.ulong).mul(DV).mul(NWG)).add(rid.mul(2 * NWG)); + const Mj = b.array(.float, @intCast(NP)); + const Sj = b.array(.float, @intCast(NP)); + const mx = b.local(.float, -flt_max / 2.0); + var load = b.openForOpts(b.constant(.short, 0), NP, 1, full); + { + const j = load.iv; + const idx = a.tiisg.to(.int).add(j.to(.int).mul(simd_width)); + const ok = idx.lt(live); + Mj.store(j, b.ternary(ok, ss.load(idx.mul(2).add(1)), b.constant(.float, -flt_max / 2.0))); + Sj.store(j, b.ternary(ok, ss.load(idx.mul(2)), b.constant(.float, 0))); + mx.assign(b.max(mx.rvalue(), Mj.load(j))); + load.close(); + } + const m = b.simdMax(mx.rvalue()); + const sum = b.local(.float, 0.0); + var weigh = b.openForOpts(b.constant(.short, 0), NP, 1, full); + { + const j = weigh.iv; + Mj.store(j, b.exp(Mj.load(j).sub(m))); + sum.assign(sum.add(Sj.load(j).mul(Mj.load(j)))); + weigh.close(); + } + const S = b.simdSum(sum.rvalue()); + const inv_sum = b.ternary(S.eq(0.0), 0.0, b.lift(1.0).div(S)); + + const htmp4 = a.htmp.to(.float4).add(rid.mul(DV4).mul(NWG)); + const acc = b.local(.float4, b.constant(.float4, 0)); + var parts = b.openForOpts(b.constant(.short, 0), NP, 1, full); + { + const j = parts.iv; + const idx = a.tiisg.to(.int).add(j.to(.int).mul(simd_width)); + var ok = b.openIf(idx.lt(live)); + { + acc.assign(acc.add(htmp4.load(i.mul(NWG).add(idx)).mul(Mj.load(j)))); + ok.close(); + } + parts.close(); + } + const v = b.simdSum(acc.rvalue()); + var lane0 = b.openIf(a.tiisg.eq(0)); + { + dst4.store(i, v.mul(inv_sum).to(.bfloat4)); + lane0.close(); } } }; @@ -984,13 +888,14 @@ pub const PagedAttentionTiled = struct { test "the paged attention kernels emit verified modules" { const allocator = std.testing.allocator; - for ([_]i32{ 128, 256, 512 }) |head_dim| { - for ([_]i32{ 4, 8, 16 }) |nsg| { - allocator.free(try FaVecPaged.Kernel.emit(allocator, .{ .head_dim = head_dim, .nsg = nsg, .block_size = 16, .position_stride = 2 * head_dim, .nwg = 1 })); + for ([_]i32{ 32, 64, 128, 256, 512 }) |head_dim| { + for ([_][2]i32{ .{ 1, 1 }, .{ 4, 1 }, .{ 2, 8 }, .{ 4, 128 } }) |launch| { + allocator.free(try FaVecPaged.Kernel.emit(allocator, .{ .head_dim = head_dim, .block_size = 16, .position_stride = 2 * head_dim, .heads = launch[0], .splits = launch[1] })); + } + for ([_]i32{ 8, 128 }) |splits| { + allocator.free(try FaVecPagedReduce.Kernel.emit(allocator, .{ .splits = splits, .head_dim = head_dim })); } } - allocator.free(try FaVecPaged.Kernel.emit(allocator, .{ .head_dim = 512, .nsg = 8, .block_size = 16, .position_stride = 1024, .nwg = 4 })); - allocator.free(try FaVecPagedReduce.Kernel.emit(allocator, .{ .nwg = 4, .head_dim = 512 })); for ([_]Type{ .half, .bfloat }) |dtype| { allocator.free(try ZeroPaddingRows.Kernel.emit(allocator, .{ .dtype = dtype })); for ([_]i32{ 32, 64, 96, 128, 256, 512 }) |head_size| { diff --git a/zml/kernels/paged_attention/mosaic_tpu/mosaic_tpu.zig b/zml/kernels/paged_attention/mosaic_tpu/mosaic_tpu.zig index ef9e852c..84716379 100644 --- a/zml/kernels/paged_attention/mosaic_tpu/mosaic_tpu.zig +++ b/zml/kernels/paged_attention/mosaic_tpu/mosaic_tpu.zig @@ -5,7 +5,7 @@ const ragged_paged = @import("platforms/tpu/ragged_paged"); const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); -const AttentionOptions = @import("../paged_attention.zig").AttentionOptions; +const AttentionOptions = @import("../options.zig").AttentionOptions; const ragged_attention = @import("ragged_attention.zig"); const log = std.log.scoped(.tpu_attention); @@ -153,32 +153,28 @@ pub const mosaic_tpu = struct { "mosaic_tpu ragged paged attention q head_dim {} exceeds kernel/cache head_dim {}", .{ q.dim(.hd), target_head_dim }, ); - var q_padded = q.pad(0, .{ .hd = zml.Tensor.Pad{ .high = target_head_dim - q.dim(.hd) } }); - q_padded._shape._sharding = q._shape._sharding; - return q_padded; + return q.pad(0, .{ .hd = zml.Tensor.Pad{ .high = target_head_dim - q.dim(.hd) } }); } inline fn prepareInputs(parameters: Parameters, q: zml.Tensor, kv_pages: zml.Tensor) PreparedInputs { - const mesh = zml.Compiler.current().resolveMesh(.{.model}); - const kv_pages_partitioned = kv_pages.withPartitioning(mesh, .{ .hkv = .model }); - const q_merged = q.merge(.{ .h = .{ .hkv, .hg } }).withPartitioning(mesh, .{ .h = .model }); + const q_merged = q.merge(.{ .h = .{ .hkv, .hg } }); const logical_head_dim: usize = @intCast(q_merged.dim(.hd)); - const cache_head_dim: usize = @intCast(kv_pages_partitioned.dim(.hd)); + const cache_head_dim: usize = @intCast(kv_pages.dim(.hd)); stdx.debug.assert( cache_head_dim == alignedKernelHeadDim(logical_head_dim), "mosaic_tpu ragged paged attention expects KV cache head_dim {} to match aligned logical head_dim {}", .{ cache_head_dim, logical_head_dim }, ); - const q_ragged = alignQueryHeadDimForKernel(q_merged, kv_pages_partitioned.dim(.hd)); - const query_start_len = parameters.query_start_len.withPartitioning(mesh, .{ .b = .replicated }); + const q_ragged = alignQueryHeadDimForKernel(q_merged, kv_pages.dim(.hd)); + const query_start_len = parameters.query_start_len; return .{ .q = q_ragged, - .kv_pages = kv_pages_partitioned, - .seq_lens = parameters.seq_lens.withPartitioning(mesh, .{ ._0 = .replicated }), - .block_table = parameters.block_table.withPartitioning(mesh, .{ ._0 = .replicated, ._1 = .replicated }), + .kv_pages = kv_pages, + .seq_lens = parameters.seq_lens, + .block_table = parameters.block_table, .query_start_len = query_start_len, - .num_seqs = activeSequenceCount(query_start_len).withPartitioning(mesh, .{ ._0 = .replicated }), + .num_seqs = activeSequenceCount(query_start_len), }; } @@ -224,57 +220,9 @@ pub const mosaic_tpu = struct { pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, kv_cache: zml.Tensor, opts: AttentionOptions) zml.Tensor { if (opts.mask == .indices) @panic("Sparse Attention is not supported on TPU"); - const prepared = prepareInputs(parameters, q, kv_cache); - - const q_out = zml.ops.manualComputation( - (struct { - q: zml.Tensor, - kv_pages: zml.Tensor, - seq_lens: zml.Tensor, - block_table: zml.Tensor, - query_start_len: zml.Tensor, - num_seqs: zml.Tensor, - opts: AttentionOptions, - parameters: Parameters, - - fn body(self: @This(), output: zml.Shape) zml.Tensor { - const prepared_inputs: PreparedInputs = .{ - .q = self.q, - .kv_pages = self.kv_pages, - .seq_lens = self.seq_lens, - .block_table = self.block_table, - .query_start_len = self.query_start_len, - .num_seqs = self.num_seqs, - }; - - const cfg = buildCfg(prepared_inputs, self.parameters, self.opts); - const q_out = raggedPagedKernelCall( - prepared_inputs.q, - prepared_inputs.kv_pages, - prepared_inputs.seq_lens, - prepared_inputs.block_table, - prepared_inputs.query_start_len, - prepared_inputs.num_seqs, - cfg, - ); - stdx.debug.assert(q_out.shape().eql(output), "mosaic_tpu ragged paged attention output shape mismatch, got {f}, expected {f}", .{ q_out.shape(), output }); - return q_out; - } - }).body, - .{ - .q = prepared.q, - .kv_pages = prepared.kv_pages, - .seq_lens = prepared.seq_lens, - .block_table = prepared.block_table, - .query_start_len = prepared.query_start_len, - .num_seqs = prepared.num_seqs, - .opts = opts, - .parameters = parameters, - }, - prepared.q.shape(), - ); - + const cfg = buildCfg(prepared, parameters, opts); + const q_out = raggedPagedKernelCall(prepared.q, prepared.kv_pages, prepared.seq_lens, prepared.block_table, prepared.query_start_len, prepared.num_seqs, cfg); const restored = restoreQueryHeads(q, q_out); stdx.debug.assert(restored.shape().eql(q.shape()), "mosaic_tpu ragged paged attention output shape mismatch, got {f}, expected {f}", .{ restored.shape(), q.shape() }); return restored; diff --git a/zml/kernels/paged_attention/options.zig b/zml/kernels/paged_attention/options.zig new file mode 100644 index 00000000..35ad4049 --- /dev/null +++ b/zml/kernels/paged_attention/options.zig @@ -0,0 +1,36 @@ +const Tensor = @import("../../tensor.zig").Tensor; + +pub const AttentionMask = union(enum) { + standard: struct { + /// Restrict attention to keys at or before the query position. + causal: bool = true, + /// Optional left window size, including the current token. + /// Non-causal attention also includes every later valid key. + sliding_window: ?u32 = null, + }, + /// Attend to keys at specific positions. + indices: Tensor, + + /// Attend to every valid key in the sequence. + pub const none: AttentionMask = .{ .standard = .{ .causal = false } }; + + pub fn isCausal(mask: AttentionMask) bool { + return switch (mask) { + .standard => |options| options.causal, + .indices => false, + }; + } + + pub fn slidingWindow(mask: AttentionMask) ?u32 { + return switch (mask) { + .standard => |options| options.sliding_window, + .indices => null, + }; + } +}; + +pub const AttentionOptions = struct { + mask: AttentionMask = .{ .standard = .{} }, + scale: ?f32 = null, + sink: ?Tensor = null, +}; diff --git a/zml/kernels/paged_attention/paged_attention.zig b/zml/kernels/paged_attention/paged_attention.zig index 99065140..4a97a38e 100644 --- a/zml/kernels/paged_attention/paged_attention.zig +++ b/zml/kernels/paged_attention/paged_attention.zig @@ -3,12 +3,14 @@ const std = @import("std"); const stdx = @import("stdx"); const zml = @import("../../zml.zig"); -const cute = @import("cute/cute.zig"); +pub const AttentionMask = @import("options.zig").AttentionMask; +pub const AttentionOptions = @import("options.zig").AttentionOptions; +pub const cute = @import("cute/cute.zig"); const metal = @import("metal/metal.zig"); const tpu = @import("mosaic_tpu/mosaic_tpu.zig"); -const triton = @import("triton/triton.zig"); -pub const tcl = @import("tcl/tcl.zig"); const stablehlo = @import("stablehlo/stablehlo.zig"); +pub const tcl = @import("tcl/tcl.zig"); +const triton = @import("triton/triton.zig"); const PagedAttention = @This(); @@ -30,7 +32,7 @@ pub const Backend = enum { pub fn auto(platform: *const zml.Platform) Backend { return switch (platform.target) { - .cuda => .triton, + .cuda => if (cute.isAvailable(platform)) .cute else .triton, .rocm => .triton, .oneapi => .triton, .tpu => .mosaic_tpu, @@ -151,41 +153,6 @@ pub const Parameters = union(Backend) { } }; -pub const AttentionMask = union(enum) { - standard: struct { - /// Restrict attention to keys at or before the query position. - causal: bool = true, - /// Optional left window size, including the current token. - /// Non-causal attention also includes every later valid key. - sliding_window: ?u32 = null, - }, - /// Attend to keys at specific positions. - indices: zml.Tensor, - - /// Attend to every valid key in the sequence. - pub const none: AttentionMask = .{ .standard = .{ .causal = false } }; - - pub fn isCausal(mask: AttentionMask) bool { - return switch (mask) { - .standard => |options| options.causal, - .indices => false, - }; - } - - pub fn slidingWindow(mask: AttentionMask) ?u32 { - return switch (mask) { - .standard => |options| options.sliding_window, - .indices => null, - }; - } -}; - -pub const AttentionOptions = struct { - mask: AttentionMask = .{ .standard = .{} }, - scale: ?f32 = null, - sink: ?zml.Tensor = null, -}; - pub const KvCache = union(enum) { split: struct { k: zml.Tensor, @@ -232,49 +199,53 @@ pub const KvCache = union(enum) { const offset = slot_mapping.remainder(.scalar(page_size, slot_mapping.dtype())); return .{ page_index, offset }; } +}; - pub fn getPage( - self: KvCache, - page_index: zml.Tensor, - ) [2]zml.Tensor { - return switch (self) { - .split => |split| .{ - split.k.slice(split.k.axis(.page), .dynSingle(page_index)), - split.v.slice(split.k.axis(.page), .dynSingle(page_index)), +pub const Partitioning = zml.ops.ManualPartitioning(Attention.forward); + +pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, kv_cache: KvCache, opts: AttentionOptions, mesh: *const zml.Mesh, partitioning: Partitioning) zml.Tensor { + _ = k; + _ = v; + const args: Attention = .{ .parameters = parameters, .q = q, .kv_cache = kv_cache, .opts = opts }; + return zml.ops.manualComputation(Attention.forward, args, q.shape(), mesh, partitioning); +} + +const Attention = struct { + parameters: Parameters, + q: zml.Tensor, + kv_cache: KvCache, + opts: AttentionOptions, + + fn forward(args: @This(), _: zml.Shape) zml.Tensor { + return switch (args.parameters) { + .cute => |p| switch (args.kv_cache) { + .split => |split| cute.pagedAttention(p, args.q, split.k, split.v, args.opts), + .dense => @panic("CuTe paged attention requires split KV pages"), + }, + .triton => |triton_parameters| switch (args.kv_cache) { + .split => |split| triton.paged.pagedAttention(triton_parameters, args.q, split.k, split.v, args.opts), + .dense => std.debug.panic("fused KV pages are only supported with the mosaic_tpu backend", .{}), + }, + .mosaic_tpu => |mosaic_tpu_parameters| switch (args.kv_cache) { + .split => std.debug.panic("split KV pages is not supported with the mosaic_tpu backend", .{}), + .dense => tpu.mosaic_tpu.pagedAttention(mosaic_tpu_parameters, args.q, args.kv_cache.dense, args.opts), + }, + .metal => |metal_parameters| switch (args.kv_cache) { + .split => |split| metal.pagedAttention(metal_parameters, args.q, split.k, split.v, args.opts), + .dense => std.debug.panic("fused KV pages are only supported with the mosaic_tpu backend", .{}), + }, + .stablehlo => |params| switch (args.kv_cache) { + .split => |split| stablehlo.pagedAttention(params, args.q, split.k, split.v, args.opts), + .dense => std.debug.panic("fused KV pages are only supported with the mosaic_tpu backend", .{}), + }, + .tcl => |params| switch (args.kv_cache) { + .split => |split| tcl.paged.pagedAttention(params, args.q, split.k, split.v, args.opts), + .dense => std.debug.panic("TCL paged attention takes separate K and V pages", .{}), }, - .dense => @panic("TODO"), }; } }; -pub fn pagedAttention(parameters: Parameters, q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, kv_cache: KvCache, opts: AttentionOptions) zml.Tensor { - _ = k; - _ = v; - return switch (parameters) { - .cute => |p| switch (kv_cache) { - .split => |split| cute.pagedAttention(p, q, split.k, split.v, opts), - .dense => @panic("CuTe paged attention requires split KV pages"), - }, - .triton => |triton_parameters| switch (kv_cache) { - .split => |split| triton.paged.pagedAttention(triton_parameters, q, split.k, split.v, opts), - .dense => std.debug.panic("fused KV pages are only supported with the mosaic_tpu backend", .{}), - }, - .mosaic_tpu => |mosaic_tpu_parameters| switch (kv_cache) { - .split => std.debug.panic("split KV pages is not supported with the mosaic_tpu backend", .{}), - .dense => tpu.mosaic_tpu.pagedAttention(mosaic_tpu_parameters, q, kv_cache.dense, opts), - }, - .metal => |metal_parameters| switch (kv_cache) { - .split => |split| metal.pagedAttention(metal_parameters, q, split.k, split.v, opts), - .dense => std.debug.panic("fused KV pages are only supported with the mosaic_tpu backend", .{}), - }, - .stablehlo => |params| stablehlo.pagedAttention(params, q, kv_cache, opts), - .tcl => |params| switch (kv_cache) { - .split => |split| tcl.paged.pagedAttention(params, q, split.k, split.v, opts), - .dense => std.debug.panic("TCL paged attention takes separate K and V pages", .{}), - }, - }; -} - test "Backend.auto selects mosaic_tpu on TPU" { const platform: zml.Platform = .{ .arena = undefined, @@ -335,21 +306,22 @@ test pagedAttention { const batch_size = active_batch_size + 1; const query_token_count = prefill_token_count + num_decode; const dt: zml.DataType = .bf16; - const model_meshe = platform.meshes.get("model").?; + const tp = zml.testing.tpMesh(); + const num_kv_heads = tp.numPartitionsForLogicalAxis(.model); const partition = .{ .hkv = .model }; const tensors: struct { q: zml.Tensor, k: zml.Tensor, v: zml.Tensor, kv_cache: KvCache } = .{ - .q = zml.Tensor.fromShape(.init(.{ .b = query_token_count, .hkv = 4, .hg = 4, .hd = 32 }, dt)).withPartitioning(model_meshe, partition), - .k = zml.Tensor.fromShape(.init(.{ .b = query_token_count, .hkv = 4, .hd = 32 }, dt)).withPartitioning(model_meshe, partition), - .v = zml.Tensor.fromShape(.init(.{ .b = query_token_count, .hkv = 4, .hd = 32 }, dt)).withPartitioning(model_meshe, partition), + .q = zml.Tensor.fromShape(.init(.{ .b = query_token_count, .hkv = num_kv_heads, .hg = 4, .hd = 32 }, dt)).withPartitioning(tp, partition), + .k = zml.Tensor.fromShape(.init(.{ .b = query_token_count, .hkv = num_kv_heads, .hd = 32 }, dt)).withPartitioning(tp, partition), + .v = zml.Tensor.fromShape(.init(.{ .b = query_token_count, .hkv = num_kv_heads, .hd = 32 }, dt)).withPartitioning(tp, partition), .kv_cache = .{ .split = .{ - .k = zml.Tensor.fromShape(.init(.{ .page = num_pages, .k_chunk = page_size, .hkv = 4, .hd = 32 }, dt)).withPartitioning(model_meshe, partition), - .v = zml.Tensor.fromShape(.init(.{ .page = num_pages, .k_chunk = page_size, .hkv = 4, .hd = 32 }, dt)).withPartitioning(model_meshe, partition), + .k = zml.Tensor.fromShape(.init(.{ .page = num_pages, .k_chunk = page_size, .hkv = num_kv_heads, .hd = 32 }, dt)).withPartitioning(tp, partition), + .v = zml.Tensor.fromShape(.init(.{ .page = num_pages, .k_chunk = page_size, .hkv = num_kv_heads, .hd = 32 }, dt)).withPartitioning(tp, partition), }, }, }; - const meshes: []const *const zml.Mesh = &.{ platform.replicated_mesh, platform.meshes.get("model").? }; + const meshes: []const *const zml.Mesh = &.{ platform.replicated_mesh, tp }; const rng_q = try platform.compileFn(allocator, io, zml.Tensor.Rng.normal, .{ tensors.q.shape(), .{} }, .{ .meshes = meshes }); defer rng_q.deinit(); const rng_k = try platform.compileFn(allocator, io, zml.Tensor.Rng.normal, .{ tensors.k.shape(), .{} }, .{ .meshes = meshes }); @@ -462,11 +434,22 @@ test pagedAttention { backend_options_args.backend = backend; const parameters: Parameters = .init(.fromBackend(backend_options_args)); + const head_spec: zml.Sharding.Partitioning.Logical = .parse(.{ .hkv = .model }); + const partitioning: Partitioning = .{ + .in_specs = .{ + .parameters = try zml.ops.partitionSpecsAlloc(arena, parameters, .replicated), + .q = head_spec, + .kv_cache = .{ .split = .{ .k = head_spec, .v = head_spec } }, + .opts = .{ .mask = .{ .standard = {} }, .sink = null }, + }, + .out_specs = head_spec, + }; + const exe = try platform.compileFn( allocator, io, pagedAttention, - .{ parameters, tensors.q, tensors.k, tensors.v, tensors.kv_cache, test_case.attention_options }, + .{ parameters, tensors.q, tensors.k, tensors.v, tensors.kv_cache, test_case.attention_options, tp, partitioning }, .{ .program_name = try std.fmt.allocPrint(arena, "paged_attention_{s}_{t}", .{ test_case.name, backend }), .meshes = meshes }, ); defer exe.deinit(); diff --git a/zml/kernels/paged_attention/stablehlo/stablehlo.zig b/zml/kernels/paged_attention/stablehlo/stablehlo.zig index ab0ffee2..55ee01d5 100644 --- a/zml/kernels/paged_attention/stablehlo/stablehlo.zig +++ b/zml/kernels/paged_attention/stablehlo/stablehlo.zig @@ -1,12 +1,12 @@ const zml = @import("../../../zml.zig"); -const KvCache = @import("../paged_attention.zig").KvCache; -const AttentionOptions = @import("../paged_attention.zig").AttentionOptions; +const AttentionOptions = @import("../options.zig").AttentionOptions; const triton = @import("../triton/triton.zig"); pub fn pagedAttention( parameters: triton.paged.Parameters, q: zml.Tensor, - kv_cache: KvCache, + k: zml.Tensor, + v: zml.Tensor, opts: AttentionOptions, ) zml.Tensor { switch (opts.mask) { @@ -14,11 +14,12 @@ pub fn pagedAttention( .indices => @panic("Sparse attention mask is not supported on StableHLO"), } - const page_size = kv_cache.split.k.dim(.k_chunk); + const page_size = k.dim(.k_chunk); + const padded_q = q.pad(0, .{ .b = zml.Tensor.Pad{ .high = page_size } }); const final_state = zml.ops.@"while"( AttentionLoop, - .{ .q = q, .kv_cache = kv_cache, .parameters = parameters, .opts = opts }, + .{ .q = padded_q, .k = k, .v = v, .parameters = parameters, .opts = opts }, .{ // i32 to match triton conventions .slot_id = .scalar(0, .i32), @@ -28,16 +29,17 @@ pub fn pagedAttention( // Note: we do all partial softmax in f32 .partial_softmax_prefill = .zeroes(q.shape().setDim(.b, page_size).withDtype(.f32)), .partial_softmax_decode = .zeroes(q.shape().setDim(.b, 1).withDtype(.f32)), - .out = .zeroes(q.shape()), + .out = .zeroes(padded_q.shape()), }, ); - return final_state.out; + return final_state.out.slice(.b, .{ .end = q.dim(.b) }); } const AttentionLoop = struct { q: zml.Tensor, - kv_cache: KvCache, + k: zml.Tensor, + v: zml.Tensor, parameters: triton.paged.Parameters, opts: AttentionOptions, @@ -65,7 +67,7 @@ const AttentionLoop = struct { const query_start = self.parameters.query_start_len.slice(0, .dynSingle(state_.seq_id)); const query_end = self.parameters.query_start_len.slice(0, .dynSingle(state_.seq_id.addConstant(1))); const seq_num_queries_: zml.Tensor = .asScalar(.sub(query_end, query_start)); - const page_size_: u32 = @intCast(self.kv_cache.split.k.dim(.k_chunk)); + const page_size_: u32 = @intCast(self.k.dim(.k_chunk)); return zml.ops.@"if"( struct { @@ -93,7 +95,7 @@ const AttentionLoop = struct { const sequence_has_more_queries: zml.Tensor = .cmp(state.q_page_idx.addConstant(1).scale(num_slots), .LT, if_ctx.seq_num_queries); const partial_softmax = while_body.attentionOnePage(state, q_offset, num_slots, page_size); - return updateState(state, num_slots, query_has_more_keys, sequence_has_more_queries, partial_softmax); + return updateState(state, num_slots, query_has_more_keys, sequence_has_more_queries, partial_softmax, while_body.nextQueryStart(state)); } /// Decode version: num_slots == 1, we multiply one query with a page of keys. @@ -112,7 +114,7 @@ const AttentionLoop = struct { const sequence_has_more_queries: zml.Tensor = .scalar(false, .bool); const partial_softmax = while_body.attentionOnePage(state, q_offset, num_slots, page_size); - return updateState(state, num_slots, query_has_more_keys, sequence_has_more_queries, partial_softmax); + return updateState(state, num_slots, query_has_more_keys, sequence_has_more_queries, partial_softmax, while_body.nextQueryStart(state)); } }, .{ @@ -133,7 +135,10 @@ const AttentionLoop = struct { .{ .b, .p }, &.{ .dynSingle(state.seq_id), .dynSingle(state.k_page_idx) }, ); - const active_k, const active_v = self.kv_cache.getPage(active_k_page_id); + + const page_axis = self.k.axis(.page); + const active_k = self.k.slice(page_axis, .dynSingle(active_k_page_id)); + const active_v = self.v.slice(page_axis, .dynSingle(active_k_page_id)); const dtype = self.q.dtype(); const attn_mask: zml.Tensor = attn_mask: { @@ -162,7 +167,11 @@ const AttentionLoop = struct { return og_softmax.merge(partial); } - pub fn updateState(state: State, num_slots: u32, query_has_more_keys: zml.Tensor, sequence_has_more_queries: zml.Tensor, softmax: PartialSoftmax) State { + fn nextQueryStart(self: While, state: State) zml.Tensor { + return self.parameters.query_start_len.slice(0, .dynSingle(state.seq_id.addConstant(1))).asScalar(); + } + + pub fn updateState(state: State, num_slots: u32, query_has_more_keys: zml.Tensor, sequence_has_more_queries: zml.Tensor, softmax: PartialSoftmax, next_query_start: zml.Tensor) State { return zml.ops.if2( query_has_more_keys, // If the query has more pages: update partial softmax with attention on current page, delay updating output. @@ -181,7 +190,7 @@ const AttentionLoop = struct { .partial_softmax_prefill = if (num_slots == 1) state.partial_softmax_prefill else .reset(softmax), .partial_softmax_decode = if (num_slots == 1) .reset(softmax) else state.partial_softmax_decode, .out = state.out.dynamicUpdateSlice(.{ .b = state.slot_id }, softmax.finalize().convert(state.out.dtype())), - .slot_id = state.slot_id.addConstant(num_slots), + .slot_id = .select(sequence_has_more_queries, state.slot_id.addConstant(num_slots), next_query_start), // If the sequence has more query chunks: increment q_page_idx, otherwise start over at 0. .q_page_idx = .select(sequence_has_more_queries, state.q_page_idx.addConstant(1), .scalar(0, .i32)), // If the sequence has more query chunks: keep seq_id, otherwise increment seq_id. diff --git a/zml/kernels/paged_attention/tcl/tcl.zig b/zml/kernels/paged_attention/tcl/tcl.zig index ce8bdb4f..d04b684e 100644 --- a/zml/kernels/paged_attention/tcl/tcl.zig +++ b/zml/kernels/paged_attention/tcl/tcl.zig @@ -1,42 +1,55 @@ -//! Furiosa RNGD paged attention in TCL, using the common/flash attention -//! algorithms from Furiosa's SDK (`furiosa-kernels 2026.4.0` reference). -//! This implementation retains FP32 normalization/accumulation and BF16 -//! contractions. The exact common SDK operation materializes scores in BF16; -//! it fails the required unit-scale tolerance and is not a selectable backend. -//! -//! Queries are cut in blocks of rows of one sequence (one token per block for -//! decode). A context of at most `chunk_keys` keys runs Furiosa's one-shot -//! attention: gather the keys by token slot, contract, mask, one softmax, -//! contract. Longer contexts run Furiosa's flash attention: a device loop over -//! chunks of `chunk_keys` keys up to the longest sequence, gathering whole -//! pages and accumulating the softmax statistics in f32. -//! -//! Decode without sinks or windows, up to `decode_keys` keys, instead runs one -//! fusible kernel per layer over the whole pages each token sees, with the -//! token metadata every layer shares computed once per forward pass. +//! Furiosa RNGD paged attention in TCL: f32 softmax and accumulation, BF16 contractions. +//! `defaultPlan` picks fusible BF16 decode kernels, else one-shot attention or a device loop over key chunks. +//! `Repro:` files are in the reproducer repository, platform/furiosa. const std = @import("std"); -const stdx = @import("stdx"); const builder = @import("dsl/tcl/builder"); +const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); const Tensor = zml.Tensor; const tcl = zml.kernel.tcl; -const paged_attention = @import("../paged_attention.zig"); +const AttentionOptions = @import("../options.zig").AttentionOptions; const triton = @import("../triton/triton.zig"); +/// WORKAROUND (TCC 2026.5.0b0): flash decode with 8192-key chunks fails to +/// lower, so decode chunks hold 4096 keys. Repro: repro_flash_decode_8k_chunks.py. +/// Remove once 8192-key chunks compile. const chunk_keys = 4096; +const flash_chunk_pairs = 512 * 2048; +const prefill_chunk_keys = 2048; const prefill_block = 32; +/// WORKAROUND (TCC 2026.5.0b0): block counts that are not a power of two slow TCC's layout +/// search or panic its allocator. Repro: repro_flash_allocator_block_group_panic.py. +/// Remove once any count compiles. +fn prefillBlocks(tokens: i64, sequences: i64, block: i64) i64 { + const blocks = @divTrunc(tokens + sequences * (block - 1), block); + return @intCast(std.math.ceilPowerOfTwo(u64, @intCast(blocks)) catch unreachable); +} + +fn prefillBlock(tokens: i64, sequences: i64) i64 { + const short = prefill_block / 2; + return if (short * prefillBlocks(tokens, sequences, short) < prefill_block * prefillBlocks(tokens, sequences, prefill_block)) short else prefill_block; +} + +/// A straight-line head over each token's first `head` keys, then a device loop over +/// `chunk`-key chunks (none when zero), with tokens in groups of `group`. +pub const Plan = struct { + head: i64, + chunk: i64, + group: i64, +}; + +pub var calibration_plan: ?Plan = null; + pub const paged = struct { pub const Options = triton.paged.Options; pub const Parameters = triton.paged.Parameters; - /// `q` is `{.b = tokens, .hkv, .hg, .hd}`, the caches - /// `{.page, .k_chunk, .hkv, .hd}`. Sinks are one logit per query head. - /// Unwritten cache entries are excluded before the contractions; callers - /// need only populate the tokens covered by their sequence lengths. - pub fn pagedAttention(parameters: Parameters, q: Tensor, k_cache: Tensor, v_cache: Tensor, opts: paged_attention.AttentionOptions) Tensor { + /// `q` is `{.b = tokens, .hkv, .hg, .hd}`, the caches `{.page, .k_chunk, .hkv, .hd}`. + /// Sinks are one logit per query head. + pub fn pagedAttention(parameters: Parameters, q: Tensor, k_cache: Tensor, v_cache: Tensor, opts: AttentionOptions) Tensor { if (opts.mask == .indices) @panic("Sparse attention masks are not supported by TCL paged attention"); stdx.debug.assert(q.rank() == 4 and k_cache.rank() == 4 and v_cache.rank() == 4 and k_cache.shape().eqlDims(v_cache.shape()), "TCL attention requires matching rank-four split K/V caches", .{}); stdx.debug.assert(q.dtype().isFloat() and k_cache.dtype().isFloat() and v_cache.dtype().isFloat(), "TCL attention requires floating-point queries and caches", .{}); @@ -46,7 +59,7 @@ pub const paged = struct { stdx.debug.assert(parameters.seq_lens.dim(.b) > 0 and parameters.block_table.dim(.p) > 0 and k_cache.dim(.page) > 0 and k_cache.dim(.k_chunk) > 0, "TCL attention requires nonempty sequence and cache capacities", .{}); stdx.debug.assert(k_cache.dim(.page) * k_cache.dim(.k_chunk) <= std.math.maxInt(i32) and q.dim(.b) <= std.math.maxInt(i32), "TCL attention indices must fit in I32", .{}); const compiler = zml.Compiler.current(); - stdx.debug.assert(compiler.platform.target == .furiosa and compiler.meshes[0].numDevices() == 1, "TCL attention requires one Furiosa chip", .{}); + stdx.debug.assert(compiler.platform.target == .furiosa, "TCL attention requires Furiosa chips", .{}); const attributes = compiler.platform.devices[0].pjrt_device.attributes(compiler.platform.pjrt_api) catch std.debug.panic("TCL attention could not verify the Furiosa device's PE range", .{}); defer attributes.deinit(); @@ -54,25 +67,33 @@ pub const paged = struct { const last_pe = attributes.get("pe_end"); stdx.debug.assert(first_pe != null and first_pe.? == .int64 and first_pe.?.int64 == 0 and last_pe != null and last_pe.? == .int64 and last_pe.?.int64 == 7, "TCL attention requires a complete 8-PE RNGD chip", .{}); + return attend(parameters, q, k_cache, v_cache, opts); + } + + fn attend(parameters: Parameters, q: Tensor, k_cache: Tensor, v_cache: Tensor, opts: AttentionOptions) Tensor { if (q.dim(.b) == 0) return Tensor.zeroes(q.shape()); const page = k_cache.dim(.k_chunk); - // TCC 2026.5 has no DPE tactic for 32-key attention. Pad to - // at least 64 keys; the extra pages remain invalid and contribute zero. + // WORKAROUND (TCC 2026.5.0b0): no DPE tactic for 32-key attention, so pad to + // at least 64 keys with invalid pages. Repro: repro_decode_32_keys_no_tactic.py. + // Remove once 16- to 48-key decode compiles. const keys = @max(parameters.block_table.dim(.p), std.math.divCeil(i64, 64, page) catch unreachable) * page; - // Chunks hold whole pages and a multiple of 64 keys: TCC 2026.5 fails - // to allocate a chunk of 4080 keys (85 pages of 48). + // WORKAROUND (TCC 2026.5.0b0): TCC fails to allocate a 4080-key chunk (85 + // pages of 48), so chunks hold whole pages and a multiple of 64 keys. Repro: + // repro_flash_allocator_block_group_panic.py. Remove once such chunks compile. const key_alignment = @divExact(page, @as(i64, @intCast(std.math.gcd(@as(u64, @intCast(page)), 64)))) * 64; const alignment = if (key_alignment <= chunk_keys) key_alignment else page; - const chunk = if (keys <= chunk_keys) keys else @min(keys, @max(1, @divTrunc(chunk_keys, alignment)) * alignment); const decode = !parameters.options_.is_prefill and q.dim(.b) == parameters.seq_lens.dim(.b); - const blocks = (std.math.divCeil(i64, q.dim(.b), prefill_block) catch unreachable) + parameters.seq_lens.dim(.b); - // Regular block capacities limit layout-search cost for the fused graph. - // Extra blocks retain zero valid lengths and invalid query positions. - const cfg: Cfg = .{ + const block = if (decode) 1 else prefillBlock(q.dim(.b), parameters.seq_lens.dim(.b)); + const block_count = if (decode) q.dim(.b) else prefillBlocks(q.dim(.b), parameters.seq_lens.dim(.b), block); + const rows = block_count * block; + // A flash chunk's query-key pairs are bounded: the largest chunks do not finish compiling. + const chunk_limit = @max(1, @divTrunc(@min(@as(i64, if (decode) chunk_keys else prefill_chunk_keys), @divTrunc(flash_chunk_pairs, rows)), alignment)) * alignment; + const chunk = @min(keys, chunk_limit); + var cfg: Cfg = .{ .tokens = q.dim(.b), .sequences = parameters.seq_lens.dim(.b), - .b = if (decode) q.dim(.b) else @intCast(std.math.ceilPowerOfTwo(u64, @intCast(blocks)) catch unreachable), - .i = if (decode) 1 else prefill_block, + .b = block_count, + .i = block, .a = (std.math.divCeil(i64, keys, chunk) catch unreachable) * chunk, .chunk = chunk, .hkv = q.dim(.hkv), @@ -86,10 +107,24 @@ pub const paged = struct { .is_causal = opts.mask.isCausal(), .window = @intCast(opts.mask.slidingWindow() orelse 0), }; - if (decode and !cfg.use_sink and cfg.window == 0 and keys <= decode_keys) { - return decodeAttention(parameters, q, k_cache, v_cache, .{ + const plain_decode = decode and !cfg.use_sink and cfg.window == 0; + // WORKAROUND (TCC 2026.5.0b0): an in-unit cache conversion and several tokens + // over a non-power-of-two page count fail to compile, so straight decode reads + // BF16 caches and pads the table. Repro: repro_sparse_propagation_transmute.py, + // repro_metadata_page_count_cover_valid_range.py. Remove once both compile. + const straight = decode and k_cache.dtype() == .bf16 and v_cache.dtype() == .bf16; + const straight_keys = if (straight and cfg.tokens > 1) @as(i64, @intCast(std.math.ceilPowerOfTwo(u64, @intCast(@divExact(keys, page))) catch unreachable)) * page else keys; + const plan = calibration_plan orelse defaultPlan(cfg, straight_keys, parameters.block_table.dim(.p), plain_decode and straight, straight, chunk); + const sink = if (opts.sink) |s| s.convert(.f32).reshape(.{ .hkv = cfg.hkv, .hg = cfg.g }) else Tensor.zeroes(.init(.{ .hkv = cfg.hkv, .hg = cfg.g }, .f32)); + if (plan.head > 0 and plan.chunk > 0) { + stdx.debug.assert(plain_decode and cfg.tokens == 1, "TCL hybrid attention takes one plain decode token", .{}); + return hybridDecodeAttention(parameters, q, k_cache, v_cache, cfg.scale, cfg.is_causal, plan).convert(q.dtype()); + } + if (plan.chunk == 0 and straight) { + return decodeAttention(parameters, q, k_cache, v_cache, sink, .{ .t = cfg.tokens, - .a = keys, + .a = straight_keys, + .group = plan.group, .hkv = cfg.hkv, .g = cfg.g, .d = cfg.d, @@ -97,34 +132,90 @@ pub const paged = struct { .pg = page, .scale = cfg.scale, .is_causal = cfg.is_causal, + .window = cfg.window, + .use_sink = cfg.use_sink, }).convert(q.dtype()); } + cfg.chunk = if (plan.chunk > 0) plan.chunk else keys; + cfg.a = (std.math.divCeil(i64, keys, cfg.chunk) catch unreachable) * cfg.chunk; const table = parameters.block_table.pad(-1, .{ .p = Tensor.Pad{ .high = @divExact(cfg.a, page) - parameters.block_table.dim(.p) } }); + // WORKAROUND (TCC 2026.5.0b0): attention over query groups of a size + // other than a power of two fails ("Inconsistent tensor group + // offsets"): pad the groups with zero queries, whose outputs are + // dropped. Repro: + // repro_flash_prefill_group6_inconsistent_offsets.py (zml/reproducer, + // platform/furiosa). Remove once any group size compiles. + const groups: i64 = @intCast(std.math.ceilPowerOfTwo(u64, @intCast(cfg.g)) catch unreachable); + const extra = groups - cfg.g; + cfg.g = groups; + const queries = q.convert(.bf16).pad(0, .{ .hg = Tensor.Pad{ .high = extra } }); const inputs: OneShot.Inputs = .{ - .q = q.convert(.bf16), + .q = queries, .pages = table, .seq_lens = parameters.seq_lens, .query_offsets = parameters.query_start_len, .k_cache = k_cache.convert(.bf16), .v_cache = v_cache.convert(.bf16), - .sink = if (opts.sink) |s| s.convert(.f32).reshape(.{ .hkv = cfg.hkv, .hg = cfg.g }) else Tensor.zeroes(.init(.{ .hkv = cfg.hkv, .hg = cfg.g }, .f32)), + .sink = sink.pad(0, .{ .hg = Tensor.Pad{ .high = extra } }), }; - const outputs: OneShot.Outputs = .{ .out = q.shape().withDtype(.bf16) }; + const outputs: OneShot.Outputs = .{ .out = queries.shape() }; const out = if (cfg.a == cfg.chunk) OneShot.call(inputs, outputs, .{ .cfg = cfg }).out else Flash.call(inputs, outputs, .{ .cfg = cfg }).out; - return out.convert(q.dtype()); + return out.slice(.hg, .{ .end = groups - extra }).convert(q.dtype()); } }; -/// Keys a decode kernel attends to in one pass: TCC 2026.5 fails on the -/// decode metadata for 131072 keys in pages of 16. +fn defaultPlan(cfg: Cfg, keys: i64, columns: i64, plain_decode: bool, straight: bool, chunk: i64) Plan { + // WORKAROUND (TCC 2026.5.0b0): the hybrid kernel's tail does not compile in + // reasonable time for 2 or 4 KV heads wider than 128, which run flash instead. + // Repro: repro_hybrid_decode_slow_compile.py. Remove once that tail compiles. + if (plain_decode and cfg.tokens == 1 and columns > std.math.divCeil(i64, straight_decode_keys, cfg.pg) catch unreachable) { + if (cfg.d <= 128 or cfg.hkv == 1 or cfg.hkv >= 8) return .{ .head = head_keys, .chunk = tail_chunk_keys, .group = 1 }; + return .{ .head = 0, .chunk = chunk, .group = cfg.b * cfg.i }; + } + if (straight and cfg.window > 0) { + const window_chunk = @max(window_chunk_keys, @as(i64, @intCast(std.math.ceilPowerOfTwo(u64, @intCast(cfg.window)) catch unreachable))); + const chunks: i64 = if (cfg.tokens < chunked_decode_tokens) window_flash_chunks else batch_window_flash_chunks; + if (window_chunk <= chunk and @rem(window_chunk, cfg.pg) == 0 and keys >= chunks * window_chunk) + return .{ .head = 0, .chunk = window_chunk, .group = cfg.tokens }; + } + if (straight and keys <= decode_keys and (cfg.tokens < chunked_decode_tokens or keys <= straight_batch_decode_keys)) + return .{ .head = keys, .chunk = 0, .group = @divExact(cfg.tokens, decodeGroups(cfg.tokens, keys)) }; + if (chunk == keys) return .{ .head = keys, .chunk = 0, .group = cfg.tokens }; + return .{ .head = 0, .chunk = chunk, .group = cfg.b * cfg.i }; +} + +const window_chunk_keys = 1024; +const window_flash_chunks = 16; +const batch_window_flash_chunks = 4; + +/// Widest block table the straight-line decode kernel takes for one token. +/// WORKAROUND (TCC 2026.5.0b0): several tokens do not use it, as TCC cannot +/// propagate gather bounds outside a loop over a token axis. Repro: +/// repro_sparse_propagation_loop.py. Remove once such kernels propagate. +const straight_decode_keys = 32768; + +/// Keys a straight-line decode kernel attends to in one pass. WORKAROUND (TCC +/// 2026.5.0b0): TCC fails on the decode metadata for 131072 keys in pages of 16. +/// Repro: repro_metadata_einsum_by_ve_no_tactic.py. Remove once that compiles. const decode_keys = 65536; +const chunked_decode_tokens = 16; +const straight_batch_decode_keys = 8192; + +/// Decode contracts the gathered cache unmasked: the select gives unused entries (NaN included) zero +/// weight. WORKAROUND (TCC 2026.5.0b0): TCC panics ("slice out of bounds") on score contractions of more +/// tokens times keys than this. Repro: repro_gather_einsum_slice_panic.py. Remove once larger products compile. +const unmasked_decode_keys = 1 << 17; + +const masked_decode_keys = 4096; + const DecodeCfg = struct { t: i64, a: i64, + group: i64, hkv: i64, g: i64, d: i64, @@ -132,14 +223,12 @@ const DecodeCfg = struct { pg: i64, scale: f32, is_causal: bool, + window: i64, + use_sink: bool, }; -/// Decode attention in Furiosa's form: the kernel gathers whole pages, then -/// runs one contraction, a masked softmax and a second contraction. It reads -/// past each sequence's end, selecting zero for those unused cache entries. -/// The token metadata depends only on the attention parameters, which every -/// layer shares: one kernel computes it once per forward pass. -fn decodeAttention(parameters: paged.Parameters, q: Tensor, k_cache: Tensor, v_cache: Tensor, cfg: DecodeCfg) Tensor { +/// The token metadata depends only on parameters every layer shares, so one kernel computes it per forward pass. +fn decodeAttention(parameters: paged.Parameters, q: Tensor, k_cache: Tensor, v_cache: Tensor, sink: Tensor, cfg: DecodeCfg) Tensor { const columns = @divExact(cfg.a, cfg.pg); const table = parameters.block_table.pad(0, .{ .p = Tensor.Pad{ .high = columns - parameters.block_table.dim(.p) } }); const meta = DecodeMeta.call(.{ .seq_lens = parameters.seq_lens, .query_offsets = parameters.query_start_len, .table = table }, .{ @@ -147,7 +236,7 @@ fn decodeAttention(parameters: paged.Parameters, q: Tensor, k_cache: Tensor, v_c .bounds = zml.Shape.init(.{ .t = cfg.t }, .i32), .mask = zml.Shape.init(.{ .t = cfg.t, .a = cfg.a }, .i32), .keyed = zml.Shape.init(.{ .t = cfg.t }, .f32), - }, .{ .cfg = .{ .t = cfg.t, .s = parameters.seq_lens.dim(.b), .ap = columns, .pg = cfg.pg, .pages = cfg.pages, .is_causal = cfg.is_causal } }); + }, .{ .cfg = .{ .t = cfg.t, .s = parameters.seq_lens.dim(.b), .ap = columns, .pg = cfg.pg, .pages = cfg.pages, .is_causal = cfg.is_causal, .window = cfg.window } }); return Decode.call(.{ .q = q.convert(.bf16), .pages = meta.pages, @@ -156,19 +245,15 @@ fn decodeAttention(parameters: paged.Parameters, q: Tensor, k_cache: Tensor, v_c .keyed = meta.keyed, .k_cache = k_cache.convert(.bf16), .v_cache = v_cache.convert(.bf16), + .sink = sink, }, .{ .out = q.shape().withDtype(.bf16) }, .{ .cfg = cfg }).out; } -const MetaCfg = struct { t: i64, s: i64, ap: i64, pg: i64, pages: i64, is_causal: bool }; +const MetaCfg = struct { t: i64, s: i64, ap: i64, pg: i64, pages: i64, is_causal: bool, window: i64 }; -/// For each token: the cache page of each column, how many columns hold keys -/// it sees, the keys it sees (below its limit, in pages inside the cache) and -/// whether it sees any. Columns past the sequence's -/// pages read page `p`, so that the masked padding spreads over many pages -/// instead of hammering one. Rows are selected by one-hot products and -/// comparisons follow `attentionMask`'s clamp arithmetic: TCC 2026.5 has no -/// tactics for small integer gathers and fails on the plugin's lowering of the -/// same comparisons for several tokens. +/// WORKAROUND (TCC 2026.5.0b0): one-hot row selection and `attentionMask`'s clamp arithmetic, as TCC +/// fails on small integer gathers and the plugin's comparisons. Repro: repro_scalar_gather_no_axis_to_pad.py, +/// repro_row_broadcast_past_4096.py. Remove once both compile. const DecodeMeta = tcl.Kernel(MetaCfg, .{ .name = "paged_decode_metadata", .inputs = &.{ "seq_lens", "query_offsets", "table" }, @@ -188,14 +273,12 @@ const DecodeMeta = tcl.Kernel(MetaCfg, .{ }); const ends = try selectMetadata(b, t.query_offsets, try selector(b, try b.arange(cfg.s + 1, .i32, &.{S}, .{ .start = 1 }), O, cfg.s + 1), O); const tokens = try b.arange(cfg.t, .i32, &.{T}, .{}); - // Padding tokens follow every sequence's end and belong to none. const seq = try countBefore(b, ends, tokens, S); const own = try selector(b, seq, S, cfg.s); var op = b.tensorOperation(.{}); const active = try op.commit(op.fetch(seq, .{}).muli(-1).addi(cfg.s).binary(.maxi, 0).binary(.mini, 1), .{}); var visible = try selectMetadata(b, t.seq_lens, own, S); if (cfg.is_causal) { - // Causal tokens see their sequence up to their own position. const end = try selectMetadata(b, ends, own, S); op = b.tensorOperation(.{}); visible = try op.commit(op.fetch(visible, .{}).subi(op.fetch(end, .{})).addi(op.fetch(tokens, .{})).addi(1), .{}); @@ -213,8 +296,6 @@ const DecodeMeta = tcl.Kernel(MetaCfg, .{ const spread = try op.commit(op.fetch(try b.arange(cfg.ap, .i32, &.{AP}, .{}), .{}).binary(.mini, cfg.pages - 1), .{}); op = b.tensorOperation(.{}); const read = try op.commit(op.where(op.fetch(used, .{}), .ne, 0, op.fetch(pages.safe, .{}), op.fetch(spread, .{ .broadcast_to = &.{ T, AP } })), .{}); - // Tokens without keys have no defined attention: the attention - // kernel zeroes them. op = b.tensorOperation(.{}); const keyed = try op.commit(op.fetch(used, .{}).muli(op.fetch(pages.valid, .{})).reduce(&.{AP}, .maxi).toFloat(), .{}); op = b.tensorOperation(.{}); @@ -222,21 +303,19 @@ const DecodeMeta = tcl.Kernel(MetaCfg, .{ op = b.tensorOperation(.{}); const position = try op.commit(op.fetch(try b.arange(cfg.ap * cfg.pg, .i32, &.{A}, .{}), .{ .broadcast_to = &.{ T, A } }), .{}); op = b.tensorOperation(.{ .tactic = .elementwise }); - const mask = try op.commit(op.fetch(position, .{}).muli(-1).addi(op.fetch(limit, .{})).binary(.maxi, 0).binary(.mini, 1).muli(op.fetch(try b.reshape(valid, &.{ T, A }, .{}), .{})), .{}); + var mask = try op.commit(op.fetch(position, .{}).muli(-1).addi(op.fetch(limit, .{})).binary(.maxi, 0).binary(.mini, 1).muli(op.fetch(try b.reshape(valid, &.{ T, A }, .{}), .{})), .{}); + if (cfg.window > 0) { + op = b.tensorOperation(.{ .tactic = .elementwise }); + mask = try op.commit(op.fetch(position, .{}).subi(op.fetch(limit, .{})).addi(cfg.window + 1).binary(.maxi, 0).binary(.mini, 1).muli(op.fetch(mask, .{})), .{}); + } b.ret(&.{ read, bounds, mask, keyed }); } }.run, }); -/// Gathers the key and value pages a token sees, contracts with the query, -/// masks keys past the sequence, normalizes in one softmax times `keyed` (zero -/// for tokens without keys, whose softmax is uniform) and contracts with the -/// values. A token's pages are gathered up to its bound, and the operations -/// reading them skip the rest: the work follows the sequence's length, not the -/// table's. Fusible: it joins the layer around it, one launch per layer. const Decode = tcl.Kernel(DecodeCfg, .{ .name = "paged_decode_attention", - .inputs = &.{ "q", "pages", "bounds", "mask", "keyed", "k_cache", "v_cache" }, + .inputs = &.{ "q", "pages", "bounds", "mask", "keyed", "k_cache", "v_cache", "sink" }, .outputs = &.{"out"}, .run = struct { fn run(b: *tcl.Builder, cfg: DecodeCfg) tcl.FinishError!void { @@ -258,26 +337,360 @@ const Decode = tcl.Kernel(DecodeCfg, .{ .keyed = .{ .dtype = .f32, .axes = &.{T} }, .k_cache = .{ .dtype = .bf16, .axes = &.{ P, PG, Hkv, D } }, .v_cache = .{ .dtype = .bf16, .axes = &.{ P, PG, Hkv, D } }, + .sink = .{ .dtype = .f32, .axes = &.{ Hkv, G } }, }); - const mask = try b.reshape(t.mask, &.{ T, AP, PG }, .{}); - const k = try maskCacheValues(b, try b.gather(t.k_cache, t.pages, P, &.{ T, AP, PG, Hkv, D }, .{ .valid_length = t.bounds }), mask); - const v = try maskCacheValues(b, try b.gather(t.v_cache, t.pages, P, &.{ T, AP, PG, Hkv, D }, .{ .valid_length = t.bounds }), mask); - const keys: []const tcl.Axis = &.{ AP, PG }; - var op = b.tensorOperation(.{}); - const scores = try op.commit(op.contract(t.q, k, &.{ Hkv, G, T, AP, PG }).mulf(cfg.scale), .{}); - op = b.tensorOperation(.{}); - const s = op.fetch(scores, .{}); - const masked = try op.commit(op.where(op.fetch(mask, .{}), .ne, 0, s, score_min), .{}); - op = b.tensorOperation(.{}); - const x = op.fetch(masked, .{}); - const e = x.subf(x.reduce(keys, .maxf)).exp(); - const p = try op.commit(e.divf(e.reduce(keys, .addf)).mulf(op.fetch(t.keyed, .{})), .{ .dtype = .bf16 }); - op = b.tensorOperation(.{}); - b.ret(&.{try op.commit(op.contract(v, p, &.{ T, Hkv, G, D }), .{ .dtype = .bf16 })}); + const sink = if (cfg.use_sink) t.sink else null; + const groups = @divExact(cfg.t, cfg.group); + if (groups == 1) return b.ret(&.{try decodeTokens(b, cfg.scale, t.q, t.pages, t.bounds, t.mask, t.keyed, t.k_cache, t.v_cache, sink)}); + const NG = b.axis("NG", groups); + const TG = b.axis("TG", @divExact(cfg.t, groups)); + var loop = b.openFor(groups, .{try b.scratchpad(.bf16, &.{ NG, TG, Hkv, G, D })}); + { + const q = try b.indexRead(try b.reshape(t.q, &.{ NG, TG, Hkv, G, D }, .{}), loop.iv); + const pages = try b.indexRead(try b.reshape(t.pages, &.{ NG, TG, AP }, .{}), loop.iv); + const bounds = try b.indexRead(try b.reshape(t.bounds, &.{ NG, TG }, .{}), loop.iv); + const mask = try b.indexRead(try b.reshape(t.mask, &.{ NG, TG, A }, .{}), loop.iv); + const keyed = try b.indexRead(try b.reshape(t.keyed, &.{ NG, TG }, .{}), loop.iv); + const out = try decodeTokens(b, cfg.scale, q, pages, bounds, mask, keyed, t.k_cache, t.v_cache, sink); + try loop.yield(.{try b.indexWrite(loop.carried[0], loop.iv, out)}); + } + b.ret(&.{try b.reshape(loop.results[0], &.{ T, Hkv, G, D }, .{})}); } }.run, }); +const unmasked_group_tokens = 4; + +/// Groups of at most `unmasked_group_tokens` tokens whose tokens times keys fit `unmasked_decode_keys`. +fn decodeGroups(tokens: i64, keys: i64) i64 { + var groups: i64 = 1; + while ((@divExact(tokens, groups) > unmasked_group_tokens or @divExact(tokens, groups) * keys > unmasked_decode_keys) and groups < tokens) { + groups += 1; + while (@rem(tokens, groups) != 0) groups += 1; + } + return groups; +} + +/// `keyed` scales the softmax: zero for tokens without keys. +fn decodeTokens(b: *tcl.Builder, scale: f32, q: builder.Tensor, pages: builder.Tensor, bounds: builder.Tensor, mask: builder.Tensor, keyed: builder.Tensor, k_cache: builder.Tensor, v_cache: builder.Tensor, sink: ?builder.Tensor) tcl.FinishError!builder.Tensor { + const T = q.axes[0]; + const Hkv = q.axes[1]; + const G = q.axes[2]; + const D = q.axes[3]; + const AP = pages.axes[1]; + const P = k_cache.axes[0]; + const PG = k_cache.axes[1]; + const keys: []const tcl.Axis = &.{ AP, PG }; + const key_mask = try b.reshape(mask, &.{ T, AP, PG }, .{}); + var k = try b.gather(k_cache, pages, P, &.{ T, AP, PG, Hkv, D }, .{ .valid_length = bounds }); + var v = try b.gather(v_cache, pages, P, &.{ T, AP, PG, Hkv, D }, .{ .valid_length = bounds }); + if (T.size == 1 and AP.size * PG.size <= masked_decode_keys) { + k = try maskCacheValues(b, k, key_mask); + v = try maskCacheValues(b, v, key_mask); + } + var op = b.tensorOperation(.{}); + const scores = try op.commit(op.contract(q, k, &.{ Hkv, G, T, AP, PG }).mulf(scale), .{}); + op = b.tensorOperation(.{}); + const s = op.fetch(scores, .{}); + const masked = try op.commit(op.where(op.fetch(key_mask, .{}), .ne, 0, s, score_min), .{}); + op = b.tensorOperation(.{}); + const p = if (sink) |logits| blk: { + const maximum = try op.commit(op.fetch(masked, .{}).reduce(keys, .maxf).maxf(op.fetch(logits, .{ .broadcast_to = &.{ Hkv, G, T } })), .{}); + op = b.tensorOperation(.{}); + const e = try op.commit(op.fetch(masked, .{}).subf(op.fetch(maximum, .{})).exp(), .{}); + op = b.tensorOperation(.{}); + const sink_exp = try op.commit(op.fetch(maximum, .{}).mulf(-1).addf(op.fetch(logits, .{})).exp(), .{}); + op = b.tensorOperation(.{}); + const denominator = try op.commit(op.fetch(e, .{}).reduce(keys, .addf).addf(op.fetch(sink_exp, .{})), .{}); + op = b.tensorOperation(.{}); + break :blk try op.commit(op.fetch(e, .{}).divf(op.fetch(denominator, .{})).mulf(op.fetch(keyed, .{})), .{ .dtype = .bf16 }); + } else blk: { + const x = op.fetch(masked, .{}); + const e = x.subf(x.reduce(keys, .maxf)).exp(); + break :blk try op.commit(e.divf(e.reduce(keys, .addf)).mulf(op.fetch(keyed, .{})), .{ .dtype = .bf16 }); + }; + op = b.tensorOperation(.{}); + return op.commit(op.contract(v, p, &.{ T, Hkv, G, D }), .{ .dtype = .bf16 }); +} + +const head_keys = 4096; +const tail_chunk_keys = 8192; + +const HybridCfg = struct { + t: i64, + s: i64, + /// Head pages, tail chunks and pages per tail chunk. + hp: i64, + nc: i64, + cp: i64, + pg: i64, + hkv: i64, + g: i64, + d: i64, + pages: i64, + scale: f32, + is_causal: bool, +}; + +fn hybridDecodeAttention(parameters: paged.Parameters, q: Tensor, k_cache: Tensor, v_cache: Tensor, scale: f32, is_causal: bool, plan: Plan) Tensor { + const pg = k_cache.dim(.k_chunk); + const t = q.dim(.b); + const s = parameters.seq_lens.dim(.b); + const hp = std.math.divCeil(i64, plan.head, pg) catch unreachable; + const cp = std.math.divCeil(i64, plan.chunk, pg) catch unreachable; + const nc = std.math.divCeil(i64, parameters.block_table.dim(.p) - hp, cp) catch unreachable; + const cfg: HybridCfg = .{ .t = t, .s = s, .hp = hp, .nc = nc, .cp = cp, .pg = pg, .hkv = q.dim(.hkv), .g = q.dim(.hg), .d = q.dim(.hd), .pages = k_cache.dim(.page), .scale = scale, .is_causal = is_causal }; + const table = parameters.block_table.pad(-1, .{ .p = Tensor.Pad{ .high = hp + nc * cp - parameters.block_table.dim(.p) } }); + const meta = HybridMeta.call(.{ .seq_lens = parameters.seq_lens, .query_offsets = parameters.query_start_len, .table = table }, .{ + .hpages = zml.Shape.init(.{ .t = t, .p = hp }, .i32), + .hmask = zml.Shape.init(.{ .t = t, .a = hp * pg }, .i32), + .hbounds = zml.Shape.init(.{ .t = t }, .i32), + .tpages = zml.Shape.init(.{ .c = nc, .t = t, .p = cp }, .i32), + .tmask = zml.Shape.init(.{ .c = nc, .t = t, .a = cp * pg }, .i32), + .tbounds = zml.Shape.init(.{ .c = nc, .t = t }, .i32), + .limits = zml.Shape.init(.{ .t = t }, .i32), + .keyed = zml.Shape.init(.{ .t = t }, .f32), + }, .{ .cfg = cfg }); + return HybridDecode.call(.{ + .q = q.convert(.bf16), + .hpages = meta.hpages, + .hmask = meta.hmask, + .hbounds = meta.hbounds, + .tpages = meta.tpages, + .tmask = meta.tmask, + .tbounds = meta.tbounds, + .limits = meta.limits, + .keyed = meta.keyed, + .k_cache = k_cache.convert(.bf16), + .v_cache = v_cache.convert(.bf16), + }, .{ .out = q.shape().withDtype(.bf16) }, .{ .cfg = cfg }).out; +} + +/// Per token, once per forward pass. Tail chunks are chunk-major, only those holding keys of the +/// longest token; `limits` is the visible key count. +const HybridMeta = tcl.Kernel(HybridCfg, .{ + .name = "paged_hybrid_metadata", + .inputs = &.{ "seq_lens", "query_offsets", "table" }, + .outputs = &.{ "hpages", "hmask", "hbounds", "tpages", "tmask", "tbounds", "limits", "keyed" }, + .run = struct { + fn run(b: *tcl.Builder, cfg: HybridCfg) tcl.FinishError!void { + const total = (cfg.hp + cfg.nc * cfg.cp) * cfg.pg; + const T = b.axis("T", cfg.t); + const S = b.axis("S", cfg.s); + const O = b.axis("O", cfg.s + 1); + const AP = b.axis("AP", cfg.hp + cfg.nc * cfg.cp); + const HP = b.axis("HP", cfg.hp); + const RP = b.axis("RP", cfg.nc * cfg.cp); + const NC = b.axis("NC", cfg.nc); + const CP = b.axis("CP", cfg.cp); + const HA = b.axis("HA", cfg.hp * cfg.pg); + const CK = b.axis("CK", cfg.cp * cfg.pg); + const t = try b.declareArgs(.{ + .seq_lens = .{ .dtype = .i32, .axes = &.{S} }, + .query_offsets = .{ .dtype = .i32, .axes = &.{O} }, + .table = .{ .dtype = .i32, .axes = &.{ S, AP } }, + }); + const ends = try selectMetadata(b, t.query_offsets, try selector(b, try b.arange(cfg.s + 1, .i32, &.{S}, .{ .start = 1 }), O, cfg.s + 1), O); + const tokens = try b.arange(cfg.t, .i32, &.{T}, .{}); + const seq = try countBefore(b, ends, tokens, S); + const own = try selector(b, seq, S, cfg.s); + var op = b.tensorOperation(.{}); + const active = try op.commit(op.fetch(seq, .{}).muli(-1).addi(cfg.s).binary(.maxi, 0).binary(.mini, 1), .{}); + var visible = try selectMetadata(b, t.seq_lens, own, S); + if (cfg.is_causal) { + const end = try selectMetadata(b, ends, own, S); + op = b.tensorOperation(.{}); + visible = try op.commit(op.fetch(visible, .{}).subi(op.fetch(end, .{})).addi(op.fetch(tokens, .{})).addi(1), .{}); + } + op = b.tensorOperation(.{}); + const limit = try op.commit(op.fetch(visible, .{}).binary(.maxi, 0).binary(.mini, total).muli(op.fetch(active, .{})), .{}); + + // WORKAROUND (TCC 2026.5.0b0): a 3-D one-hot product computes garbage when the one-hot is read + // first, so the selection is 2-D per chunk. Repro: repro_onehot_einsum_by_ve_wrong_values_run.py. + // Remove once it is exact on RNGD. + const rows1 = try b.reshape(try b.slice(t.table, AP, cfg.hp, &.{ S, RP }, .{}), &.{ S, NC, CP }, .{}); + op = b.tensorOperation(.{}); + const rows = try op.commit(op.fetch(rows1, .{}), .{ .axes = &.{ NC, S, CP } }); + // Trips never exceed nc: `limit` is at most the table's keys. + const trips = try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(limit), cfg.cp * cfg.pg - 1 - cfg.hp * cfg.pg), cfg.cp * cfg.pg); + var loop = b.openFor(trips, .{ try b.scratchpad(.i32, &.{ NC, T, CP }), try b.scratchpad(.i32, &.{ NC, T, CK }), try b.scratchpad(.i32, &.{ NC, T }), try b.full(0, .i32, &.{T}) }); + { + const reads, const masks, const counts, const seen = loop.carried; + const first = try b.symExpr(.add, try b.symExpr(.mul, loop.iv, cfg.cp * cfg.pg), cfg.hp * cfg.pg); + const chunk = try hybridColumns(b, cfg, own, try b.indexRead(rows, loop.iv), limit, first, CP, CK); + op = b.tensorOperation(.{}); + const any = try op.commit(op.fetch(seen, .{}).binary(.maxi, op.fetch(chunk.any, .{})), .{}); + try loop.yield(.{ + try b.indexWrite(reads, loop.iv, chunk.pages), + try b.indexWrite(masks, loop.iv, chunk.mask), + try b.indexWrite(counts, loop.iv, chunk.bounds), + any, + }); + } + const head = try hybridColumns(b, cfg, own, try b.slice(t.table, AP, 0, &.{ S, HP }, .{}), limit, 0, HP, HA); + op = b.tensorOperation(.{}); + const keyed = try op.commit(op.fetch(head.any, .{}).binary(.maxi, op.fetch(loop.results[3], .{})).toFloat(), .{}); + b.ret(&.{ head.pages, head.mask, head.bounds, loop.results[0], loop.results[1], loop.results[2], limit, keyed }); + } + }.run, +}); + +/// Columns start at key `first`; sparse gathers never read past a token's `bounds`. +fn hybridColumns(b: *tcl.Builder, cfg: HybridCfg, own: builder.Tensor, table: builder.Tensor, limit: builder.Tensor, first: anytype, cols: tcl.Axis, keys: tcl.Axis) tcl.FinishError!struct { pages: builder.Tensor, mask: builder.Tensor, bounds: builder.Tensor, any: builder.Tensor } { + const T = own.axes[0]; + const S = own.axes[1]; + const PG = b.axis("PG", cfg.pg); + const pages = try validIndices(b, try selectMetadata(b, table, own, S), cfg.pages); + var op = b.tensorOperation(.{}); + const firsts = try op.commit(op.fetch(try b.arange(cols.size * cfg.pg, .i32, &.{cols}, .{ .step = cfg.pg }), .{}).addi(if (@TypeOf(first) == builder.Tensor) op.fetch(first, .{}) else first), .{}); + op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); + const used = try op.commit(op.fetch(limit, .{}).subi(op.fetch(firsts, .{})).binary(.maxi, 0).binary(.mini, 1), .{}); + op = b.tensorOperation(.{}); + const bounds = try op.commit(op.fetch(used, .{}).reduce(&.{cols}, .addi), .{}); + op = b.tensorOperation(.{}); + const any = try op.commit(op.fetch(used, .{}).muli(op.fetch(pages.valid, .{})).reduce(&.{cols}, .maxi), .{}); + op = b.tensorOperation(.{}); + const key_ids = try op.commit(op.fetch(try b.arange(keys.size, .i32, &.{keys}, .{}), .{}).addi(if (@TypeOf(first) == builder.Tensor) op.fetch(first, .{}) else first), .{}); + const to: []const tcl.Axis = &.{ T, cols, PG }; + op = b.tensorOperation(.{}); + const position = try op.commit(op.fetch(try b.reshape(key_ids, &.{ cols, PG }, .{}), .{ .broadcast_to = to }), .{}); + op = b.tensorOperation(.{}); + const valid = try op.commit(op.fetch(pages.valid, .{ .broadcast_to = to }), .{}); + op = b.tensorOperation(.{ .tactic = .elementwise }); + const mask = try op.commit(op.fetch(position, .{}).muli(-1).addi(op.fetch(limit, .{})).binary(.maxi, 0).binary(.mini, 1).muli(op.fetch(valid, .{})), .{}); + return .{ .pages = pages.safe, .mask = try b.reshape(mask, &.{ T, keys }, .{}), .bounds = bounds, .any = any }; +} + +const SoftmaxState = struct { max: builder.Tensor, sum: builder.Tensor, out: builder.Tensor }; + +const HybridDecode = tcl.Kernel(HybridCfg, .{ + .name = "paged_hybrid_decode_attention", + .inputs = &.{ "q", "hpages", "hmask", "hbounds", "tpages", "tmask", "tbounds", "limits", "keyed", "k_cache", "v_cache" }, + .outputs = &.{"out"}, + .run = struct { + fn run(b: *tcl.Builder, cfg: HybridCfg) tcl.FinishError!void { + b.fuse(); + const a = hybridAxes(b, cfg); + const t = try b.declareArgs(.{ + .q = .{ .dtype = .bf16, .axes = &.{ a.T, a.Hkv, a.G, a.D } }, + .hpages = .{ .dtype = .i32, .axes = &.{ a.T, a.HP } }, + .hmask = .{ .dtype = .i32, .axes = &.{ a.T, a.HA } }, + .hbounds = .{ .dtype = .i32, .axes = &.{a.T} }, + .tpages = .{ .dtype = .i32, .axes = &.{ a.NC, a.T, a.CP } }, + .tmask = .{ .dtype = .i32, .axes = &.{ a.NC, a.T, a.CK } }, + .tbounds = .{ .dtype = .i32, .axes = &.{ a.NC, a.T } }, + .limits = .{ .dtype = .i32, .axes = &.{a.T} }, + .keyed = .{ .dtype = .f32, .axes = &.{a.T} }, + .k_cache = .{ .dtype = .bf16, .axes = &.{ a.P, a.PG, a.Hkv, a.D } }, + .v_cache = .{ .dtype = .bf16, .axes = &.{ a.P, a.PG, a.Hkv, a.D } }, + }); + const tail = try hybridTail(b, cfg, a, t.q, t.tpages, t.tmask, t.tbounds, t.limits, t.k_cache, t.v_cache); + b.ret(&.{try hybridHead(b, cfg, a, t.q, t.hpages, t.hmask, t.hbounds, t.keyed, t.k_cache, t.v_cache, tail)}); + } + }.run, +}); + +const HybridAxes = struct { T: tcl.Axis, HP: tcl.Axis, HA: tcl.Axis, NC: tcl.Axis, CP: tcl.Axis, CK: tcl.Axis, Hkv: tcl.Axis, G: tcl.Axis, D: tcl.Axis, P: tcl.Axis, PG: tcl.Axis }; + +fn hybridAxes(b: *tcl.Builder, cfg: HybridCfg) HybridAxes { + return .{ + .T = b.axis("T", cfg.t), + .HP = b.axis("HP", cfg.hp), + .HA = b.axis("HA", cfg.hp * cfg.pg), + .NC = b.axis("NC", cfg.nc), + .CP = b.axis("CP", cfg.cp), + .CK = b.axis("CK", cfg.cp * cfg.pg), + .Hkv = b.axis("Hkv", cfg.hkv), + .G = b.axis("G", cfg.g), + .D = b.axis("D", cfg.d), + .P = b.axis("P", cfg.pages), + .PG = b.axis("PG", cfg.pg), + }; +} + +fn hybridTail(b: *tcl.Builder, cfg: HybridCfg, a: HybridAxes, q: builder.Tensor, tpages: builder.Tensor, tmask: builder.Tensor, tbounds: builder.Tensor, limits: builder.Tensor, k_cache: builder.Tensor, v_cache: builder.Tensor) tcl.FinishError!SoftmaxState { + const stats: []const tcl.Axis = &.{ a.Hkv, a.G, a.T }; + const trips = try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(limits), cfg.cp * cfg.pg - 1 - cfg.hp * cfg.pg), cfg.cp * cfg.pg); + var loop = b.openFor(trips, .{ try b.full(score_min, .f32, stats), try b.full(0, .f32, stats), try b.full(0, .f32, &.{ a.T, a.Hkv, a.G, a.D }) }); + { + const max, const sum, const out = loop.carried; + const mask = try b.reshape(try b.indexRead(tmask, loop.iv), &.{ a.T, a.CP, a.PG }, .{}); + const bounds = try b.indexRead(tbounds, loop.iv); + const pages = try b.indexRead(tpages, loop.iv); + const k = try b.gather(k_cache, pages, a.P, &.{ a.T, a.CP, a.PG, a.Hkv, a.D }, .{ .valid_length = bounds }); + const v = try b.gather(v_cache, pages, a.P, &.{ a.T, a.CP, a.PG, a.Hkv, a.D }, .{ .valid_length = bounds }); + const merged = try mergeSoftmax(b, .{ .max = max, .sum = sum, .out = out }, try chunkSoftmax(b, cfg, q, k, v, mask)); + try loop.yield(.{ merged.max, merged.sum, merged.out }); + } + return .{ .max = loop.results[0], .sum = loop.results[1], .out = loop.results[2] }; +} + +fn hybridHead(b: *tcl.Builder, cfg: HybridCfg, a: HybridAxes, q: builder.Tensor, hpages: builder.Tensor, hmask: builder.Tensor, hbounds: builder.Tensor, keyed: builder.Tensor, k_cache: builder.Tensor, v_cache: builder.Tensor, tail: SoftmaxState) tcl.FinishError!builder.Tensor { + const head_mask = try b.reshape(hmask, &.{ a.T, a.HP, a.PG }, .{}); + const head_k = try b.gather(k_cache, hpages, a.P, &.{ a.T, a.HP, a.PG, a.Hkv, a.D }, .{ .valid_length = hbounds }); + const head_v = try b.gather(v_cache, hpages, a.P, &.{ a.T, a.HP, a.PG, a.Hkv, a.D }, .{ .valid_length = hbounds }); + const keys: []const tcl.Axis = &.{ a.HP, a.PG }; + var op = b.tensorOperation(.{}); + const scores = try op.commit(op.contract(q, head_k, &.{ a.Hkv, a.G, a.T, a.HP, a.PG }).mulf(cfg.scale), .{}); + op = b.tensorOperation(.{}); + const sv = op.fetch(scores, .{}); + const masked = try op.commit(op.where(op.fetch(head_mask, .{}), .ne, 0, sv, score_min), .{}); + op = b.tensorOperation(.{}); + const max = try op.commit(op.fetch(masked, .{}).reduce(keys, .maxf).maxf(op.fetch(tail.max, .{})), .{}); + const alpha = try expDifference(b, tail.max, max); + op = b.tensorOperation(.{}); + const carried = try op.commit(op.fetch(tail.sum, .{}).mulf(op.fetch(alpha, .{})), .{}); + op = b.tensorOperation(.{}); + const denominator = try op.commit(op.fetch(masked, .{}).subf(op.fetch(max, .{})).exp().reduce(keys, .addf).addf(op.fetch(carried, .{})), .{}); + op = b.tensorOperation(.{}); + const p = try op.commit(op.fetch(masked, .{}).subf(op.fetch(max, .{})).exp().divf(op.fetch(denominator, .{})).mulf(op.fetch(keyed, .{})), .{ .dtype = .bf16 }); + op = b.tensorOperation(.{}); + const head = try op.commit(op.contract(head_v, p, &.{ a.T, a.Hkv, a.G, a.D }), .{}); + op = b.tensorOperation(.{}); + const factor = try op.commit(op.fetch(alpha, .{}).divf(op.fetch(denominator, .{})), .{}); + op = b.tensorOperation(.{}); + return op.commit(op.fetch(tail.out, .{}).mulf(op.fetch(factor, .{})).addf(op.fetch(head, .{})).mulf(op.fetch(keyed, .{})), .{ .dtype = .bf16 }); +} + +fn chunkSoftmax(b: *tcl.Builder, cfg: HybridCfg, q: builder.Tensor, k: builder.Tensor, v: builder.Tensor, mask: builder.Tensor) tcl.FinishError!SoftmaxState { + const T = q.axes[0]; + const Hkv = q.axes[1]; + const G = q.axes[2]; + const D = q.axes[3]; + const keys = mask.axes[1..]; + var op = b.tensorOperation(.{}); + const scores = try op.commit(op.contract(q, k, &.{ Hkv, G, T, keys[0], keys[1] }).mulf(cfg.scale), .{}); + op = b.tensorOperation(.{}); + const sv = op.fetch(scores, .{}); + const masked = try op.commit(op.where(op.fetch(mask, .{}), .ne, 0, sv, score_min), .{}); + op = b.tensorOperation(.{}); + const max = try op.commit(op.fetch(masked, .{}).reduce(keys, .maxf), .{}); + const exps = try maskedExp(b, masked, max, mask); + op = b.tensorOperation(.{}); + const sum = try op.commit(op.fetch(exps, .{}).reduce(keys, .addf), .{}); + op = b.tensorOperation(.{}); + const p = try op.commit(op.fetch(exps, .{}).mulf(1.0), .{ .dtype = .bf16 }); + op = b.tensorOperation(.{}); + const out = try op.commit(op.contract(v, p, &.{ T, Hkv, G, D }), .{}); + return .{ .max = max, .sum = sum, .out = out }; +} + +fn mergeSoftmax(b: *tcl.Builder, x: SoftmaxState, y: SoftmaxState) tcl.FinishError!SoftmaxState { + var op = b.tensorOperation(.{ .tactic = .elementwise }); + const max = try op.commit(op.fetch(x.max, .{}).maxf(op.fetch(y.max, .{})), .{}); + const alpha = try expDifference(b, x.max, max); + const beta = try expDifference(b, y.max, max); + op = b.tensorOperation(.{}); + const xs = try op.commit(op.fetch(x.sum, .{}).mulf(op.fetch(alpha, .{})), .{}); + op = b.tensorOperation(.{}); + const sum = try op.commit(op.fetch(y.sum, .{}).mulf(op.fetch(beta, .{})).addf(op.fetch(xs, .{})), .{}); + op = b.tensorOperation(.{}); + const xo = try op.commit(op.fetch(x.out, .{}).mulf(op.fetch(alpha, .{})), .{}); + op = b.tensorOperation(.{}); + const out = try op.commit(op.fetch(y.out, .{}).mulf(op.fetch(beta, .{})).addf(op.fetch(xo, .{})), .{}); + return .{ .max = max, .sum = sum, .out = out }; +} + const Cfg = struct { tokens: i64, sequences: i64, @@ -299,11 +712,9 @@ const Cfg = struct { window: i32 = -1, }; -/// Attention tuning inherited from the original TCL backend, informed by -/// `furiosa.kernels.llama.config` and `common.attention`. Keep the compiler's -/// default lowering mode and omit the adaptive-einsum and half-VRF settings -/// removed in TCC 2026.5. +/// Prefill bounds the tactic shapes searched (fewer for narrow heads): the compute-bound hint compiles too slowly. fn configure(b: *tcl.Builder, cfg: Cfg) void { + const shapes: i64 = if (cfg.hkv * cfg.g * cfg.d >= 1024) 16 else 8; if (cfg.decode) b.compilerConfig(.{ .enable_einsum_fusion = true, .padding_policy = .{ .Small = 1.1 }, @@ -326,11 +737,10 @@ fn configure(b: *tcl.Builder, cfg: Cfg) void { .reshape_einsum_mode = .{ .Reshape = .{ .permute = false } }, .tensor_unit_bridge_threshold_in_page = 12, .allow_external_operators = false, - .enable_tactic_pruning = false, .scheduler_beam_search = true, .allow_reduce_by_ve_cluster_chip_reduce = false, .allow_reduce_by_ve_cluster_chip_reduce_base_population = false, - .tactic_hint = .{ .ForLlmModelComputeBound = 200 }, + .tactic_hint = .{ .DefaultWithMaxNumUniqueShapes = shapes }, .dma_preference = 1.2, }); } @@ -390,11 +800,24 @@ fn declareInputs(b: *tcl.Builder, a: Axes) tcl.FinishError!NativeInputs { const Prepared = struct { q: builder.Tensor, + /// Page table rows: one per block for decode, one per sequence for prefill (blocks select theirs with `pick`). pages: builder.Tensor, + pick: ?builder.Tensor, qpos: builder.Tensor, kvlen: builder.Tensor, row: builder.Tensor, active: builder.Tensor, + /// Per page row: the first query position and the keys. + first: builder.Tensor, + keys: builder.Tensor, + + /// WORKAROUND (TCC 2026.5.0b0): TCC has no tactic for selecting whole rows of 8192 pages, so flash + /// attention selects per chunk. Repro: repro_metadata_einsum_by_ve_no_tactic.py. Remove once such + /// products compile. + fn blockPages(self: Prepared, b: *tcl.Builder, a: Axes, pages: builder.Tensor) tcl.FinishError!builder.Tensor { + const pick = self.pick orelse return pages; + return selectMetadata(b, pages, pick, a.S); + } }; fn countBefore(b: *tcl.Builder, ends: builder.Tensor, positions: builder.Tensor, axis: tcl.Axis) tcl.FinishError!builder.Tensor { @@ -414,15 +837,16 @@ fn selector(b: *tcl.Builder, indices: builder.Tensor, axis: tcl.Axis, count: i64 return op.commit(op.where(distance, .eq, 0, 1, 0), .{}); } -// Keep metadata exact in I32. Scalar gathers and four-byte-offset slices do -// not have native tactics for these shapes in TCC 2026.5. +// Keep metadata exact in I32. WORKAROUND (TCC 2026.5.0b0): no native tactics for +// scalar gathers or four-byte-offset slices. Repro: repro_scalar_gather_no_axis_to_pad.py, +// repro_i32_slice_odd_offset.py. Remove once both compile. fn selectMetadata(b: *tcl.Builder, values: builder.Tensor, selected: builder.Tensor, axis: tcl.Axis) tcl.FinishError!builder.Tensor { const op = b.tensorOperation(.{ .tactic = if (values.axes.len > 1) .einsum_by_ve else .reduce_by_ve }); return op.commit(op.fetch(selected, .{}).muli(op.fetch(values, .{})).reduce(&.{axis}, .addi), .{}); } -/// Keep adjacent I32 reads inside the native unit. Exposing offsets[1:] as a -/// separate argument would require alignment staging at every invocation. +/// Reads adjacent I32 offsets inside the native unit: a separate `offsets[1:]` argument would +/// need alignment staging at every invocation. fn prepareQueries(b: *tcl.Builder, cfg: Cfg, a: Axes, t: NativeInputs) tcl.FinishError!Prepared { const start_indices = try b.arange(cfg.sequences, .i32, &.{a.S}, .{}); const end_indices = try b.arange(cfg.sequences + 1, .i32, &.{a.S}, .{ .start = 1 }); @@ -432,8 +856,7 @@ fn prepareQueries(b: *tcl.Builder, cfg: Cfg, a: Axes, t: NativeInputs) tcl.Finis const qlen = try op.commit(op.fetch(ends, .{}).subi(op.fetch(starts, .{})), .{}); const tokens = try b.arange(cfg.tokens, .i32, &.{a.T}, .{}); const seq_of_token = try countBefore(b, ends, tokens, a.S); - // All sequence ends precede a padding token. Reuse this vector instead of - // bringing a symbolic scalar reduction back into a vector operation. + // Padding tokens follow all sequence ends; this vector avoids a symbolic scalar reduction in a vector operation. op = b.tensorOperation(.{}); const token_tail = try op.commit(op.fetch(seq_of_token, .{}).subi(cfg.sequences), .{}); op = b.tensorOperation(.{}); @@ -453,16 +876,19 @@ fn prepareQueries(b: *tcl.Builder, cfg: Cfg, a: Axes, t: NativeInputs) tcl.Finis return .{ .q = try b.reshape(q, &.{ a.B, a.I, a.Hkv, a.G, a.D }, .{}), .pages = try b.reshape(t.pages, &.{ a.B, a.AP }, .{}), + .pick = null, .qpos = try b.reshape(qpos, &.{ a.B, a.I }, .{}), .kvlen = try b.reshape(lengths, &.{a.B}, .{}), .row = token_seq, .active = active, + .first = try b.reshape(qpos, &.{a.B}, .{}), + .keys = try b.reshape(lengths, &.{a.B}, .{}), }; } - const shift = std.math.log2_int(u64, prefill_block); + const shift = std.math.log2_int(u64, @intCast(cfg.i)); op = b.tensorOperation(.{}); - const blocks = try op.commit(op.fetch(qlen, .{}).addi(prefill_block - 1).binary(.shr_arith, shift), .{}); + const blocks = try op.commit(op.fetch(qlen, .{}).addi(cfg.i - 1).binary(.shr_arith, shift), .{}); op = b.tensorOperation(.{}); const block_end = try op.commit(op.fetch(blocks, .{}).cumsum(&.{a.S}), .{}); op = b.tensorOperation(.{}); @@ -482,7 +908,7 @@ fn prepareQueries(b: *tcl.Builder, cfg: Cfg, a: Axes, t: NativeInputs) tcl.Finis const key_length = try selectMetadata(b, t.seq_lens, pick, a.S); const local_rows = try b.arange(cfg.i, .i32, &.{a.I}, .{}); op = b.tensorOperation(.{}); - const first_row = try op.commit(op.fetch(block_ids, .{}).subi(op.fetch(block_begin, .{})).muli(prefill_block), .{}); + const first_row = try op.commit(op.fetch(block_ids, .{}).subi(op.fetch(block_begin, .{})).muli(cfg.i), .{}); op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); const qi = try op.commit(op.fetch(first_row, .{}).addi(op.fetch(local_rows, .{})), .{}); op = b.tensorOperation(.{}); @@ -500,14 +926,15 @@ fn prepareQueries(b: *tcl.Builder, cfg: Cfg, a: Axes, t: NativeInputs) tcl.Finis const qpos = try op.commit(op.where(live, .ne, 0, op.fetch(pos, .{}), -1), .{}); op = b.tensorOperation(.{}); const kvlen = try op.commit(op.fetch(key_length, .{}).muli(op.fetch(live_block, .{})), .{}); - const table = try selectMetadata(b, t.pages, pick, a.S); const token_pick = try selector(b, token_seq, a.S, cfg.sequences); const token_begin = try selectMetadata(b, starts, token_pick, a.S); const token_block = try selectMetadata(b, block_start, token_pick, a.S); op = b.tensorOperation(.{}); - const row = try op.commit(op.fetch(token_block, .{}).muli(prefill_block).addi(op.fetch(tokens, .{})).subi(op.fetch(token_begin, .{})).binary(.maxi, 0).binary(.mini, cfg.b * cfg.i - 1), .{}); - return .{ .q = q, .pages = table, .qpos = qpos, .kvlen = kvlen, .row = row, .active = active }; + const row = try op.commit(op.fetch(token_block, .{}).muli(cfg.i).addi(op.fetch(tokens, .{})).subi(op.fetch(token_begin, .{})).binary(.maxi, 0).binary(.mini, cfg.b * cfg.i - 1), .{}); + op = b.tensorOperation(.{}); + const first = try op.commit(op.fetch(t.seq_lens, .{}).subi(op.fetch(qlen, .{})), .{}); + return .{ .q = q, .pages = t.pages, .pick = pick, .qpos = qpos, .kvlen = kvlen, .row = row, .active = active, .first = first, .keys = t.seq_lens }; } fn restoreQueries(b: *tcl.Builder, cfg: Cfg, a: Axes, prepared: Prepared, output: builder.Tensor) tcl.FinishError!builder.Tensor { @@ -533,8 +960,7 @@ fn validIndices(b: *tcl.Builder, indices: builder.Tensor, count: i64) tcl.Finish return .{ .safe = safe, .valid = valid }; } -/// Cache validity is per sequence, not per query: causal/window masking stays -/// on scores, without replicating K/V for every row of a prefill block. +/// Validity is per sequence, not per query, so K/V are not replicated per row of a prefill block. fn cacheMask(b: *tcl.Builder, valid_pages: builder.Tensor, kvlen: builder.Tensor, position: builder.Tensor) tcl.FinishError!builder.Tensor { var op = b.tensorOperation(.{}); const pos = try op.commit(op.fetch(position, .{ .broadcast_to = valid_pages.axes }), .{}); @@ -542,17 +968,15 @@ fn cacheMask(b: *tcl.Builder, valid_pages: builder.Tensor, kvlen: builder.Tensor return op.commit(op.fetch(pos, .{}).muli(-1).addi(op.fetch(kvlen, .{})).binary(.maxi, 0).binary(.mini, 1).muli(op.fetch(valid_pages, .{})), .{}); } -/// A true select excludes NaN/Inf in unwritten slots; multiplying by zero -/// would still propagate them into the contraction. Materialize only BF16. +/// A select, not a multiply by zero: NaN/Inf in unwritten slots must not reach the contraction. fn maskCacheValues(b: *tcl.Builder, values: builder.Tensor, valid: builder.Tensor) tcl.FinishError!builder.Tensor { const op = b.tensorOperation(.{ .tactic = .elementwise }); const data = op.fetch(values, .{}); return op.commit(op.where(op.fetch(valid, .{ .broadcast_to = values.axes }), .ne, 0, data, 0.0), .{ .dtype = .bf16 }); } -/// Build the sequence, query, page and window mask for the current keys. fn attentionMask(b: *tcl.Builder, cfg: Cfg, qpos: builder.Tensor, kvlen: builder.Tensor, position: builder.Tensor, valid_keys: builder.Tensor) tcl.FinishError!builder.Tensor { - const axes = &.{ qpos.axes[0], qpos.axes[1], position.axes[0] }; + const axes = &.{ qpos.axes[0], qpos.axes[1], position.axes[position.axes.len - 1] }; const read: builder.TensorOperation.FetchOptions = .{ .broadcast_to = axes }; // Elementwise operations require their first read to cover the output // axes, and each VE instruction must consume the current accumulator. @@ -577,8 +1001,6 @@ fn attentionMask(b: *tcl.Builder, cfg: Cfg, qpos: builder.Tensor, kvlen: builder return mask; } -/// Keep the contraction accumulator in f32 through masking and softmax. -/// A finite sentinel also makes entirely masked chunks safe to rescale. fn maskedScores(b: *tcl.Builder, a: Axes, cfg: Cfg, q: builder.Tensor, k: builder.Tensor, mask: builder.Tensor, keys: tcl.Axis) tcl.FinishError!builder.Tensor { const scores_axes = &.{ a.Hkv, a.G, a.B, a.I, keys }; var op = b.tensorOperation(.{}); @@ -591,8 +1013,7 @@ fn maskedScores(b: *tcl.Builder, a: Axes, cfg: Cfg, q: builder.Tensor, k: builde fn maskedExp(b: *tcl.Builder, scores: builder.Tensor, maximum: builder.Tensor, mask: builder.Tensor) tcl.FinishError!builder.Tensor { const op = b.tensorOperation(.{}); const e = op.fetch(scores, .{}).subf(op.fetch(maximum, .{})).exp(); - // Subtracting the sentinel from itself gives exp(0), so masking only - // the logits is insufficient for empty rows and fully masked chunks. + // A sentinel minus itself gives exp(0), so masking only the logits fails for empty rows and fully masked chunks. return op.commit(op.where(op.fetch(mask, .{}), .ne, 0, e, 0.0), .{}); } @@ -622,7 +1043,7 @@ const OneShot = tcl.Kernel(Cfg, .{ const prepared = try prepareQueries(b, cfg, a, t); const kc = try b.reshape(t.k_cache, &.{ Nb, a.Hkv, a.D }, .{}); const vc = try b.reshape(t.v_cache, &.{ Nb, a.Hkv, a.D }, .{}); - const page_indices = try validIndices(b, prepared.pages, cfg.pages); + const page_indices = try validIndices(b, try prepared.blockPages(b, a, prepared.pages), cfg.pages); const within_page = try b.arange(cfg.pg, .i32, &.{a.PG}, .{}); var slot_op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); const flat_slots = try slot_op.commit(slot_op.fetch(page_indices.safe, .{}).muli(cfg.pg).addi(slot_op.fetch(within_page, .{})), .{}); @@ -630,7 +1051,6 @@ const OneShot = tcl.Kernel(Cfg, .{ slot_op = b.tensorOperation(.{}); const valid = try slot_op.commit(slot_op.fetch(page_indices.valid, .{ .broadcast_to = &.{ a.B, a.AP, a.PG } }), .{}); const valid_keys = try b.reshape(valid, &.{ a.B, A }, .{}); - // Match the original slot sanitization before reading invalid pages. slot_op = b.tensorOperation(.{}); const slots = try slot_op.commit(slot_op.where(valid_keys, .ne, 0, slot_op.fetch(expanded_slots, .{}), 0), .{}); const valid_length = if (cfg.decode) prepared.kvlen else null; @@ -642,8 +1062,6 @@ const OneShot = tcl.Kernel(Cfg, .{ const masked = try maskedScores(b, a, cfg, prepared.q, k, mask, A); var op = b.tensorOperation(.{}); const p = if (cfg.decode) compact: { - // Furiosa's common.softmax keeps normalization inside one VE - // operation. FP32 intermediates only narrow at the DPE input. const exp_sink = if (cfg.use_sink) sink: { const maximum = try op.commit(op.fetch(masked, .{}).reduce(&.{A}, .maxf).maxf(op.fetch(t.sink, .{})), .{}); break :sink try expDifference(b, t.sink, maximum); @@ -656,13 +1074,11 @@ const OneShot = tcl.Kernel(Cfg, .{ var sum = e.reduce(&.{A}, .addf); if (exp_sink) |s| sum = sum.addf(op.fetch(s, .{})); const probabilities = try op.commit(e.divf(sum), .{ .dtype = .bf16 }); - // Empty rows have uniform probabilities from finite sentinels. - // Clear masked probabilities before contracting with values. + // Empty rows have uniform probabilities from finite sentinels: clear masked ones before the value contraction. op = b.tensorOperation(.{}); break :compact try op.commit(op.where(op.fetch(mask, .{}), .ne, 0, op.fetch(probabilities, .{}), 0.0), .{ .dtype = .bf16 }); } else tiled: { - // Separate reductions schedule better for multi-query blocks - // with the current compiler; keep their intermediate state FP32. + // Separate reductions schedule better for multi-query blocks with the current compiler. var m = op.fetch(masked, .{}).reduce(&.{A}, .maxf); if (cfg.use_sink) m = m.maxf(op.fetch(t.sink, .{})); const maximum = try op.commit(m, .{}); @@ -695,28 +1111,66 @@ const Flash = tcl.Kernel(Cfg, .{ const Cp = b.axis("Cp", @divExact(cfg.chunk, cfg.pg)); const t = try declareInputs(b, a); const prepared = try prepareQueries(b, cfg, a, t); - const by_batch = try b.reshape(prepared.pages, &.{ a.B, C, Cp }, .{}); + const rows = prepared.pages.axes[0]; + const by_row = try b.reshape(prepared.pages, &.{ rows, C, Cp }, .{}); const page_op = b.tensorOperation(.{}); - const pages = try page_op.commit(page_op.fetch(by_batch, .{}), .{ .axes = &.{ C, a.B, Cp } }); + const pages = try page_op.commit(page_op.fetch(by_row, .{}), .{ .axes = &.{ C, rows, Cp } }); const stats = &.{ a.Hkv, a.G, a.B, a.I }; const out_axes = &.{ a.B, a.I, a.Hkv, a.G, a.D }; var op = b.tensorOperation(.{}); - const n = try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(prepared.kvlen), cfg.chunk - 1), cfg.chunk); + const shift = std.math.log2_int(u64, @intCast(cfg.chunk)); + // With a window, each page row's loop starts at the chunk holding its first visible key. TCL divides + // by shifts, so chunks that are not a power of two run every chunk and the mask applies the window. + const window = if (cfg.window > 0 and std.math.isPowerOfTwo(cfg.chunk)) blk: { + op = b.tensorOperation(.{}); + const start = try op.commit(op.fetch(prepared.first, .{}).subi(cfg.window - 1).binary(.maxi, 0).binary(.shr_arith, shift), .{}); + op = b.tensorOperation(.{}); + const count = try op.commit(op.fetch(prepared.keys, .{}).addi(cfg.chunk - 1).binary(.shr_arith, shift).subi(op.fetch(start, .{})), .{}); + // WORKAROUND (TCC 2026.5.0b0): TCC reads the loop index only as an index, so each trip's chunk per + // row is built before the loop. Repro: repro_loop_index_as_data.py. Remove once the loop index + // compiles as data. + const trip = try b.arange(C.size, .i32, &.{C}, .{}); + op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); + const chunks = try op.commit(op.fetch(trip, .{}).addi(op.fetch(start, .{})), .{ .axes = &.{ C, rows } }); + break :blk .{ .chunks = chunks, .trips = try b.reduceMaxI32(count) }; + } else null; + const n = if (window) |w| w.trips else try b.symExpr(.div, try b.symExpr(.add, try b.reduceMaxI32(prepared.kvlen), cfg.chunk - 1), cfg.chunk); var loop = b.openFor(n, .{ try b.full(0, .f32, stats), try b.full(0, .f32, out_axes), try b.full(score_min, .f32, stats) }); { const sum_acc, const qkv_acc, const max_acc = loop.carried; - const indices = try validIndices(b, try b.indexRead(pages, loop.iv), cfg.pages); + const local_pos = try b.arange(cfg.chunk, .i32, &.{Ck}, .{}); + const local_pages = try b.arange(@divExact(cfg.chunk, cfg.pg), .i32, &.{Cp}, .{}); + var row_pages: builder.Tensor = undefined; + var position: builder.Tensor = undefined; + var first_keys: builder.Tensor = undefined; + if (window) |w| { + const chunk_index = try b.indexRead(w.chunks, loop.iv); + op = b.tensorOperation(.{}); + const read_index = try op.commit(op.fetch(chunk_index, .{}).binary(.mini, @divExact(cfg.a, cfg.chunk) - 1), .{}); + // WORKAROUND (TCC 2026.5.0b0): a batched gather of the chunk does not lower + // for some shapes. Repro: repro_batched_gather_all_tactics_failed.py. Remove + // once batched gathers of several rows compile. + row_pages = try selectMetadata(b, by_row, try selector(b, read_index, C, C.size), C); + op = b.tensorOperation(.{}); + const row_offset = try op.commit(op.fetch(chunk_index, .{}).muli(cfg.chunk), .{}); + const offset = if (prepared.pick) |pick| try selectMetadata(b, row_offset, pick, a.S) else row_offset; + op = b.tensorOperation(.{}); + position = try op.commit(op.fetch(local_pos, .{ .broadcast_to = &.{ a.B, Ck } }).addi(op.fetch(offset, .{})), .{}); + op = b.tensorOperation(.{}); + first_keys = try op.commit(op.fetch(local_pages, .{ .broadcast_to = &.{ a.B, Cp } }).muli(cfg.pg).addi(op.fetch(offset, .{})), .{}); + } else { + row_pages = try b.indexRead(pages, loop.iv); + const offset = try b.symExpr(.mul, loop.iv, cfg.chunk); + op = b.tensorOperation(.{}); + position = try op.commit(op.fetch(local_pos, .{}).addi(op.fetch(offset, .{})), .{}); + op = b.tensorOperation(.{}); + first_keys = try op.commit(op.fetch(local_pages, .{}).muli(cfg.pg).addi(op.fetch(offset, .{})), .{}); + } + const indices = try validIndices(b, try prepared.blockPages(b, a, row_pages), cfg.pages); const idx = indices.safe; op = b.tensorOperation(.{}); const valid_pages = try op.commit(op.fetch(indices.valid, .{ .broadcast_to = &.{ a.B, Cp, a.PG } }), .{}); const valid_keys = try b.reshape(valid_pages, &.{ a.B, Ck }, .{}); - const offset = try b.symExpr(.mul, loop.iv, cfg.chunk); - const local_pos = try b.arange(cfg.chunk, .i32, &.{Ck}, .{}); - op = b.tensorOperation(.{}); - const position = try op.commit(op.fetch(local_pos, .{}).addi(op.fetch(offset, .{})), .{}); - const local_pages = try b.arange(@divExact(cfg.chunk, cfg.pg), .i32, &.{Cp}, .{}); - op = b.tensorOperation(.{}); - const first_keys = try op.commit(op.fetch(local_pages, .{}).muli(cfg.pg).addi(op.fetch(offset, .{})), .{}); op = b.tensorOperation(.{ .tactic = .einsum_by_ve }); const remaining = try op.commit(op.fetch(prepared.kvlen, .{}).subi(op.fetch(first_keys, .{})), .{}); op = b.tensorOperation(.{}); @@ -724,9 +1178,8 @@ const Flash = tcl.Kernel(Cfg, .{ op = b.tensorOperation(.{}); const cvl = try op.commit(op.fetch(live_pages, .{}).reduce(&.{Cp}, .addi), .{}); const mask = try attentionMask(b, cfg, prepared.qpos, prepared.kvlen, position, valid_keys); - const valid_cache = try cacheMask(b, valid_keys, prepared.kvlen, position); - const k = try maskCacheValues(b, try b.reshape(try b.gather(t.k_cache, idx, a.P, &.{ a.B, Cp, a.PG, a.Hkv, a.D }, .{ .valid_length = cvl }), &.{ a.B, Ck, a.Hkv, a.D }, .{}), valid_cache); - const v = try maskCacheValues(b, try b.reshape(try b.gather(t.v_cache, idx, a.P, &.{ a.B, Cp, a.PG, a.Hkv, a.D }, .{ .valid_length = cvl }), &.{ a.B, Ck, a.Hkv, a.D }, .{}), valid_cache); + const k = try b.reshape(try b.gather(t.k_cache, idx, a.P, &.{ a.B, Cp, a.PG, a.Hkv, a.D }, .{ .valid_length = cvl }), &.{ a.B, Ck, a.Hkv, a.D }, .{}); + const v = try b.reshape(try b.gather(t.v_cache, idx, a.P, &.{ a.B, Cp, a.PG, a.Hkv, a.D }, .{ .valid_length = cvl }), &.{ a.B, Ck, a.Hkv, a.D }, .{}); const masked = try maskedScores(b, a, cfg, prepared.q, k, mask, Ck); op = b.tensorOperation(.{}); const max_cur = try op.commit(op.fetch(masked, .{}).reduce(&.{Ck}, .maxf), .{}); @@ -751,7 +1204,6 @@ const Flash = tcl.Kernel(Cfg, .{ var sum = loop.results[0]; var numerator = loop.results[1]; if (cfg.use_sink) { - // Merge the zero-valued sink once, after all key chunks. op = b.tensorOperation(.{}); const maximum = try op.commit(op.fetch(loop.results[2], .{}).maxf(op.fetch(t.sink, .{})), .{}); const alpha = try expDifference(b, loop.results[2], maximum); @@ -922,6 +1374,7 @@ const TestDims = struct { page: usize = 16, uniform: bool = false, uniform_scores: bool = false, + spikes: []const usize = &.{}, dtype: zml.DataType = .bf16, }; @@ -1030,7 +1483,7 @@ test "tcl paged attention cache masking emits a BF16 select without query replic try std.testing.expect(std.mem.indexOf(u8, ir, "#tcl.ve_opcode<\"mulf\">") == null); try std.testing.expect(std.mem.indexOf(u8, ir, "#tcl.symbol<\"I\">") == null); - const decode = try Decode.emit(std.testing.allocator, .{ .t = 2, .a = 64, .hkv = 2, .g = 4, .d = 32, .pages = 8, .pg = 16, .scale = 0.125, .is_causal = true }); + const decode = try Decode.emit(std.testing.allocator, .{ .t = 2, .a = 64, .group = 2, .hkv = 2, .g = 4, .d = 32, .pages = 8, .pg = 16, .scale = 0.125, .is_causal = true, .window = 0, .use_sink = false }); defer std.testing.allocator.free(decode); try std.testing.expectEqual(@as(usize, 2), std.mem.count(u8, decode, "\"tcl.dpe\"")); try std.testing.expect(std.mem.indexOf(u8, decode, "tcl.fusible") != null); @@ -1143,6 +1596,32 @@ test "tcl paged attention chunked decode on furiosa" { try runCases(257, 514, &cases, .{ .hkv = 1, .g = 1, .d = 32 }); } +test "tcl paged attention hybrid decode on furiosa" { + if (zml.testing.env().target != .furiosa) return error.SkipZigTest; + try runCases(2049, 2101, &.{ + .{ .prefill = false, .seqs = &.{.{ 1, 300 }}, .poison = .both }, + .{ .prefill = false, .seqs = &.{.{ 1, 4097 }}, .poison = .both }, + .{ .prefill = false, .seqs = &.{.{ 1, 32784 }}, .poison = .both, .poison_value = std.math.inf(f32) }, + .{ .prefill = false, .seqs = &.{.{ 0, 0 }}, .padding = 1, .poison = .both }, + .{ .prefill = false, .seqs = &.{.{ 1, 17 }}, .mask_all = true, .poison = .both }, + }, .{ .batch = 1, .hkv = 1, .g = 2, .d = 32 }); + try runCases(2049, 2101, &.{.{ .prefill = false, .seqs = &.{.{ 1, 32784 }}, .poison = .both }}, .{ .batch = 1, .hkv = 1, .g = 2, .d = 32, .uniform_scores = true, .spikes = &.{ 4095, 4096, 12287, 12288, 32783 } }); +} + +test "tcl paged attention flash decode of many tokens on furiosa" { + if (zml.testing.env().target != .furiosa) return error.SkipZigTest; + const seqs = [_][2]i32{ .{ 1, 32784 }, .{ 1, 4097 }, .{ 1, 1 }, .{ 0, 0 } } ++ @as([12][2]i32, @splat(.{ 1, 300 })); + try runCases(2049, 2600, &.{ + .{ .prefill = false, .seqs = &seqs, .padding = 1, .poison = .both }, + .{ .prefill = false, .seqs = &seqs, .padding = 1, .poison = .both, .opts = .{ .sliding_window = 1024 } }, + .{ .prefill = false, .seqs = &seqs, .padding = 1, .poison = .both, .opts = .{ .sliding_window = 5000, .sink = &.{ 0, -std.math.inf(f32) } } }, + }, .{ .batch = 16, .hkv = 1, .g = 2, .d = 32 }); + try runCases(2049, 2700, &.{ + .{ .prefill = true, .seqs = &.{ .{ 37, 32784 }, .{ 5, 9000 } }, .padding = 2, .poison = .both, .opts = .{ .sliding_window = 1024 } }, + .{ .prefill = true, .seqs = &.{ .{ 70, 32784 }, .{ 0, 0 } }, .padding = 1, .poison = .both, .opts = .{ .sliding_window = 3000 } }, + }, .{ .batch = 2, .hkv = 1, .g = 2, .d = 32 }); +} + test "tcl paged attention uniform probabilities on furiosa" { if (zml.testing.env().target != .furiosa) return error.SkipZigTest; try runCases(257, 514, &.{.{ .prefill = false, .seqs = &.{ .{ 1, 64 }, .{ 1, 65 }, .{ 1, 66 }, .{ 0, 0 } }, .padding = 1 }}, .{ .hkv = 1, .g = 1, .d = 32, .uniform_scores = true }); @@ -1172,11 +1651,11 @@ fn runCases(comptime max_pages: usize, comptime num_pages: usize, cases: []const const params: paged.Parameters = .init(.{ .batch_size = batch, .max_num_pages = max_pages, .max_seqlen_q = 128, .is_prefill = case.prefill }); const q_t: Tensor = .init(.{ .b = tokens, .hkv = hkv, .hg = g, .hd = d }, dims.dtype); const Mod = struct { - pub fn forward(p: paged.Parameters, q: Tensor, k: Tensor, v: Tensor, opts: paged_attention.AttentionOptions) Tensor { - return paged_attention.pagedAttention(.{ .tcl = p }, q, k, v, .{ .split = .{ .k = k, .v = v } }, opts); + pub fn forward(p: paged.Parameters, q: Tensor, k: Tensor, v: Tensor, opts: AttentionOptions) Tensor { + return paged.pagedAttention(p, q, k, v, opts); } }; - const opts: paged_attention.AttentionOptions = .{ + const opts: AttentionOptions = .{ .mask = .{ .standard = .{ .causal = case.opts.is_causal, .sliding_window = if (case.opts.sliding_window > 0) @intCast(case.opts.sliding_window) else null, @@ -1209,9 +1688,16 @@ fn runCases(comptime max_pages: usize, comptime num_pages: usize, cases: []const for (kc) |*v| v.* = .fromF32(if (dims.uniform or dims.uniform_scores) 0 else r.floatNorm(f32)); for (vc, 0..) |*v, i| v.* = .fromF32(if (dims.uniform) @as(f32, @floatFromInt(i % d)) / d else r.floatNorm(f32)); + if (dims.spikes.len > 0) { + for (case.seqs[0..batch], 0..) |seq, row| { + for (0..@intCast(seq[1])) |key| { + const slot: usize = @intCast(table[row][key / page] * @as(i32, page) + @as(i32, @intCast(key % page))); + const spike = std.mem.indexOfScalar(usize, dims.spikes, key) != null; + @memset(vc[slot * hkv * d ..][0 .. hkv * d], bf16.fromF32(if (spike) 32768 else 0)); + } + } + } if (case.poison != .none) { - // Mark the union of live physical slots before poisoning: reordered - // or shared pages must retain every token a live sequence can read. const live = try allocator.alloc(bool, num_pages * page); defer allocator.free(live); @memset(live, false); @@ -1243,11 +1729,11 @@ fn runCases(comptime max_pages: usize, comptime num_pages: usize, cases: []const .query_start_len = try .fromBytes(io, platform, params.query_start_len.shape(), std.mem.sliceAsBytes(&starts)), }; defer zml.Buffer.deinitAll(paged.Parameters, &bp); - var opts_d: zml.Bufferized(paged_attention.AttentionOptions) = .{ + var opts_d: zml.Bufferized(AttentionOptions) = .{ .mask = .{ .standard = {} }, .sink = if (case.opts.sink) |s| try .fromBytes(io, platform, opts.sink.?.shape(), std.mem.sliceAsBytes(s)) else null, }; - defer zml.Buffer.deinitAll(paged_attention.AttentionOptions, &opts_d); + defer zml.Buffer.deinitAll(AttentionOptions, &opts_d); var result = try exe.eval(allocator, io, .{ bp, bq, bk, bv, opts_d }); defer result.deinit(); const host = try result.toSliceAlloc(allocator, io); diff --git a/zml/kernels/paged_attention/triton/triton.zig b/zml/kernels/paged_attention/triton/triton.zig index 1ccbeba8..2a12e012 100644 --- a/zml/kernels/paged_attention/triton/triton.zig +++ b/zml/kernels/paged_attention/triton/triton.zig @@ -4,10 +4,10 @@ const stdx = @import("stdx"); const zml = @import("../../../zml.zig"); const triton = zml.kernel.triton; -const AttentionOptions = @import("../paged_attention.zig").AttentionOptions; +const AttentionOptions = @import("../options.zig").AttentionOptions; const kernels = @import("unified_attention.zig"); -const kernels_sparse = @import("sparse_unified_attention.zig"); const kernels_oneapi = @import("unified_attention_oneapi.zig"); +const kernels_sparse = @import("sparse_unified_attention.zig"); const log = std.log.scoped(.@"zml/attention/triton"); @@ -274,92 +274,57 @@ pub const paged = struct { return pagedSparseAttention(parameters, q, k_cache, v_cache, opts); } - const output = zml.ops.manualComputation( - (struct { - q: zml.Tensor, - k_cache: zml.Tensor, - v_cache: zml.Tensor, - block_table: zml.Tensor, - seq_lens: zml.Tensor, - query_start_len: zml.Tensor, - opts: AttentionOptions, - options: Options, + const cu_count = getCuCount(); + const num_heads: usize = @intCast(q.dim(.hkv) * q.dim(.hg)); + const num_kv_heads: usize = @intCast(k_cache.dim(.hkv)); + const num_queries_per_kv: usize = num_heads / num_kv_heads; + // Intel decode: pack exactly one GQA group per tile (block_q == 1) so the + // single decode query token doesn't carry masked-out fp32 acc lanes. + // oneAPI decode keeps one GQA group per tile, padded to a power of two so tt.make_range emits legal Triton IR. + const block_m: usize = if (!parameters.options_.is_prefill and isOneapiTarget()) + std.math.ceilPowerOfTwoAssert(usize, num_queries_per_kv) + else if (num_queries_per_kv <= 16) 16 else std.math.ceilPowerOfTwoAssert(usize, num_queries_per_kv); + const block_q: usize = block_m / num_queries_per_kv; + const num_tokens: usize = @intCast(q.dim(.b)); + const num_seqs: usize = @intCast(parameters.block_table.dim(.b)); + const total_q_blocks: usize = num_tokens / block_q + num_seqs; + const target_num_prgms: usize = cu_count * 4; + const num_2d_prgms: usize = total_q_blocks * num_kv_heads; - fn body(self: @This(), _: zml.Shape) zml.Tensor { - const parameters_: Parameters = .{ - .block_table = self.block_table, - .seq_lens = self.seq_lens, - .query_start_len = self.query_start_len, - .options_ = self.options, - }; + const sliding_window: usize = opts.mask.slidingWindow() orelse 0; - const cu_count = getCuCount(); - const num_heads: usize = @intCast(self.q.dim(.hkv) * self.q.dim(.hg)); - const num_kv_heads: usize = @intCast(self.k_cache.dim(.hkv)); - const num_queries_per_kv: usize = num_heads / num_kv_heads; - // Intel decode: pack exactly one GQA group per tile (block_q == 1) so the - // single decode query token doesn't carry masked-out fp32 acc lanes. - // oneAPI decode keeps one GQA group per tile, padded to a power of two so tt.make_range emits legal Triton IR. - const block_m: usize = if (!self.options.is_prefill and isOneapiTarget()) - std.math.ceilPowerOfTwoAssert(usize, num_queries_per_kv) - else if (num_queries_per_kv <= 16) 16 else std.math.ceilPowerOfTwoAssert(usize, num_queries_per_kv); - const block_q: usize = block_m / num_queries_per_kv; - const num_tokens: usize = @intCast(self.q.dim(.b)); - const num_seqs: usize = @intCast(parameters_.block_table.dim(.b)); - const total_q_blocks: usize = num_tokens / block_q + num_seqs; - const target_num_prgms: usize = cu_count * 4; - const num_2d_prgms: usize = total_q_blocks * num_kv_heads; + const paged_attention_opts: PagedAttentionOptions = .{ + .cu_count = getCuCount(), + .all_decode = !parameters.options_.is_prefill, + .num_tokens = num_tokens, + .num_heads = num_heads, + .num_kv_heads = num_kv_heads, + .head_dim = @intCast(q.dim(.hd)), + .batch_size = @intCast(parameters.block_table.dim(.b)), + .block_size = @intCast(k_cache.dim(.k_chunk)), + .num_blocks = @intCast(k_cache.dim(.page)), + .max_num_block_per_seq = @intCast(parameters.block_table.dim(.p)), + .sliding_window = sliding_window, + .block_m = block_m, + .block_q = block_q, + .total_q_blocks = total_q_blocks, + .target_num_prgms = target_num_prgms, + .num_2d_prgms = num_2d_prgms, + .max_seqlen_q = parameters.options_.max_seqlen_q, + .scale = opts.scale, + }; - const sliding_window: usize = self.opts.mask.slidingWindow() orelse 0; - - const paged_attention_opts: PagedAttentionOptions = .{ - .cu_count = getCuCount(), - .all_decode = !self.options.is_prefill, - .num_tokens = num_tokens, - .num_heads = num_heads, - .num_kv_heads = num_kv_heads, - .head_dim = @intCast(self.q.dim(.hd)), - .batch_size = @intCast(parameters_.block_table.dim(.b)), - .block_size = @intCast(self.k_cache.dim(.k_chunk)), - .num_blocks = @intCast(self.k_cache.dim(.page)), - .max_num_block_per_seq = @intCast(parameters_.block_table.dim(.p)), - .sliding_window = sliding_window, - .block_m = block_m, - .block_q = block_q, - .total_q_blocks = total_q_blocks, - .target_num_prgms = target_num_prgms, - .num_2d_prgms = num_2d_prgms, - .max_seqlen_q = self.options.max_seqlen_q, - .scale = self.opts.scale, - }; - - const use_2d_kernel = use2dKernel( - paged_attention_opts.all_decode, - paged_attention_opts.batch_size, - paged_attention_opts.num_kv_heads, - ); - const output = if (use_2d_kernel) - pagedAttention2d(parameters_, self.q, self.k_cache, self.v_cache, self.opts, paged_attention_opts) - else if (isOneapiTarget()) - pagedAttention3dOneapi(parameters_, self.q, self.k_cache, self.v_cache, self.opts, paged_attention_opts) - else - pagedAttention3d(parameters_, self.q, self.k_cache, self.v_cache, self.opts, paged_attention_opts); - - return output; - } - }).body, - .{ - .q = q, - .k_cache = k_cache, - .v_cache = v_cache, - .block_table = parameters.block_table, - .seq_lens = parameters.seq_lens, - .query_start_len = parameters.query_start_len, - .opts = opts, - .options = parameters.options_, - }, - q.shape(), + const use_2d_kernel = use2dKernel( + paged_attention_opts.all_decode, + paged_attention_opts.batch_size, + paged_attention_opts.num_kv_heads, ); + const output = if (use_2d_kernel) + pagedAttention2d(parameters, q, k_cache, v_cache, opts, paged_attention_opts) + else if (isOneapiTarget()) + pagedAttention3dOneapi(parameters, q, k_cache, v_cache, opts, paged_attention_opts) + else + pagedAttention3d(parameters, q, k_cache, v_cache, opts, paged_attention_opts); return output; } @@ -372,51 +337,18 @@ pub const paged = struct { stdx.debug.assert(indices.shape().hasTags(.{ .b, .topk }), "Sparse indices expect to have .b, .topk tags", .{}); stdx.debug.assert(q.dtype() == k_cache.dtype() and q.dtype() == v_cache.dtype(), "Sparse attention requires matching query and KV dtypes", .{}); - return zml.ops.manualComputation( - (struct { - q: zml.Tensor, - k_cache: zml.Tensor, - v_cache: zml.Tensor, - block_table: zml.Tensor, - seq_lens: zml.Tensor, - query_start_len: zml.Tensor, - opts: AttentionOptions, - options: Options, - - fn body(self: @This(), _: zml.Shape) zml.Tensor { - const parameters_: Parameters = .{ - .block_table = self.block_table, - .seq_lens = self.seq_lens, - .query_start_len = self.query_start_len, - .options_ = self.options, - }; - - const config = selectSparseConfig( - @intCast(self.opts.mask.indices.dim(.topk)), - @intCast(self.block_table.dim(.b)), - @intCast(self.q.dim(.hkv)), - getCuCount(), - !self.options.is_prefill, - ); - - return if (config.num_segments == 1) - pagedSparseAttention2d(parameters_, self.q, self.k_cache, self.v_cache, self.opts, config) - else - pagedSparseAttention3d(parameters_, self.q, self.k_cache, self.v_cache, self.opts, config); - } - }).body, - .{ - .q = q, - .k_cache = k_cache, - .v_cache = v_cache, - .block_table = parameters.block_table, - .seq_lens = parameters.seq_lens, - .query_start_len = parameters.query_start_len, - .opts = opts, - .options = parameters.options_, - }, - q.shape(), + const config = selectSparseConfig( + @intCast(opts.mask.indices.dim(.topk)), + @intCast(parameters.block_table.dim(.b)), + @intCast(q.dim(.hkv)), + getCuCount(), + !parameters.options_.is_prefill, ); + + return if (config.num_segments == 1) + pagedSparseAttention2d(parameters, q, k_cache, v_cache, opts, config) + else + pagedSparseAttention3d(parameters, q, k_cache, v_cache, opts, config); } pub fn pagedSparseAttention2d(parameters: Parameters, q: zml.Tensor, k_cache: zml.Tensor, v_cache: zml.Tensor, opts: AttentionOptions, tuning: SparseConfig) zml.Tensor { diff --git a/zml/kernels/paged_attention/triton/unified_attention.zig b/zml/kernels/paged_attention/triton/unified_attention.zig index 39018f1b..7b2b9428 100644 --- a/zml/kernels/paged_attention/triton/unified_attention.zig +++ b/zml/kernels/paged_attention/triton/unified_attention.zig @@ -1020,6 +1020,25 @@ fn reduceSegments( const query_token_idx = k.programId(.x); const query_head_idx = k.programId(.y); + const offs_d = k.arange(0, HEAD_SIZE_PADDED, .i32); + const dim_mask: Value = if (HEAD_SIZE_PADDED != HEAD_SIZE) + offs_d.lt(@as(i32, @intCast(HEAD_SIZE))) + else + k.full(&.{HEAD_SIZE_PADDED}, 1, .i1); + const output_offset = query_token_idx.to(.i64).mul(output_stride_0) + .add(query_head_idx.to(.i64).mul(output_stride_1)) + .add(offs_d.to(.i64)); + + // The launch covers the padded query capacity, but the attention pass only + // writes scratch for active tokens. Zero the tail before looking up a + // sequence or dividing by its segment size (zero for inactive sequences). + const active_tokens = k.load(query_start_len_ptr.addPtr(num_seqs)); + const inactive = query_token_idx.ge(active_tokens); + k.storeOpts(output_ptr.addPtr(output_offset), k.zeros(&.{HEAD_SIZE_PADDED}, config.o_dtype), .{ + .mask = dim_mask.bitAnd(inactive), + }); + k.returnIf(inactive); + const seq_idx = findSeqIdx(k, query_start_len_ptr, query_token_idx, num_seqs, BLOCK_Q, false); const seq_len = k.load(seq_lens_ptr.addPtr(seq_idx)); @@ -1029,12 +1048,6 @@ fn reduceSegments( const seg_range = k.arange(0, NUM_SEGMENTS_PER_SEQ, .i32); const segm_mask = seg_range.lt(act_num_segments); - const offs_d = k.arange(0, HEAD_SIZE_PADDED, .i32); - const dim_mask: Value = if (HEAD_SIZE_PADDED != HEAD_SIZE) - offs_d.lt(@as(i32, @intCast(HEAD_SIZE))) - else - k.full(&.{HEAD_SIZE_PADDED}, 1, .i1); - // segm_offset = query_token_idx*NQH*NSPS + query_head_idx*NSPS + arange(NSPS) const tok_stride: i64 = NUM_QUERY_HEADS * NUM_SEGMENTS_PER_SEQ; const head_stride: i32 = @intCast(NUM_SEGMENTS_PER_SEQ); @@ -1100,9 +1113,6 @@ fn reduceSegments( ); } - const output_offset = query_token_idx.to(.i64).mul(output_stride_0) - .add(query_head_idx.to(.i64).mul(output_stride_1)) - .add(offs_d.to(.i64)); k.storeOpts( output_ptr.addPtr(output_offset), acc.to(config.o_dtype), diff --git a/zml/meta.zig b/zml/meta.zig index 1eefea11..84fb9d36 100644 --- a/zml/meta.zig +++ b/zml/meta.zig @@ -143,7 +143,7 @@ test MapType { /// The `to` parameter must be passed with mutable pointer, and tensor data need to be mutable if callback needs it. /// `mapAlloc` tries as much as possible to respect the conversions made by Zig itself. /// For example it can convert from a comptime array to a runtime slice. -/// `mapAlloc` can allocate new slices to write the result if the result struct requires it. +/// `mapAlloc` allocates slices and single-item pointers needed to hold mapped values. /// The caller is owning said allocations, using an `ArenaAllocator` might help tracking them. /// /// Note: to avoid infinite loop, mapAlloc doesn't look for `From` fields inside `To` struct. @@ -180,6 +180,11 @@ pub fn mapAlloc(comptime cb: anytype, allocator: std.mem.Allocator, ctx: FnParam return; } + if (FromStruct == ToStruct and !comptime Contains(FromStruct, From)) { + to.* = from; + return; + } + if (@sizeOf(ToStruct) == 0) return; switch (type_info_to) { @@ -226,7 +231,11 @@ pub fn mapAlloc(comptime cb: anytype, allocator: std.mem.Allocator, ctx: FnParam } to.* = items; }, - else => try mapAlloc(cb, allocator, ctx, from.*, to.*), + else => { + const item = try allocator.create(ptr_info.child); + try mapAlloc(cb, allocator, ctx, from.*, item); + to.* = item; + }, }, .slice => { const items = try allocator.alloc(@typeInfo(ToStruct).pointer.child, from.len); @@ -267,6 +276,7 @@ test mapAlloc { field: A, array: [2]A, slice: []const A, + pointer: *const A, other: u8, tag: [*:0]const u8, // We want to allow conversion from comptime to runtime, because Zig type inference works like this. @@ -278,6 +288,7 @@ test mapAlloc { field: B, array: [2]B, slice: []const B, + pointer: *const B, other: u8, tag: [*:0]const u8, static_val: u8, @@ -291,6 +302,7 @@ test mapAlloc { .other = 7, .tag = "feature", .slice = &.{ .{ .a = 9 }, .{ .a = 10 } }, + .pointer = &.{ .a = 13 }, .field_with_empty = .{ .{ .a = 9 }, .{} }, }; var bb: BB = undefined; @@ -298,6 +310,7 @@ test mapAlloc { try mapAlloc(A.convert, testing.allocator, {}, aa, &bb); defer testing.allocator.free(bb.slice); defer testing.allocator.free(bb.static_slice); + defer testing.allocator.destroy(bb.pointer); try testing.expectEqual(4, bb.field.b); try testing.expectEqual(5, bb.array[0].b); @@ -309,6 +322,7 @@ test mapAlloc { try testing.expectEqual(10, bb.slice[1].b); try testing.expectEqual(11, bb.static_slice[0].b); try testing.expectEqual(12, bb.static_slice[1].b); + try testing.expectEqual(13, bb.pointer.b); } /// Visit a given type `T` and: diff --git a/zml/nn.zig b/zml/nn.zig index 56fd8b51..eceb316b 100644 --- a/zml/nn.zig +++ b/zml/nn.zig @@ -2,11 +2,13 @@ const std = @import("std"); const mlir = @import("mlir"); +const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); const constants = @import("constants.zig"); const DataType = @import("dtype.zig").DataType; const meta = @import("meta.zig"); +const mlirx = @import("mlirx.zig"); const ops = @import("ops.zig"); const quantization = @import("quantization.zig"); pub const Quantization = quantization.Quantization; @@ -226,7 +228,7 @@ test "block128 scaled dot layouts" { const io = std.testing.io; const platform = zml.testing.env(); // FIXME: Add rocm when we have a PJRT plugin with native fp8 dot support (by the end of Sept 2026) - if (platform.target != .cuda) return error.SkipZigTest; + if (platform.target != .cuda and platform.target != .furiosa) return error.SkipZigTest; const dtype: DataType = if (platform.target == .rocm and @import("platform.zig").rocm.computeCapability(platform) == .gfx942) .f8e4m3fnuz else .f8e4m3fn; const Local = struct { const Outputs = struct { actual: Tensor, expected: Tensor, linear: Tensor, linear_expected: Tensor }; @@ -320,13 +322,23 @@ fn onesGrid(shape: Shape, dt: DataType) Shape { } fn scaledDotReference(in: []const Tensor, out_shape: Shape) Tensor { - return ops.customCall( - "zml$scaled_dot_unmatched", - .{ in[0], in[1], in[2], in[3] }, - out_shape, - {}, - .{ .has_side_effect = false }, - ); + // Keep the unmatched-composite marker in the graph for backend lowering. + const ctx = zml.Compiler.current(); + const op = stablehlo.custom_call( + ctx.mlir_ctx, + &.{ in[0].value(), in[1].value(), in[2].value(), in[3].value() }, + &.{mlirx.Type.rankedTensor(ctx.mlir_ctx, out_shape)}, + .{ + .call_target_name = "zml$scaled_dot_unmatched", + .has_side_effect = false, + .backend_config = .{ .typed_ffi = .dict(ctx.mlir_ctx, &.{}) }, + }, + ctx.location, + ).appendTo(ctx.currentScope().block); + if (ctx.manual_computation_depth > 0 and ctx.partitioner == .gspmd) { + op.setAttributeByName("mhlo.sharding", .string(ctx.mlir_ctx, "{manual}")); + } + return Tensor._result(out_shape, op.result(0)); } fn applyGlobalScale(acc: Tensor, igs: ?Tensor, wgs: ?Tensor) Tensor { @@ -2354,8 +2366,8 @@ pub fn softplus( // Using log(1 + exp(x)) directly loses precision for x < ~-16.6, causing strongly negative logits to produce zero scores. const z = x.exp(); const u = z.addConstant(1); - const log1p = u.cmp(.EQ, .scalar(1, .f32)).select(z, z.mul(u.log()).div(u.subConstant(1))); - return x.cmp(.GT, .scalar(opts.threshold, .f32)).select(x, log1p); + const log1p = u.cmp(.EQ, .scalar(1, x.dtype())).select(z, z.mul(u.log()).div(u.subConstant(1))); + return x.cmp(.GT, .scalar(opts.threshold, x.dtype())).select(x, log1p); } test "softplus" { @@ -2365,14 +2377,11 @@ test "softplus" { const platform = zml.testing.env(); const input: Tensor = .init(.{ .x = inputs.len }, .f32); const Local = struct { - fn call(x: Tensor) Tensor { - return softplus(x, .{ .threshold = 2.0 }); + fn call(x: Tensor, dtype: DataType) Tensor { + return softplus(x.convert(dtype), .{ .threshold = 2.0 }).convert(x.dtype()); } }; - var exe = try platform.compileFn(std.testing.allocator, std.testing.io, Local.call, .{input}, .{}); - defer exe.deinit(); - var input_buffer: zml.Buffer = try .fromBytes( std.testing.io, platform, @@ -2381,13 +2390,18 @@ test "softplus" { ); defer input_buffer.deinit(); - var result = try exe.eval(std.testing.allocator, std.testing.io, .{input_buffer}); - defer result.deinit(); + for ([_]DataType{ .f32, .bf16 }) |dtype| { + var exe = try platform.compileFn(std.testing.allocator, std.testing.io, Local.call, .{ input, dtype }, .{}); + defer exe.deinit(); - try zml.testing.expectClose( - std.testing.io, - Slice.init(input.shape(), std.mem.sliceAsBytes(&expected)), - result, - .{}, - ); + var result = try exe.eval(std.testing.allocator, std.testing.io, .{input_buffer}); + defer result.deinit(); + + try zml.testing.expectClose( + std.testing.io, + Slice.init(input.shape(), std.mem.sliceAsBytes(&expected)), + result, + .{}, + ); + } } diff --git a/zml/ops.zig b/zml/ops.zig index b1a26d23..f8fdba7e 100644 --- a/zml/ops.zig +++ b/zml/ops.zig @@ -8,13 +8,6 @@ const platforms = @import("platforms"); const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); -const manual_computation = @import("ops/manual_computation.zig"); -pub const manualComputation = manual_computation.manualComputation; - -test { - _ = manual_computation; -} - const Buffer = @import("buffer.zig").Buffer; const Bufferized = @import("mem.zig").Bufferized; const Compiler = @import("Compiler.zig"); @@ -23,15 +16,113 @@ const CustomCallBuffer = @import("pjrtx.zig").CustomCallBuffer; const DataType = @import("dtype.zig").DataType; const meta = @import("meta.zig"); const mlirx = @import("mlirx.zig"); +const manual_computation = @import("ops/manual_computation.zig"); +pub const manualComputation = manual_computation.manualComputation; +pub const ManualPartitioning = manual_computation.ManualPartitioning; +pub const PartitionSpecs = manual_computation.PartitionSpecs; +pub const partitionSpecs = manual_computation.partitionSpecs; +pub const partitionSpecsAlloc = manual_computation.partitionSpecsAlloc; const Platform = @import("platform.zig").Platform; const Shape = @import("shape.zig").Shape; const ShapeToCustomCallBuffer = @import("pjrtx.zig").ShapeToCustomCallBuffer; const Sharding = @import("Sharding.zig"); const Tensor = @import("tensor.zig").Tensor; const TensorToCustomCallBuffer = @import("pjrtx.zig").TensorToCustomCallBuffer; + +test { + _ = manual_computation; +} + const log = std.log.scoped(.@"zml/Compiler"); +/// Reduces local tensor values across all processes in the computation. +/// For each tensor index, `func_` combines the values at that index from every +/// process. The result is identical on all processes. +/// +/// Inside a manual computation, `inputs` describes the local tensors at each +/// partition. For example, summing `[1, 2]` and `[3, 4]` from two partitions gives +/// `[4, 6]` at each partition. The local tensor dimensions stay unchanged. +/// +/// * `inputs` is a `Tensor`, a nonempty tuple of tensors, or an `[N]Tensor`. +/// Each input must have matching shapes and dtypes across processes. +/// * `func_` describes how to combine two scalar contributions: +/// - For one input, it must be `fn (left: Tensor, right: Tensor) Tensor`, such +/// as `Tensor.add`. This also applies to a one-element tuple or array. +/// - For multiple inputs, it takes one `ReduceArgs` per input and returns a +/// tuple or array with one scalar tensor per input, in the same order. +/// +/// Returns the same container type as `inputs`, preserving tensor attributes. +/// With one process, returns `inputs` unchanged. +/// +/// The reduction uses one group containing all processes, including mesh replicas. +/// To reduce along one logical mesh axis, use `allReduceOnAxis`. pub fn allReduce(inputs: anytype, comptime func_: anytype) AllReduceReturnType(@TypeOf(inputs)) { + const count = Compiler.current().meshes[0].numDevices(); + var devices: [Platform.MAX_NUM_DEVICES]i64 = undefined; + for (devices[0..count], 0..) |*device, i| device.* = @intCast(i); + return allReduceGroups(inputs, func_, devices[0..count], count); +} + +/// Reduces local tensor values within process groups defined by `axis` in `mesh`. +/// A group contains processes with the same coordinates on every other logical +/// mesh axis. Separate replicas of the mesh belong to separate groups. Within +/// each group, `func_` combines corresponding elements and the result is identical +/// on all member processes. +/// +/// Use this inside a manual computation to combine partial results along `axis` +/// while preserving the independent computations along the other mesh axes. +/// +/// * `inputs` and `func_` follow `allReduce`'s tensor and reducer contracts. +/// * `mesh` is the logical mesh used by the manual computation and must be +/// included in the compilation's meshes. +/// * `axis` names an axis of that mesh, for example `.experts`. It can be an enum +/// literal or a mesh axis tag. The axis must exist in `mesh`. +/// +/// Returns the same container type with tensor attributes as +/// `inputs`. An axis of size one returns `inputs` unchanged. +/// +/// For a mesh with `.data = 2` and `.experts = 4`, the following sums four local +/// expert contributions independently for each of the two data shards: +/// +/// ```zig +/// const output = allReduceOnAxis(local_output, Tensor.add, mesh, .experts); +/// ``` +/// +/// Each expert shard now has the complete output for its data shard. To expose +/// this result through `manualComputation`, an output spec such as +/// `.parse(.{ .t = .data })` keeps tokens split across `.data` and declares the +/// result replicated along `.experts`. +pub fn allReduceOnAxis(inputs: anytype, comptime func_: anytype, mesh: *const Sharding.Mesh, axis: anytype) AllReduceReturnType(@TypeOf(inputs)) { + const logical_axis = mesh.resolveLogicalAxis(Shape.toTag(axis)) orelse @panic("Unknown logical reduction axis"); + const size = mesh.grid.axis_sizes[logical_axis]; + var groups: [Platform.MAX_NUM_DEVICES]i64 = undefined; + var replica_indices: [Platform.MAX_NUM_DEVICES]usize = @splat(0); + const replicas = mesh.numReplicas(); + for (0..mesh.numDevices()) |device| { + const coords = mesh.grid.logical_coords[device]; + var tile: usize = 0; + var group: usize = 0; + for (0..mesh.logical.axes.len) |dim| { + tile = tile * mesh.grid.axis_sizes[dim] + coords[dim]; + if (dim != logical_axis) group = group * mesh.grid.axis_sizes[dim] + coords[dim]; + } + group = group * replicas + replica_indices[tile]; + replica_indices[tile] += 1; + groups[group * size + coords[logical_axis]] = @intCast(device); + } + return allReduceGroups(inputs, func_, groups[0..mesh.numDevices()], size); +} + +/// Reduces corresponding tensor elements independently within each process group. +/// `inputs`, `func_` and the return value follow `allReduce`'s contract. +/// +/// `groups` identifies processes by their executable partition IDs, as returned +/// by `partitionId()`. Each consecutive `group_size` IDs defines one group. +/// +/// For example, `groups = &.{ 0, 2, 1, 3 }` and `group_size = 2` combine partitions +/// 0 and 2 together, and partitions 1 and 3 together. Both members of each group +/// receive that group's result. Groups of size one return `inputs` unchanged. +fn allReduceGroups(inputs: anytype, comptime func_: anytype, groups: []const i64, group_size: usize) AllReduceReturnType(@TypeOf(inputs)) { const ctx = Compiler.current(); const mlir_ctx = ctx.mlir_ctx; @@ -75,8 +166,7 @@ pub fn allReduce(inputs: anytype, comptime func_: anytype) AllReduceReturnType(@ else => @compileError("zml.ops.allReduce expects Tensor, tuple of Tensor, or [N]Tensor inputs"), }; - const num_devices = ctx.meshes[0].numDevices(); - if (num_devices <= 1) return inputs; + if (group_size <= 1) return inputs; const reducer_block = b: { const ArgsTypes: [input_tensors.len]type = @splat(ReduceArgs); @@ -125,14 +215,9 @@ pub fn allReduce(inputs: anytype, comptime func_: anytype) AllReduceReturnType(@ break :b block; }; - var replica_group_storage: [constants.MAX_RANK]i64 = undefined; - for (0..@intCast(num_devices)) |i| { - replica_group_storage[i] = @intCast(i); - } - const replica_group_slice = replica_group_storage[0..@intCast(num_devices)]; const replica_groups_attr = mlir.Attribute.denseElements( - .rankedTensor(&.{ 1, num_devices }, .int(mlir_ctx, .i64)), - replica_group_slice, + .rankedTensor(&.{ @intCast(groups.len / group_size), @intCast(group_size) }, .int(mlir_ctx, .i64)), + groups, ); const handle = ctx.nextChannelId(); @@ -202,6 +287,20 @@ fn AllReduceReturnType(comptime InputsT: type) type { }; } +/// Returns the current process's zero based coordinate on `axis` in `mesh` +/// as a scalar u32. +/// +/// Example: for a mesh {.data = 2, .experts = 2}, each data group has expert coordinates +/// 0 and 1. Use this coordinate inside a manual computation to identify a shard +/// along one axis while `partitionId()` identifies a partition across the whole mesh. +pub fn axisIndex(mesh: *const Sharding.Mesh, axis: anytype) Tensor { + const logical_axis = mesh.resolveLogicalAxis(Shape.toTag(axis)) orelse @panic("Unknown logical mesh axis"); + var coords: [Platform.MAX_NUM_DEVICES]u32 = undefined; + for (coords[0..mesh.numDevices()], 0..) |*coord, i| coord.* = mesh.grid.logical_coords[i][logical_axis]; + return Tensor.constantTensor(.init(.{ .device = mesh.numDevices() }, .u32), std.mem.sliceAsBytes(coords[0..mesh.numDevices()])) + .gather(.{ .device = partitionId() }, .{}); +} + pub fn partitionId() Tensor { const ctx = Compiler.current(); const op = stablehlo.partition_id(ctx.mlir_ctx, ctx.location).appendTo(ctx.currentScope().block); @@ -830,10 +929,9 @@ fn fromMlirOperationWithTags(op: *const mlir.Operation, base: anytype) @TypeOf(b fn cb(inner_ctx: *LocalContext, tensor: *Tensor) void { var new = Tensor.fromMlirValue(inner_ctx.op.result(inner_ctx.index)); stdx.debug.internalAssert(new.rank() == tensor.rank(), "expected operand result to have rank {} but got {f}", .{ tensor.rank(), new }); - // copy tags and mesh info over + // Copy tags; the partitioner infers result sharding. // some ops can change dims eg reduceWindow, so we trust mlir here. new._shape._tags = tensor._shape._tags; - new._shape._sharding = tensor._shape._sharding; tensor.* = new; inner_ctx.index += 1; } @@ -1434,7 +1532,15 @@ pub const CuteOps = struct { block: ?[3]i32 = null, zeroed_outputs: []const i32 = &.{}, scalars: []const i64 = &.{}, + symmetric_workspace: ?SymmetricWorkspace = null, output_operand_aliases: []const stablehlo.CustomCallOpts.OutputOperandAlias = &.{}, + + /// Named per-device workspace shared across CuTe calls and executables. + /// Every rank allocates `bytes`; XLA passes peer addresses and the rank. + pub const SymmetricWorkspace = struct { + name: []const u8, + bytes: i64, + }; }; /// A `__gpu$xla.gpu.cute` custom call: XLA compiles `ir` with the CuTe compiler @@ -1455,7 +1561,7 @@ pub fn cute(inputs: anytype, outputs: anytype, opts: CuteOps) [outputs.len]Tenso res_types[i] = mlirx.Type.rankedTensor(mlir_ctx, output); } - var attrs: stdx.BoundedArray(mlir.NamedAttribute, 7) = .empty; + var attrs: stdx.BoundedArray(mlir.NamedAttribute, 8) = .empty; attrs.appendSliceAssumeCapacity(&.{ .named(mlir_ctx, "name", .string(mlir_ctx, opts.name)), .named(mlir_ctx, "kernel_type", .string(mlir_ctx, "cute")), @@ -1480,6 +1586,13 @@ pub fn cute(inputs: anytype, outputs: anytype, opts: CuteOps) [outputs.len]Tenso for (opts.scalars, scalar_attrs) |value, *attribute| attribute.* = .int(mlir_ctx, .i64, value); attrs.appendAssumeCapacity(.named(mlir_ctx, "scalars", .array(mlir_ctx, scalar_attrs))); } + if (opts.symmetric_workspace) |workspace| { + stdx.debug.assert(workspace.name.len > 0 and workspace.bytes > 0, "zml.ops.cute({s}): symmetric_workspace needs a name and a positive size", .{opts.name}); + attrs.appendAssumeCapacity(.named(mlir_ctx, "symmetric_workspace", .dict(mlir_ctx, &.{ + .named(mlir_ctx, "name", .string(mlir_ctx, workspace.name)), + .named(mlir_ctx, "bytes", .int(mlir_ctx, .i64, workspace.bytes)), + }))); + } const backend_config: *const mlir.Attribute = .dict(mlir_ctx, attrs.constSlice()); const op = stablehlo.custom_call( @@ -2647,11 +2760,6 @@ pub fn CustomCall( comptime params: struct { /// Name of the custom call, must be unique. name: [:0]const u8, - /// Custom call will see the data of a single device. So for a sharded buffer, it will only see a single partition. This applies - /// for both the input and output buffers. By default, the partitioning algorithm will assume nothing on the custom call input/output, - /// so everything will be forced as replicated. If do you know / ensure that the custom call is consistent with the provided input/output - /// sharding, this setting will avoid the shuffling of the buffer's data. - sharding_aware: bool, /// Whether the function has any side-effect, meaning that XLA cannot re-order it/optimize it out. A typical example is print. has_side_effect: bool, /// Similar to reuseBuffer on tensors, tells XLA that the custom call re-uses an input buffer for its output. @@ -2685,23 +2793,7 @@ pub fn CustomCall( .output_operand_aliases = comptime customCallOutputOperandAliases(I, O, params.output_operand_aliases), .compute_on_host = params.compute_on_host, }; - if (params.sharding_aware) { - return shardingAwareTypedCustomCall( - params.name, - opts, - input_tensors, - output_shapes, - attributes, - ); - } else { - return typedCustomCall( - params.name, - opts, - input_tensors, - output_shapes, - attributes, - ); - } + return typedCustomCall(params.name, opts, input_tensors, output_shapes, attributes); } fn handler(call_frame: *pjrt.ffi.CallFrame) callconv(.c) ?*pjrt.ffi.Error { @@ -2729,44 +2821,6 @@ pub fn CustomCallOutputOperandAliases(I: type, O: type) type { ); } -pub fn shardingAwareTypedCustomCall( - comptime target_name: [:0]const u8, - comptime opts: CustomCallOptions, - input: anytype, - output: anytype, - attributes: anytype, -) ShapeToTensor(@TypeOf(output)) { - const Input = @TypeOf(input); - const Output = @TypeOf(output); - const Attributes = @TypeOf(attributes); - - var output_shapes: [@typeInfo(Output).@"struct".field_names.len]Shape = undefined; - inline for (@typeInfo(Output).@"struct".field_names, 0..) |field_name, i| { - output_shapes[i] = @field(output, field_name); - } - - const Handler = struct { - input: Input, - attributes: Attributes, - - fn body(self: @This(), sharded_output_shapes: []const Shape) []const Tensor { - return typedCustomCall(target_name, opts, self.input, sharded_output_shapes, self.attributes); - } - }; - const output_tensors = manualComputation( - Handler.body, - .{ .input = input, .attributes = attributes }, - @as([]const Shape, &output_shapes), - ); - - // Convert the slice back to a struct - var out: ShapeToTensor(Output) = undefined; - inline for (@typeInfo(Output).@"struct".field_names, 0..) |field_name, i| { - @field(out, field_name) = output_tensors[i]; - } - return out; -} - pub fn typedCustomCall( target_name: [:0]const u8, opts: CustomCallOptions, @@ -2781,7 +2835,7 @@ pub fn typedCustomCall( const ctx = Compiler.current(); const allocator = ctx.arena.allocator(); - stdx.debug.assert(!opts.has_side_effect or ctx.manual_computation_depth > 0, "side-effect customCall '{s}' must be emitted inside manualComputation", .{target_name}); + stdx.debug.assert(ctx.manual_computation_depth > 0, "customCall '{s}' must be emitted inside manualComputation", .{target_name}); const input_tensors: []const Tensor = switch (@typeInfo(Input)) { .@"struct" => |struct_info| b: { @@ -2791,7 +2845,6 @@ pub fn typedCustomCall( } break :b &input_tensors; }, - // Extra case to support []const Tensor from mesheAwareTypedCustomCall .pointer => |pointer_info| b: { if (pointer_info.size != .slice) @compileError("Expected input slice"); break :b input; @@ -2807,7 +2860,6 @@ pub fn typedCustomCall( } break :b &output_shapes; }, - // Extra case to support []const Shape from mesheAwareTypedCustomCall .pointer => |pointer_info| b: { if (pointer_info.size != .slice) @compileError("Expected input slice"); break :b output; @@ -2872,7 +2924,7 @@ pub fn typedCustomCall( ctx.location, ).appendTo(ctx.currentScope().block); - if (ctx.manual_computation_depth > 0 and ctx.partitioner == .gspmd) { + if (ctx.partitioner == .gspmd) { op.setAttributeByName("mhlo.sharding", .string(ctx.mlir_ctx, "{manual}")); } @@ -2884,7 +2936,6 @@ pub fn typedCustomCall( } return out; }, - // Extra case to support []const Shape from mesheAwareTypedCustomCall .pointer => |pointer_info| { if (pointer_info.size != .slice) @compileError("Expected input slice"); var out: []Tensor = allocator.alloc(Tensor, output_shapes.len) catch unreachable; @@ -2937,7 +2988,7 @@ test customCall { const platform = zml.testing.env(); const tp = zml.testing.tpMesh(); - var comp: zml.Compiler = .init(std.testing.allocator, std.testing.io, platform, .{}); + var comp: zml.Compiler = .init(std.testing.allocator, std.testing.io, platform, .{ .meshes = &.{tp} }); defer comp.deinit(); comp.activate(); defer comp.deactivate(); @@ -2948,9 +2999,17 @@ test customCall { const shape = zml.Shape.init(.{128}, .bf16).withPartitioning(tp, .{ ._0 = .model }); const input = Tensor.constant(zml.DataType.bf16.constant(0)).broad(shape); - const output = customCall("my_custom_call", .{input}, .{zml.Shape.init(.{128}, .bf16)}, .{}, .{ - .has_side_effect = false, - .output_operand_aliases = &.{.{ .output_index = 0, .operand_index = 0 }}, + const Local = struct { + fn body(x: Tensor, output_shape: Shape) Tensor { + return customCall("my_custom_call", .{x}, .{output_shape}, .{}, .{ + .has_side_effect = false, + .output_operand_aliases = &.{.{ .output_index = 0, .operand_index = 0 }}, + }); + } + }; + const output = manualComputation(Local.body, input, shape, tp, .{ + .in_specs = .parse(.{ ._0 = .model }), + .out_specs = .parse(.{ ._0 = .model }), }); try zml.testing.expectEqualShapes(input.shape(), output.shape()); diff --git a/zml/ops/manual_computation.zig b/zml/ops/manual_computation.zig index c9596f98..a0ec0037 100644 --- a/zml/ops/manual_computation.zig +++ b/zml/ops/manual_computation.zig @@ -1,49 +1,131 @@ const std = @import("std"); -const builtin = @import("builtin"); const mlir = @import("mlir"); const shardy = @import("mlir/dialects/shardy"); const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); -const zml = struct { - // Nested to disambiguate from stdx.meta - const meta = @import("../meta.zig"); - const testing = @import("../testing.zig"); -}; - const Compiler = @import("../Compiler.zig"); -const DataType = @import("../dtype.zig").DataType; +const meta = @import("../meta.zig"); const mlirx = @import("../mlirx.zig"); const Shape = @import("../shape.zig").Shape; const Sharding = @import("../Sharding.zig"); const Tensor = @import("../tensor.zig").Tensor; + const log = std.log.scoped(.@"zml/Compiler"); -/// ZML operator describes computation across several devices, -/// but `manualComputation` allows to partition the devices in small group, -/// split the `Tensor` on those partitions, -/// and describes sub-computation that work inside a partition. +/// Partitioning declarations for the tensors in `T`. +/// Each `Tensor` becomes a `Sharding.Partitioning.Logical`. +/// For example, `struct { x: Tensor, bias: ?Tensor, scale: f32 }` becomes +/// `struct { x: Logical, bias: ?Logical }`. /// -/// See: -/// https://github.com/openxla/shardy/blob/main/docs/sdy_dialect.md#sdymanual_computation-sdymanualcomputationop +/// Initialize individual fields to declare different shardings, or use +/// `partitionSpecs` to apply one spec to every tensor in a value. +pub fn PartitionSpecs(comptime T: type) type { + return meta.MapRestrict(Tensor, Sharding.Partitioning.Logical).map(T); +} + +/// Declares the same sharding for a group of tensors at a `manualComputation` +/// boundary. /// -/// * `body_fn` must be `fn (inputs: I, output_shapes: O) R` where: -/// - `I` is user defined and contains inputs tensors and any metadata required by the computation -/// - `O` contains the partionned output shapes, types must be `void`, `Shape` or `[]const Shape` for 0, 1 or many outputs -/// - `R` must be `void`, `Tensor, or `[]const Tensor` for 0, 1 or many outputs -/// * `inputs` must be using the same mesh (using .withPartitioning), or be fully replicated -/// * `outputs` contains the global output shapes, -/// and `manualComputation` will pass to `body_fn` the shapes of each local output shard. +/// ```zig +/// .in_specs = .{ +/// .query = .parse(.{ .h = .model }), +/// .cache = partitionSpecs(cache, .parse(.{ .h = .model })), +/// .metadata = partitionSpecs(metadata, .replicated), +/// } +/// ``` /// -/// This depart from the rest of ZML where output shapes are generally inferred, -/// but in this case since `body_fn` explains how to compute parts of the result, -/// it's not clear how to put them back together, -/// hence why the user is asked to explicitly provided the expected output shape. +/// Requires an active compiler and uses its arena for allocations. Use +/// `partitionSpecsAlloc` outside graph construction. Use named struct fields +/// when mixing tensors with non-tensor metadata. +pub fn partitionSpecs(tensors: anytype, spec: Sharding.Partitioning.Logical) PartitionSpecs(@TypeOf(tensors)) { + return partitionSpecsAlloc(Compiler.current().arena.allocator(), tensors, spec) catch @panic("OOM"); +} + +/// Builds the same declarations as `partitionSpecs` using an explicit allocator. +pub fn partitionSpecsAlloc(allocator: std.mem.Allocator, tensors: anytype, spec: Sharding.Partitioning.Logical) std.mem.Allocator.Error!PartitionSpecs(@TypeOf(tensors)) { + var specs: PartitionSpecs(@TypeOf(tensors)) = undefined; + try meta.mapAlloc(struct { + fn assign(spec_: Sharding.Partitioning.Logical, _: Tensor) Sharding.Partitioning.Logical { + return spec_; + } + }.assign, allocator, spec, tensors, &specs); + return specs; +} + +/// Describes how the inputs and outputs of `body_fn` are partitioned across a mesh. +/// Returns the type of the `partitioning` argument to `manualComputation`. +/// +/// * `in_specs` specifies the sharding of each tensor in the body's first argument. +/// It follows the argument's structure, omitting non-tensor metadata. +/// * `out_specs` specifies the sharding of each shape in the body's second argument. +/// It follows the output shapes' structure and can be omitted for a `void` argument. +/// +/// Each spec is a `Sharding.Partitioning.Logical`. A mapping such as +/// `.parse(.{ .hidden = .model })` splits the tensor's `.hidden` dimension across +/// the mesh's `.model` axis. Omitted dimensions are replicated; `.replicated` +/// replicates the entire tensor. The specs must not contain open dimensions. +/// Slice lengths, optional presence and active union tags must match the arguments. +/// +/// For a body taking `struct { x: Tensor, bias: Tensor, scale: f32 }` and a single +/// output `Shape`, the partitioning can be declared as: +/// +/// ```zig +/// const partitioning: ManualPartitioning(body) = .{ +/// .in_specs = .{ .x = .parse(.{ .hidden = .model }), .bias = .replicated }, +/// .out_specs = .parse(.{ .hidden = .model }), +/// }; +/// ``` +pub fn ManualPartitioning(comptime body_fn: anytype) type { + const Logical = Sharding.Partitioning.Logical; + + const Inputs = stdx.meta.FnParam(body_fn, 0); + const OutputShapes = stdx.meta.FnParam(body_fn, 1); + + const InputSpecs = PartitionSpecs(Inputs); + const OutputSpecs = meta.MapRestrict(Shape, Logical).map(OutputShapes); + + if (OutputSpecs == void) return struct { + in_specs: InputSpecs, + out_specs: void = {}, + }; + + return struct { + in_specs: InputSpecs, + out_specs: OutputSpecs, + }; +} + +/// Describes a computation on local shards of tensors distributed across `mesh`. +/// ZML operations normally describe work on global tensors. The partitioner +/// distributes the inputs according to `in_specs`, and `body_fn` describes the work +/// performed on each shard. The results are exposed to the surrounding graph as +/// global tensors with the sharding declared by `out_specs`. +/// +/// * `body_fn` must be `fn (inputs: I, output_shapes: O) R`, where: +/// - `I` contains the local input tensors and any metadata needed by the body. +/// - `O` is `void`, `Shape`, or a container of shapes describing the local outputs. +/// - `R` is `void`, `Tensor`, or a container of tensors matching those local shapes +/// in traversal order. +/// * `inputs` contains the global tensors and metadata. The partitioner reshards +/// the tensors as needed to satisfy `in_specs`; metadata is passed to the body unchanged. +/// * `output_shapes` contains the global output shapes. The body receives their +/// local shard shapes, computed using `out_specs` and the sizes of the mesh axes. +/// * `mesh` is the logical mesh whose axis names are referenced by the specs. +/// * `partitioning` declares the input and output shardings; see `ManualPartitioning`. +/// Each partitioned dimension must be divisible by the size of its mesh axis. +/// +/// For example, splitting a dimension of size 64 across a mesh axis of size 8 gives +/// the body a dimension of size 8. A replicated dimension keeps its global size. +/// The caller supplies the global output shapes because the body only describes +/// local results: their shapes alone do not determine the global tensor shapes. pub fn manualComputation( comptime body_fn: anytype, inputs: stdx.meta.FnParam(body_fn, 0), output_shapes: stdx.meta.FnParam(body_fn, 1), + mesh: *const Sharding.Mesh, + partitioning: ManualPartitioning(body_fn), ) stdx.meta.FnReturn(body_fn) { const Inputs = stdx.meta.FnParam(body_fn, 0); const OutputShapes = stdx.meta.FnParam(body_fn, 1); @@ -54,13 +136,42 @@ pub fn manualComputation( defer arena_state.deinit(); const arena = arena_state.allocator(); - var owned_inputs: Inputs = inputs; - const flat_input_tensors: []*Tensor = zml.meta.collectMutPtrs(Tensor, arena, &owned_inputs) catch @panic("OOM"); - defer arena.free(flat_input_tensors); + // The body may return containers borrowed from its inputs or output shapes. + const body_arena = compiler.arena.allocator(); - var owned_output_shapes: OutputShapes = output_shapes; - const flat_output_shapes: []*Shape = zml.meta.collectMutPtrs(Shape, arena, &owned_output_shapes) catch @panic("OOM"); - defer arena.free(flat_output_shapes); + var owned_inputs: Inputs = undefined; + meta.mapAlloc((struct { + fn copy(_: void, tensor: Tensor) Tensor { + return tensor; + } + }).copy, body_arena, {}, inputs, &owned_inputs) catch @panic("OOM"); + + const flat_input_tensors: []*Tensor = meta.collectMutPtrs(Tensor, arena, &owned_inputs) catch @panic("OOM"); + const in_specs = meta.collectValues(Sharding.Partitioning.Logical, arena, &partitioning.in_specs) catch @panic("OOM"); + stdx.debug.assert(flat_input_tensors.len == in_specs.len, "manualComputation has {} input tensors but {} input specs", .{ flat_input_tensors.len, in_specs.len }); + + for (flat_input_tensors, in_specs) |tensor, spec| { + tensor._shape = tensor.shape().withPartitioning(mesh, spec); + stdx.debug.assert(!tensor.shape().containsPartitionSpec(.open), "manualComputation input sharding must be closed: {f}", .{tensor.shape()}); + } + + var local_output_shapes: OutputShapes = undefined; + meta.mapAlloc((struct { + fn copy(_: void, shape: Shape) Shape { + return shape; + } + }).copy, body_arena, {}, output_shapes, &local_output_shapes) catch @panic("OOM"); + + const flat_local_output_shapes = meta.collectMutPtrs(Shape, arena, &local_output_shapes) catch @panic("OOM"); + const out_specs = meta.collectValues(Sharding.Partitioning.Logical, arena, &partitioning.out_specs) catch @panic("OOM"); + stdx.debug.assert(flat_local_output_shapes.len == out_specs.len, "manualComputation has {} output shapes but {} output specs", .{ flat_local_output_shapes.len, out_specs.len }); + + const flat_output_shapes = arena.alloc(Shape, flat_local_output_shapes.len) catch @panic("OOM"); + for (flat_output_shapes, flat_local_output_shapes, out_specs) |*global, local, spec| { + global.* = local.withPartitioning(mesh, spec); + stdx.debug.assert(!global.containsPartitionSpec(.open), "manualComputation output sharding must be closed: {f}", .{global}); + local.* = Sharding.shardedShape(global.*) catch std.debug.panic("can't shard output {f} for {f}", .{ global, mesh }); + } const WrappedInputs = struct { inputs: *const Inputs, @@ -72,15 +183,17 @@ pub fn manualComputation( outputs_ptr.* = body_fn(wrapped.inputs.*, wrapped.output_shapes.*); - return try zml.meta.collectMutPtrs(Tensor, allocator, outputs_ptr); + return try meta.collectMutPtrs(Tensor, allocator, outputs_ptr); } }; var outputs_struct: Outputs = undefined; - var wrapped_inputs: WrappedInputs = .{ .inputs = &owned_inputs, .output_shapes = &owned_output_shapes }; + var wrapped_inputs: WrappedInputs = .{ .inputs = &owned_inputs, .output_shapes = &local_output_shapes }; + _ = manualComputationInternal( compiler, arena, + mesh, @typeName(@TypeOf(body_fn)), &wrapped_inputs, flat_input_tensors, @@ -90,6 +203,7 @@ pub fn manualComputation( ) catch |err| switch (err) { error.OutOfMemory => @panic("OOM"), }; + return outputs_struct; } @@ -100,27 +214,21 @@ const LocalComputationFn = *const fn (std.mem.Allocator, *const anyopaque, *anyo noinline fn manualComputationInternal( compiler: *Compiler, arena: std.mem.Allocator, + mesh: *const Sharding.Mesh, functor_name: []const u8, input_struct: *anyopaque, input_tensors: []*Tensor, output_struct: *anyopaque, - output_shapes: []*Shape, + global_output_shapes: []const Shape, local_fn: LocalComputationFn, ) error{OutOfMemory}![]*Tensor { const global_scope = compiler.currentScope(); const mlir_ctx = compiler.mlir_ctx; - const mesh = switch (chooseCommonMesh(input_tensors, compiler.platform.replicated_mesh)) { - .mesh => |m| m, - .incompatible => |incomp| std.debug.panic("manualComputation expects all inputs and outputs to use the same mesh, got input {}: {} and input {}: {}", .{ incomp[0], input_tensors[incomp[0]], incomp[1], input_tensors[incomp[1]] }), - }; - const manual_axes = collectManualAxes(mesh, input_tensors); - log.debug("manualComputation({s}) will use mesh {s}, manual axes: {f}", .{ functor_name, mesh.name, stdx.fmt.stringsZ(manual_axes.constSlice()) }); + log.debug("manualComputation({s}) will use mesh {s}", .{ functor_name, mesh.name }); const global_input_shapes = try arena.alloc(Shape, input_tensors.len); for (global_input_shapes, input_tensors) |*shape, tensor| shape.* = tensor._shape; - const global_output_shapes = try arena.alloc(Shape, output_shapes.len); - for (global_output_shapes, output_shapes) |*shape, shape_ptr| shape.* = shape_ptr.*; const input_types = try arena.alloc(*const mlir.Type, input_tensors.len); for (input_types, input_tensors) |*input_ty, tensor_ptr| { @@ -134,11 +242,17 @@ noinline fn manualComputationInternal( const input_values = try arena.alloc(*const mlir.Value, input_tensors.len); for (input_values, input_tensors) |*input_val, tensor_ptr| { + if (compiler.partitioner == .gspmd) { + tensor_ptr.* = tensor_ptr.withPartitioning(mesh, tensor_ptr.shape()._sharding.partition); + } input_val.* = global_scope.value(tensor_ptr.*); } - const manual_block = mlir.Block.init(input_types, input_locs); - errdefer manual_block.deinit(); + const manual_block = switch (compiler.partitioner) { + .shardy => mlir.Block.init(input_types, input_locs), + .gspmd => global_scope.block, + }; + errdefer if (compiler.partitioner == .shardy) manual_block.deinit(); const local_scope = compiler.pushBlock(manual_block); defer local_scope.pop(); @@ -150,17 +264,14 @@ noinline fn manualComputationInternal( }; } - for (output_shapes, global_output_shapes) |shape_ptr, global| { - shape_ptr.* = Sharding.shardedShape(global) catch std.debug.panic("can't shard output {f} for {f}", .{ global, mesh }); - } - compiler.manual_computation_depth += 1; defer compiler.manual_computation_depth -= 1; const local_outputs = try local_fn(arena, input_struct, output_struct); stdx.debug.assert(local_outputs.len == global_output_shapes.len, "manualComputation body returned {} values, expected {}", .{ local_outputs.len, global_output_shapes.len }); - for (0.., local_outputs, output_shapes) |i, local_tensor, local_shape| { - stdx.debug.assert(local_shape.eql(local_tensor._shape), "manualComputation body output {d} expected {f}, got {f}", .{ i, local_shape.*, local_tensor._shape }); + for (0.., local_outputs, global_output_shapes) |i, local_tensor, global_shape| { + const expected_local = Sharding.shardedShape(global_shape) catch std.debug.panic("can't shard output {f} for {f}", .{ global_shape, mesh }); + stdx.debug.assert(expected_local.eql(local_tensor._shape), "manualComputation body output {d} expected {f}, got {f}", .{ i, expected_local, local_tensor._shape }); } const global_output_types = try arena.alloc(*const mlir.Type, global_output_shapes.len); @@ -214,97 +325,14 @@ noinline fn manualComputationInternal( }, }; - // reuse local_outputs buffer to store the final result. + // Returned tensors describe the operation's global output shapes and sharding. for (0.., local_outputs, global_output_shapes) |o, output_tensor_ptr, output_shape| { - output_tensor_ptr.* = Tensor._resultPropagateSharding(output_shape, op.result(o)); + output_tensor_ptr.* = Tensor._result(output_shape, op.result(o)); + output_tensor_ptr._shape._sharding = output_shape._sharding; } return local_outputs; } -const ChosenMesh = union(enum) { - mesh: *const Sharding.Mesh, - incompatible: [2]u32, -}; - -fn chooseCommonMesh(inputs: []const *const Tensor, fallback: *const Sharding.Mesh) ChosenMesh { - var mesh: ?*const Sharding.Mesh = null; - var mesh_source_id: u32 = @intCast(inputs.len); - for (0.., inputs) |i, input| { - if (input.shape()._sharding.mesh) |shape_mesh| { - if (mesh) |m| { - if (m != shape_mesh) - return .{ .incompatible = .{ mesh_source_id, @intCast(i) } }; - } else { - mesh = shape_mesh; - mesh_source_id = @intCast(i); - } - } - } - return .{ .mesh = mesh orelse fallback }; -} - -test chooseCommonMesh { - const fallback: Sharding.Mesh = .initTest(.fallback, .{.all}); - const first: Sharding.Mesh = .initTest(.first, .{ .data, .model }); - const second: Sharding.Mesh = .initTest(.second, .{ .experts, .model }); - const replicated = Tensor.init(.{ .a = 8 }, .f32); - const sharded = Tensor.init(.{ .a = 8 }, .f32).withPartitioning(&first, .{ .a = .data }); - const other = Tensor.init(.{ .a = 8 }, .f32).withPartitioning(&second, .{ .a = .experts }); - - try std.testing.expectEqual(ChosenMesh{ .mesh = &fallback }, chooseCommonMesh(&.{}, &fallback)); - try std.testing.expectEqual(ChosenMesh{ .mesh = &fallback }, chooseCommonMesh(&.{&replicated}, &fallback)); - try std.testing.expectEqual(ChosenMesh{ .mesh = &first }, chooseCommonMesh(&.{&sharded}, &fallback)); - try std.testing.expectEqual(ChosenMesh{ .mesh = &first }, chooseCommonMesh(&.{ &replicated, &sharded, &sharded }, &fallback)); - try std.testing.expectEqual(ChosenMesh{ .incompatible = .{ 0, 1 } }, chooseCommonMesh(&.{ &sharded, &other }, &fallback)); -} - -fn collectManualAxes(mesh: *const Sharding.Mesh, inputs: []const *const Tensor) stdx.BoundedArray(Shape.Tag, Sharding.MAX_MESH_RANK) { - var used: [Sharding.MAX_MESH_RANK]bool = @splat(false); - for (inputs) |input| { - const sharding = input.shape()._sharding; - if (sharding.mesh == null) continue; - std.debug.assert(sharding.mesh.? == mesh); - for (0..input.rank()) |dim| { - if (sharding.partition.get(dim).meshAxis()) |axis| used[axis] = true; - } - } - - var axes: stdx.BoundedArray(Shape.Tag, Sharding.MAX_MESH_RANK) = .empty; - for (mesh.logical.axes.constSlice(), 0..) |axis, i| { - if (used[i]) axes.appendAssumeCapacity(axis); - } - return axes; -} - -test collectManualAxes { - const mesh: Sharding.Mesh = .initTest(.manual, .{ .data, .model, .unused }); - var data = Tensor.init(.{ .a = 8 }, .f32).withPartitioning(&mesh, .{ .a = .data }); - var model = Tensor.init(.{ .a = 8 }, .f32).withPartitioning(&mesh, .{ .a = .model }); - const replicated = Tensor.init(.{ .a = 8 }, .f32); - const open = Tensor.init(.{ .a = 8 }, .f32).withPartitioning(&mesh, .{ .a = .open }); - - const axes = collectManualAxes(&mesh, &.{ &model, &replicated, &data, &model, &open }); - data = replicated; - model = replicated; - try std.testing.expectEqual(@as(usize, 2), axes.len); - try std.testing.expectEqualStrings("data", std.mem.span(axes.get(0))); - try std.testing.expectEqualStrings("model", std.mem.span(axes.get(1))); - try std.testing.expectEqual(@as(usize, 0), collectManualAxes(&mesh, &.{ &replicated, &open }).len); - try std.testing.expectEqual(@as(usize, 0), collectManualAxes(&mesh, &.{}).len); - - const registry: *mlir.DialectRegistry = try .init(); - defer registry.deinit(); - shardy.registerDialect(registry); - const ctx: *mlir.Context = try .init(.{ .registry = registry, .threading = false }); - defer ctx.deinit(); - ctx.loadAllAvailableDialects(); - const expected = shardy.ManualAxesAttr.get(ctx, &.{ - .init(ctx, "data"), - .init(ctx, "model"), - }).attribute(); - try std.testing.expect(shardy_.manualAxesAttr(ctx, axes.constSlice()).eql(expected)); -} - const gspmd = struct { fn globalToLocal(mlir_ctx: *mlir.Context, scope: *Compiler.Scope, local_tensor: Tensor, local_shape: Shape, location: *const mlir.Location) Tensor { const local_type = mlirx.Type.rankedTensor(mlir_ctx, local_shape); @@ -347,54 +375,172 @@ const gspmd = struct { } }; -const shardy_ = struct { - pub fn manualAxesAttr(mlir_ctx: *mlir.Context, axes: []const Shape.Tag) *const mlir.Attribute { - var attrs: stdx.BoundedArray(*const mlir.StringAttribute, Sharding.MAX_MESH_RANK) = .empty; - for (axes) |axis| attrs.appendAssumeCapacity(.init(mlir_ctx, std.mem.span(axis))); - return shardy.ManualAxesAttr.get(mlir_ctx, attrs.constSlice()).attribute(); - } -}; - -test "manualComputation passes local output shapes to the callback" { +test manualComputation { + const zml = @import("../zml.zig"); const allocator = std.testing.allocator; const io = std.testing.io; const platform = zml.testing.env(); const mesh = zml.testing.tpMesh(); - const partitions = mesh.numPartitionsForLogicalAxis(.model); - if (partitions == 1) return error.SkipZigTest; - var compiler: Compiler = .init(allocator, io, platform, .{ .meshes = &.{mesh} }); - defer compiler.deinit(); - compiler.activate(); - defer compiler.deactivate(); - - const block = mlir.Block.init(&.{}, &.{}); - defer block.deinit(); - const scope = compiler.pushBlock(block); - defer scope.pop(); - - const global_shape = Shape.init(.{ .b = 4, .h = 3 * partitions, .hd = 128 }, .bf16) - .withPartitioning(mesh, .{ .h = .model }); - const local_shape = Shape.init(.{ .b = 4, .h = 3, .hd = 128 }, .bf16); - const input = Tensor.zeroes(global_shape).withPartitioning(mesh, .{ .h = .model }); - var received_shape: ?Shape = null; - const Context = struct { - input: Tensor, - received_shape: *?Shape, + const LocalSum = struct { + x: Tensor, + bias: Tensor, + scale: i32, fn body(self: @This(), output_shape: Shape) Tensor { - self.received_shape.* = output_shape; - return self.input; + return self.x.sum(.x).scale(self.scale).add(self.bias).broad(output_shape); + } + + fn forward(x: Tensor, mesh_: *const Sharding.Mesh) Tensor { + return manualComputation( + body, + .{ .x = x, .bias = .scalar(10, .i32), .scale = 2 }, + x.shape(), + mesh_, + .{ .in_specs = .{ .x = .parse(.{ .x = .model }), .bias = .replicated }, .out_specs = .parse(.{ .x = .model }) }, + ); } }; - const output = manualComputation( - Context.body, - .{ .input = input, .received_shape = &received_shape }, - global_shape, - ); - try std.testing.expect(received_shape != null); - try std.testing.expect(received_shape.?.eql(local_shape)); - try std.testing.expect(output.shape().eql(global_shape)); - try std.testing.expectEqualDeep(global_shape._sharding, output.shape()._sharding); + // Each shard sums two elements and broadcasts the result to its local output shape. + const shape: Shape = .init(.{ .x = 2 * mesh.numPartitionsForLogicalAxis(.model) }, .i32); + const values = try allocator.alloc(i32, shape.count()); + defer allocator.free(values); + + for (values, 0..) |*value, i| value.* = @intCast(i + 1); + + var input = try zml.Buffer.fromBytes(io, platform, shape, std.mem.sliceAsBytes(values)); + defer input.deinit(); + + for ([_]Sharding.Partitioner{ .shardy, .gspmd }) |partitioner| { + const exe = try platform.compileFn(allocator, io, LocalSum.forward, .{ .fromShape(shape), mesh }, .{ + .meshes = &.{ platform.replicated_mesh, mesh }, + .partitioner = partitioner, + }); + defer exe.deinit(); + + var output = try exe.eval(allocator, io, .{input}); + defer output.deinit(); + + const actual = try output.toSliceAlloc(allocator, io); + defer actual.free(allocator); + + for (actual.constItems(i32), 0..) |value, i| { + const pair = values[i / 2 * 2 ..][0..2]; + try std.testing.expectEqual((pair[0] + pair[1]) * 2 + 10, value); + } + + try std.testing.expect(output.shape().eqlWithTags(shape)); + try std.testing.expectEqualDeep(shape.withPartitioning(mesh, .{ .x = .model })._sharding, output.shape()._sharding); + } +} + +test "manualComputation composes across logical and physical meshes" { + const zml = @import("../zml.zig"); + if (comptime !zml.Target.cpu.isEnabled()) return error.SkipZigTest; + const allocator = std.testing.allocator; + const io = std.testing.io; + + const Topologies = enum { flat, torus }; + + const Model = struct { + fn increment(x: Tensor, _: Shape) Tensor { + return x.addConstant(1); + } + + fn forward(x: Tensor, mesh: *const Sharding.Mesh) Tensor { + const partitioning: ManualPartitioning(increment) = .{ + .in_specs = partitionSpecs(x, .parse(.{ .x = .model })), + .out_specs = .parse(.{ .x = .model }), + }; + + const first = manualComputation(increment, x, x.shape(), mesh, partitioning); + const converted = first.convert(.f32); + return manualComputation(increment, converted, converted.shape(), mesh, partitioning); + } + }; + + const shape: Shape = .init(.{ .x = 16 }, .i32); + + var values: [16]i32 = undefined; + var expected: [16]f32 = undefined; + for (&values, &expected, 0..) |*value, *result, i| { + value.* = @intCast(i); + result.* = @floatFromInt(i + 2); + } + + for ([_]Topologies{ .flat, .torus }) |topology| { + const platform = try zml.Platform.init(allocator, io, .cpu, .{ + .cpu = .{ .device_count = 8 }, + .physical_mesh = if (topology == .torus) .{ .custom = Sharding.PhysicalMesh.torus3d } else .auto, + }); + defer platform.deinit(allocator, io); + + // Fold three physical axes into one logical axis, or split one into three. + const logical: Sharding.LogicalMesh = if (topology == .torus) + try .init(.{.model}, &.{.balanced}, &.{8}) + else + try Sharding.LogicalMesh.init(.{ .data, .model, .pipeline }, &.{ .low_bandwidth, .high_bandwidth, .balanced }, &.{ 2, 2, 2 }); + + const mesh = try platform.registerMesh("manual", logical); + + var input: zml.Buffer = try .fromBytes(io, platform, shape, std.mem.asBytes(&values)); + defer input.deinit(); + + for ([_]Sharding.Partitioner{ .shardy, .gspmd }) |partitioner| { + const exe = try platform.compileFn(allocator, io, Model.forward, .{ Tensor.fromShape(shape), mesh }, .{ + .meshes = &.{ platform.replicated_mesh, mesh }, + .partitioner = partitioner, + }); + defer exe.deinit(); + + var output = try exe.eval(allocator, io, .{input}); + defer output.deinit(); + + try std.testing.expectEqual(expected, try output.getValue([16]f32, io)); + try std.testing.expect(output.shape().eqlWithTags(shape.withDtype(.f32))); + try std.testing.expectEqualDeep(shape.withPartitioning(mesh, .{ .x = .model })._sharding, output.shape()._sharding); + } + } +} + +test "partitionSpecs preserves optional tensors and active cache variants" { + const zml = @import("../zml.zig"); + const Logical = Sharding.Partitioning.Logical; + const Inputs = struct { + layers: []const Tensor, + bias: ?Tensor, + cache: union(enum) { disabled, values: [2]Tensor }, + scale: f32, + }; + + var arena: std.heap.ArenaAllocator = .init(std.testing.allocator); + defer arena.deinit(); + + const tensor: Tensor = .fromShape(.init(.{ .h = 8, .d = 16 }, .f32)); + var inputs: Inputs = .{ + .layers = &.{ tensor, tensor }, + .bias = null, + .cache = .{ .values = .{ tensor, tensor } }, + .scale = 2, + }; + const heads: Logical = .parse(.{ .h = .model }); + const specs = try zml.ops.partitionSpecsAlloc(arena.allocator(), inputs, heads); + const expected: zml.ops.PartitionSpecs(Inputs) = .{ + .layers = &.{ heads, heads }, + .bias = null, + .cache = .{ .values = .{ heads, heads } }, + }; + try std.testing.expectEqualDeep(expected, specs); + + inputs.layers = &.{}; + inputs.bias = tensor; + inputs.cache = .disabled; + const replicated = try zml.ops.partitionSpecsAlloc(arena.allocator(), inputs, .replicated); + const expected_replicated: zml.ops.PartitionSpecs(Inputs) = .{ + .layers = &.{}, + .bias = .replicated, + .cache = .disabled, + }; + try std.testing.expectEqualDeep(expected_replicated, replicated); } diff --git a/zml/platform.zig b/zml/platform.zig index 3b52c08e..54fb90e4 100644 --- a/zml/platform.zig +++ b/zml/platform.zig @@ -8,6 +8,8 @@ pub const Target = @import("platforms").Platform; const constants = @import("constants.zig"); const Exe = @import("exe.zig").Exe; +const flashinfer_cutlass_contract = @import("kernels/moe/flashinfer_cutlass/contract.zig"); +const flashinfer_cutlass_runtime = @import("kernels/moe/flashinfer_cutlass/runtime.zig"); const pjrtx = @import("pjrtx.zig"); const profiler_ = @import("profiling/profiler.zig"); const Sharding = @import("Sharding.zig"); @@ -257,12 +259,12 @@ pub const State = union(Target) { furiosa: void, pub const CudaState = struct { - fi_cutlass_moe_runners: ?*zml.moe.cutlass_flashinfer.Runners = null, + flashinfer_cutlass_moe_runners: ?*flashinfer_cutlass_runtime.Runners = null, fn deinit(self: *CudaState) void { - if (self.fi_cutlass_moe_runners) |runners| { + if (self.flashinfer_cutlass_moe_runners) |runners| { runners.deinit(); - self.fi_cutlass_moe_runners = null; + self.flashinfer_cutlass_moe_runners = null; } } }; @@ -381,16 +383,9 @@ pub const Platform = struct { switch (target) { .cuda => { - if (zml.moe.cutlass_flashinfer.load(arena, io, platform)) { - zml.moe.cutlass_flashinfer.register(platform) catch |err| { - log.warn( - "Failed to register FlashInfer CUTLASS MoE custom calls: {}", - .{err}, - ); - }; - } else |err| { - log.warn("Failed to load FlashInfer CUTLASS MoE: {}", .{err}); - } + cuda.initFlashinferCutlass(platform, io) catch |err| { + log.warn("Failed to initialize Flashinfer CUTLASS MoE: {}", .{err}); + }; }, else => {}, } @@ -865,6 +860,12 @@ pub const cuda = struct { } }; + /// SM count of an addressable CUDA device; unavailable attributes return null. + pub fn coreCount(platform: *const zml.Platform, device_index: usize) ?u32 { + const attribute = platform.devices[device_index].pjrt_desc.attribute(platform.pjrt_api, "core_count") orelse return null; + return std.math.cast(u32, attribute.int64); + } + pub fn computeCapability(platform: *const zml.Platform) ?ComputeCapability { if (platform.target != .cuda) return null; const devices = platform.pjrt_client.devices(platform.pjrt_api); @@ -877,6 +878,29 @@ pub const cuda = struct { } } else null; } + + fn initFlashinferCutlass(platform: *Platform, io: std.Io) !void { + if (comptime !Target.cuda.isEnabled()) return error.UnsupportedPlatform; + const cc = computeCapability(platform) orelse return error.UnsupportedArchitecture; + + const allocator = platform.arena.allocator(); + const runners = try allocator.create(flashinfer_cutlass_runtime.Runners); + errdefer allocator.destroy(runners); + + try flashinfer_cutlass_runtime.load(allocator, io, cc.sm()); + + runners.* = .{}; + errdefer runners.deinit(); + + try platform.registerFfi(.{ + .name = flashinfer_cutlass_contract.name, + .handler = flashinfer_cutlass_runtime.handler, + .traits = .{ .command_buffer_compatible = true }, + .platform_name = if (flashinfer_cutlass_contract.compute_on_host) "Host" else null, + }); + + platform.state.cuda.flashinfer_cutlass_moe_runners = runners; + } }; pub const rocm = struct { @@ -991,12 +1015,15 @@ fn printCallback(call_frame: *pjrt.ffi.CallFrame) callconv(.c) ?*pjrt.ffi.Error fn printCallbackInner(call_frame: *pjrt.ffi.CallFrame) !?*pjrt.ffi.Error { if (call_frame.registeringHook()) return null; + const buffers = call_frame.args.buffers(); + if (buffers.len <= 0) return null; + const pjrt_api: *pjrt.Api = @ptrFromInt(getScalarAttributeAs(u64, call_frame, "pjrt_api").?); const pjrt_client: *pjrt.Client = @ptrFromInt(getScalarAttributeAs(u64, call_frame, "pjrt_client").?); const device_ordinal: usize = @intCast(try call_frame.ctx.getDeviceOrdinal(call_frame.api)); - const buffer = call_frame.args.buffers()[0]; + const buffer = buffers[0]; const shape = shapeFromFfiBuffer(buffer); // NOTE(Corentin): This is a hack. We take the first non device memory, hoping that it's host visible, @@ -1020,9 +1047,13 @@ fn printCallbackInner(call_frame: *pjrt.ffi.CallFrame) !?*pjrt.ffi.Error { }, }, }); + defer pjrt_buffer.deinit(pjrt_api); pjrt_buffer = try pjrt_buffer.copyToMemory(pjrt_api, first_non_device_memory); - try pjrt_buffer.readyEvent(pjrt_api).awaitRaw(pjrt_api); + + const event = pjrt_buffer.readyEvent(pjrt_api); + defer event.deinit(pjrt_api); + try event.awaitRaw(pjrt_api); const host_visible_data: [*]u8 = @ptrCast(@alignCast(try pjrt_buffer.opaqueDeviceMemoryDataPointer(pjrt_api))); diff --git a/zml/quantization.zig b/zml/quantization.zig index 89e1b210..db20619a 100644 --- a/zml/quantization.zig +++ b/zml/quantization.zig @@ -129,6 +129,8 @@ pub fn quantizeInput(quantization: Quantization, input: Tensor, axis: Shape.Tag, .fp8_block32, .fp8_block128 => blk: { const dtype: DataType = switch (platform.target) { .cuda => .f8e4m3fn, + // RNGD contracts matching FP8 pairs in 128-wide K blocks. + .furiosa => if (quantization.scheme == .fp8_block128) .f8e4m3fn else return null, .rocm => if (platform_mod.rocm.computeCapability(platform)) |capability| switch (capability.architecture()) { .cdna4, .rdna4 => .f8e4m3fn, .cdna3 => .f8e4m3fnuz, diff --git a/zml/safetensors.zig b/zml/safetensors.zig index 63791ea9..3827fe27 100644 --- a/zml/safetensors.zig +++ b/zml/safetensors.zig @@ -2,68 +2,337 @@ const std = @import("std"); const DataType = @import("dtype.zig").DataType; const Shape = @import("shape.zig").Shape; -const Dims = Shape.DimsArray; - -const StringBuilder = std.ArrayListUnmanaged(u8); - -pub const Tensors = std.StringArrayHashMapUnmanaged(Tensor); -pub const Metadatas = std.StringArrayHashMapUnmanaged(Metadata); const log = std.log.scoped(.@"zml/safetensors"); -const BYTES_HEADER = 8; +// TODO: consider promoting this to the top level struct +// TODO: Feels like we are missing a mapping file -> tensors that would be helfpful for loading weights +/// Maps tensor names to their shape and positions in different sub files. +pub const TensorRegistry = struct { + arena: std.heap.ArenaAllocator, -pub const FileType = enum { - index, - safetensors, - unknown, + tensors: std.StringArrayHashMapUnmanaged(Tensor), + + /// Metadata found in `__metadata__` field of the safetensors header. + /// This is rarely use and the information there is not standardized, so we let user code handle it. + /// Here we just store the plain bytes, reach out if you need this to be already parsed. + metadata: []const u8, + + pub fn init(allocator: std.mem.Allocator) TensorRegistry { + return .{ + .arena = std.heap.ArenaAllocator.init(allocator), + .tensors = .{}, + .metadata = "", + }; + } + + pub fn deinit(self: *TensorRegistry) void { + self.arena.deinit(); + } + + pub fn fromRepo(allocator: std.mem.Allocator, io: std.Io, repo: std.Io.Dir) !TensorRegistry { + const file_type, const file = try resolveModelEntrypoint(io, repo); + defer file.close(io); + + var registry: TensorRegistry = .init(allocator); + switch (file_type) { + .safetensors => try registry.parseSafetensors(io, allocator, file), + .index => try registry.parseIndexJson(io, allocator, repo, file), + .full_index => try registry.parseFullIndexJson(io, allocator, repo, file), + .unknown => return error.InvalidPath, + } + return registry; + } + + pub fn fromPath(allocator: std.mem.Allocator, io: std.Io, path: []const u8) !TensorRegistry { + const file_type: FileType = .resolve(path); + switch (file_type) { + .safetensors => { + // open directly + const file = try std.Io.Dir.openFile(.cwd(), io, path, .{}); + var registry: TensorRegistry = .init(allocator); + try registry.parseSafetensors(io, allocator, file); + return registry; + }, + .full_index, + .index, + => { + // resolve repo as parent folder then open + const repo_path = std.fs.path.dirname(path) orelse return ModelPathResolutionError.InvalidPath; + const repo = std.Io.Dir.openDir(.cwd(), io, repo_path, .{}) catch |e| { + return switch (e) { + error.FileNotFound => ModelPathResolutionError.FileNotFound, + else => ModelPathResolutionError.InvalidPath, + }; + }; + const file = try std.Io.Dir.openFile(.cwd(), io, path, .{}); + + var registry: TensorRegistry = .init(allocator); + try switch (file_type) { + .index => registry.parseIndexJson(io, allocator, repo, file), + .full_index => registry.parseFullIndexJson(io, allocator, repo, file), + .safetensors, .unknown => unreachable, + }; + return registry; + }, + .unknown => { + // the path is possibly a folder, let's try that + const repo = std.Io.Dir.openDir(.cwd(), io, path, .{}) catch |e| { + return switch (e) { + error.FileNotFound => ModelPathResolutionError.FileNotFound, + else => ModelPathResolutionError.InvalidPath, + }; + }; + return try fromRepo(allocator, io, repo); + }, + } + } + + /// Register a tensor, duplicate tensor name and uri. + /// If you create a lot of tensors sharing the same uri, + /// considere calling `registry.tensors` directly and avoid duplicated uris. + pub fn registerTensor(self: *TensorRegistry, tensor: Tensor) !void { + const allocator = self.arena.allocator(); + + const owned_name = try allocator.dupe(u8, tensor.name); + errdefer allocator.free(owned_name); + const owned_uri = try allocator.dupe(u8, tensor.file_uri); + errdefer allocator.free(owned_uri); + + const owned_tensor: Tensor = .{ + .name = owned_name, + .file_uri = owned_uri, + .shape = tensor.shape, + .offset = tensor.offset, + }; + try self.tensors.put(allocator, owned_name, owned_tensor); + } + + pub fn reader( + self: *TensorRegistry, + io: std.Io, + tensor_name: []const u8, + buffer: []u8, + ) !TensorReader { + const tensor = self.tensors.get(tensor_name) orelse { + log.err("Tensor {s} not found in registry", .{tensor_name}); + return error.TensorNotFound; + }; + + return try .init(io, tensor, buffer, .{}); + } + + pub fn totalBytes(self: *TensorRegistry) u64 { + var total: u64 = 0; + + var it = self.tensors.iterator(); + while (it.next()) |entry| { + total += entry.value_ptr.byteSize(); + } + + return total; + } + + /// Parse a given .safetensors file and create entries for each tensor + pub fn parseSafetensors( + self: *TensorRegistry, + io: std.Io, + scratch_allocator: std.mem.Allocator, + file: std.Io.File, + ) !void { + var tmp_state = std.heap.ArenaAllocator.init(scratch_allocator); + defer tmp_state.deinit(); + const tmp_arena = tmp_state.allocator(); + + const self_arena = self.arena.allocator(); + var owned_file_name = try self_arena.alloc(u8, std.fs.max_path_bytes); + const file_name_len = try file.realPath(io, owned_file_name); + // Try to release extra memory, but we don't care if it fails + _ = self_arena.resize(owned_file_name, file_name_len); + owned_file_name = owned_file_name[0..file_name_len]; + + const header_buf = try tmp_arena.alloc(u8, 64 * 1024); + var file_reader: std.Io.File.Reader = file.reader(io, header_buf); + + const json_header_length: u64 = try file_reader.interface.takeInt(u64, .little); + const BYTES_HEADER = 8; + const data_start_offset = BYTES_HEADER + json_header_length; + const json_data = try tmp_arena.alloc(u8, @intCast(json_header_length)); + + try file_reader.interface.readSliceAll(json_data); + const headers: std.json.Value = try std.json.parseFromSliceLeaky( + std.json.Value, + tmp_arena, + json_data, + .{}, + ); + + try self.tensors.ensureUnusedCapacity(self_arena, headers.object.count()); + for (headers.object.keys(), headers.object.values()) |key, value| { + if (std.mem.eql(u8, key, "__metadata__")) { + // value belong to tmp arena, duplicate it into self arena. + var owned_metadata: std.Io.Writer.Allocating = .init(self_arena); + try std.json.fmt(value, .{ .whitespace = .minified }).format(&owned_metadata.writer); + self.metadata = owned_metadata.written(); + continue; + } + + const owned_name = try self_arena.dupe(u8, key); + const tensor: Tensor = try .parse(value.object, owned_file_name, owned_name, data_start_offset); + self.tensors.putAssumeCapacity(owned_name, tensor); + } + } + + pub fn parseIndexJson( + self: *TensorRegistry, + io: std.Io, + gpa: std.mem.Allocator, + repo: std.Io.Dir, + index_json_file: std.Io.File, + ) !void { + var tmp_state: std.heap.ArenaAllocator = .init(gpa); + defer tmp_state.deinit(); + const tmp = tmp_state.allocator(); + + const file_reader_buf = try tmp.alloc(u8, 64 * 1024); + var file_reader: std.Io.File.Reader = index_json_file.reader(io, file_reader_buf); + var json_reader: std.json.Reader = .init(tmp, &file_reader.interface); + const index = try std.json.parseFromTokenSourceLeaky( + std.json.Value, + tmp, + &json_reader, + .{ .allocate = .alloc_if_needed }, + ); + + var safetensors_files: std.StringArrayHashMapUnmanaged(void) = .empty; + try safetensors_files.ensureUnusedCapacity(tmp, 64); + if (index.object.get("weight_map")) |wm| { + for (wm.object.values()) |file_name| { + try safetensors_files.put(tmp, file_name.string, {}); + } + } else { + log.warn("No weight_map attribute found in index", .{}); + } + + for (safetensors_files.keys()) |file_name| { + const file = try repo.openFile(io, file_name, .{}); + defer file.close(io); + try self.parseSafetensors(io, tmp, file); + } + } + + pub fn parseFullIndexJson( + self: *TensorRegistry, + io: std.Io, + gpa: std.mem.Allocator, + repo: std.Io.Dir, + index_json_file: std.Io.File, + ) !void { + var tmp_state: std.heap.ArenaAllocator = .init(gpa); + defer tmp_state.deinit(); + const tmp = tmp_state.allocator(); + + const self_arena = self.arena.allocator(); + + var repo_uri_buf = try tmp.alloc(u8, 1024); + const repo_name = repo_uri_buf[0..try repo.realPath(io, repo_uri_buf)]; + + // Note: this is relatively inefficient, we read the full file, copy most of it's content, + // then reduce to TensorRegistry format. + // We could update TensorRegistry as we stream the json. + // Zig std json API are a bit awkward around streaming though. + const file_reader_buf = try tmp.alloc(u8, 64 * 1024); + var file_reader = index_json_file.reader(io, file_reader_buf); + var json_reader: std.json.Reader = .init(tmp, &file_reader.interface); + const index = try std.json.parseFromTokenSourceLeaky( + std.json.Value, + tmp, + &json_reader, + .{ .allocate = .alloc_if_needed }, + ); + + var realpaths: std.StringArrayHashMapUnmanaged([]const u8) = .empty; + try realpaths.ensureUnusedCapacity(tmp, 64); + const wm = index.object.get("weight_map") orelse return error.InvalidJson; + + try self.tensors.ensureUnusedCapacity(self_arena, wm.object.count()); + for (wm.object.keys(), wm.object.values()) |key, value| { + const tensor_name = try self_arena.dupe(u8, key); + const object = value.object; + const filename = value.object.get("filename").?.string; + const filename_entry = try realpaths.getOrPut(tmp, filename); + const file_uri = if (filename_entry.found_existing) + filename_entry.value_ptr.* + else uri: { + const uri = try std.fs.path.join(self_arena, &.{ repo_name, filename }); + filename_entry.value_ptr.* = uri; + break :uri uri; + }; + + const tensor_info: Tensor = try .parse(object, file_uri, tensor_name, 0); + self.tensors.putAssumeCapacity(tensor_name, tensor_info); + } + + if (index.object.get("metadata")) |metadata| { + // value belong to tmp arena, duplicate it into self arena. + var owned_metadata: std.Io.Writer.Allocating = .init(self_arena); + try std.json.fmt(metadata, .{ .whitespace = .minified }).format(&owned_metadata.writer); + self.metadata = owned_metadata.written(); + } + } + + pub fn saveAsFullIndexJson( + self: *const TensorRegistry, + io: std.Io, + gpa: std.mem.Allocator, + repo: std.Io.Dir, + out: *std.Io.Writer, + ) !void { + var repo_uri_buf = try gpa.alloc(u8, std.fs.max_path_bytes); + defer gpa.free(repo_uri_buf); + const repo_name = repo_uri_buf[0..try repo.realPath(io, repo_uri_buf)]; + var total_size: u64 = 0; + for (self.tensors.values()) |value| total_size += value.byteSize(); + + try out.print( + \\{{ + \\"metadata": {{"num_tensors":{d},"total_size":{d}}}, + \\"weight_map": {{ + \\ + , .{ self.tensors.count(), total_size }); + + const n = self.tensors.count(); + for (0.., self.tensors.keys(), self.tensors.values()) |i, key, value| { + std.debug.assert(std.mem.startsWith(u8, value.file_uri, repo_name)); + const local_name = value.file_uri[repo_name.len + 1 ..]; + const end_offset = value.offset + value.byteSize(); + const safetensors_dtype: []const u8 = dt: { + for (safetensors_dtype_to_zml_dtype.keys(), safetensors_dtype_to_zml_dtype.values()) |st_dt, zml_dt| { + if (zml_dt == value.shape._dtype) break :dt st_dt; + } + // Note: we could bitcast to u8, + // but reaching this panic implies the user didn't loaded this TensorRegistry + // from a safetensors file, and manually created the dict. + // In which case, they know what they are doing, + // and can easily insert the bitcast themselves. + std.debug.panic("Safetensors doesn't support data type {}", .{value.shape._dtype}); + }; + const comma = if (i + 1 == n) "" else ","; + try out.print( + \\ "{s}": {{"filename":"{s}","shape":{f},"data_offsets":[{d},{d}],"dtype":"{s}"}}{s} + \\ + , .{ key, local_name, std.json.fmt(value.shape.dims(), .{ .whitespace = .minified }), value.offset, end_offset, safetensors_dtype, comma }); + } + + try out.writeAll( + \\} + \\} + \\ + ); + try out.flush(); + } }; -pub fn fetchRegistry( - allocator: std.mem.Allocator, - io: std.Io, - repo: std.Io.Dir, - model: std.Io.File, - model_path: []const u8, -) !TensorRegistry { - const file_type = resolveFiletype(model_path); - - return switch (file_type) { - .index => blk: { - const file_reader_buf = try allocator.alloc(u8, 64 * 1024); - defer allocator.free(file_reader_buf); - - var index_reader = model.reader(io, file_reader_buf); - - var safetensors_index = try parseSafetensorsIndex( - allocator, - &index_reader.interface, - ); - defer safetensors_index.deinit(); - - var registry: TensorRegistry = try .initWithMetadata(allocator, safetensors_index.metadata); - - try parseSafetensorsIndexFiles( - allocator, - io, - ®istry, - safetensors_index, - repo, - ); - - break :blk registry; - }, - .safetensors => blk: { - var registry: TensorRegistry = .init(allocator); - - try parseSafetensors(allocator, io, ®istry, model); - - break :blk registry; - }, - else => return error.InvalidPath, - }; -} - pub const TensorReader = struct { tensor: Tensor, file: std.Io.File, @@ -238,6 +507,30 @@ pub const Tensor = struct { return TensorReader.init(io, self, buffer, opts); } + pub fn parse(object: std.json.ObjectMap, file_uri: []const u8, name: []const u8, base_offset: u64) !Tensor { + const shape_field = object.get("shape").?.array; + if (shape_field.items.len > Shape.MAX_RANK) { + log.warn("Can't load tensor {s}, too many dims: {}", .{ name, shape_field.items.len }); + return error.TooManyDimensions; + } + + const offset_field = object.get("data_offsets").?; + const start: u64 = @intCast(offset_field.array.items[0].integer); + const dtype = try stringToDtype(object.get("dtype").?.string); + + var dims: Shape.DimsArray = .empty; + for (shape_field.items) |d| { + dims.appendAssumeCapacity(d.integer); + } + + return .{ + .file_uri = file_uri, + .name = name, + .shape = .init(dims.slice(), dtype), + .offset = base_offset + start, + }; + } + pub fn format(self: Tensor, writer: *std.Io.Writer) !void { try writer.print("Tensor(name={s} shape={f} size={d}, offset={d}, file_uri={s})", .{ self.name, @@ -249,621 +542,100 @@ pub const Tensor = struct { } }; -pub const TensorRegistry = struct { - arena: std.heap.ArenaAllocator, +pub const FileType = enum { + full_index, + index, + safetensors, + unknown, - tensors: Tensors, - metadata: Metadatas, - - mutex: std.Io.Mutex = .init, - - pub fn init(allocator: std.mem.Allocator) TensorRegistry { - return .{ - .arena = std.heap.ArenaAllocator.init(allocator), - .tensors = .{}, - .metadata = .{}, - }; - } - - pub fn initWithMetadata( - allocator: std.mem.Allocator, - metadata: Metadatas, - ) !TensorRegistry { - var self: TensorRegistry = .{ - .arena = std.heap.ArenaAllocator.init(allocator), - .tensors = .{}, - .metadata = .empty, - }; - - try self.mergeMetadata(metadata); - - return self; - } - - pub fn fromRepo( - allocator: std.mem.Allocator, - io: std.Io, - repo: std.Io.Dir, - ) !TensorRegistry { - const entrypoint_file, const entrypoint_path = try resolveModelEntrypoint(io, repo); - defer entrypoint_file.close(io); - return try fetchRegistry(allocator, io, repo, entrypoint_file, entrypoint_path); - } - - pub fn fromPath( - allocator: std.mem.Allocator, - io: std.Io, - path: []const u8, - ) !TensorRegistry { - var repo = try resolveModelRepo(io, path); - defer repo.close(io); - - const entrypoint_file, const entrypoint_path = if (std.mem.endsWith(u8, path, ".safetensors.index.json") or std.mem.endsWith(u8, path, ".safetensors")) - .{ try repo.openFile(io, path, .{ .mode = .read_only }), path } - else - try resolveModelEntrypoint(io, repo); - defer entrypoint_file.close(io); - - return try fetchRegistry(allocator, io, repo, entrypoint_file, entrypoint_path); - } - - pub fn deinit(self: *TensorRegistry) void { - const allocator = self.arena.allocator(); - self.tensors.deinit(allocator); - self.metadata.deinit(allocator); - self.arena.deinit(); - } - - pub fn mergeMetadata( - self: *TensorRegistry, - other: Metadatas, - ) !void { - // self.mutex.lock(); - // defer self.mutex.unlock(); - - const allocator = self.arena.allocator(); - - var it = other.iterator(); - while (it.next()) |entry| { - const key = try allocator.dupe(u8, entry.key_ptr.*); - const value = try entry.value_ptr.*.clone(allocator); - - const gop = try self.metadata.getOrPut(allocator, key); - if (gop.found_existing) { - gop.value_ptr.*.deinit(allocator); - gop.value_ptr.* = value; - log.debug("Overwrote existing metadata key={s} with value={f}", .{ key, value }); - } else { - gop.value_ptr.* = value; - log.debug("Added new metadata key={s} with value={f}", .{ key, value }); - } + pub fn resolve(path: []const u8) FileType { + if (std.mem.endsWith(u8, path, ".safetensors.index.json")) { + return .index; + } else if (std.mem.endsWith(u8, path, ".safetensors.full_index.json")) { + return .full_index; + } else if (std.mem.endsWith(u8, path, ".safetensors")) { + return .safetensors; + } else { + return .unknown; } } - - pub fn registerTensor( - self: *TensorRegistry, - tensor: Tensor, - ) !void { - const allocator = self.arena.allocator(); - - // self.mutex.lock(); - // defer self.mutex.unlock(); - - var tensor_copy = tensor; - - tensor_copy.name = try allocator.dupe(u8, tensor.name); - tensor_copy.file_uri = try allocator.dupe(u8, tensor.file_uri); - - try self.tensors.put(allocator, tensor_copy.name, tensor_copy); - } - - pub fn reader( - self: *TensorRegistry, - io: std.Io, - tensor_name: []const u8, - buffer: []u8, - ) !TensorReader { - const tensor = self.tensors.get(tensor_name) orelse { - log.err("Tensor {s} not found in registry", .{tensor_name}); - return error.TensorNotFound; - }; - - return try .init(io, tensor, buffer, .{}); - } - - pub fn iterator(self: *TensorRegistry) Tensors.Iterator { - return self.tensors.iterator(); - } - - pub fn totalBytes(self: *TensorRegistry) u64 { - var total: u64 = 0; - - var it = self.tensors.iterator(); - while (it.next()) |entry| { - total += entry.value_ptr.byteSize(); - } - - return total; - } }; -pub fn parseSafetensors( - allocator: std.mem.Allocator, - io: std.Io, - registry: *TensorRegistry, - file: std.Io.File, -) !void { - var arena = std.heap.ArenaAllocator.init(allocator); - defer arena.deinit(); - - const header_buf = try allocator.alloc(u8, 64 * 1024); - defer allocator.free(header_buf); - - var reader = file.reader(io, header_buf); - - var file_uri_buf: [1024]u8 = undefined; - const file_uri_len = try file.realPath(io, &file_uri_buf); - const file_uri = file_uri_buf[0..file_uri_len]; - - const json_header_length: u64 = try reader.interface.takeInt(u64, .little); - const data_start_offset = BYTES_HEADER + json_header_length; - const json_data = try arena.allocator().alloc(u8, @intCast(json_header_length)); - - try reader.interface.readSliceAll(json_data); - - const headers: std.json.Value = try std.json.parseFromSliceLeaky( - std.json.Value, - arena.allocator(), - json_data, - .{}, - ); - - var it = headers.object.iterator(); - while (it.next()) |entry| { - const key = entry.key_ptr.*; - const value = entry.value_ptr.*; - - if (std.mem.eql(u8, key, "__metadata__")) { - // Parse metadata into a local arena to avoid concurrent allocations - // against registry.arena (which is not thread-safe). - var local_arena = std.heap.ArenaAllocator.init(allocator); - defer local_arena.deinit(); - - const metas = try parseMetadata(local_arena.allocator(), value); - try registry.mergeMetadata(metas); - - continue; - } - - const shape_field = value.object.get("shape").?.array; - - if (shape_field.items.len > Shape.MAX_RANK) { - log.warn("Can't load tensor {s}, too many dims: {}", .{ key, shape_field.items.len }); - return error.TooManyDimensions; - } - - const offset_field = value.object.get("data_offsets").?; - const start: u64 = @intCast(offset_field.array.items[0].integer); - const dtype = try stringToDtype(value.object.get("dtype").?.string); - - var dims: Dims = .empty; - for (shape_field.items) |d| { - dims.appendAssumeCapacity(d.integer); - } - - const tensor: Tensor = .{ - .file_uri = file_uri, - .name = key, - .shape = .init(dims.slice(), dtype), - .offset = data_start_offset + start, - }; - - try registry.registerTensor(tensor); - } -} - -const SafetensorsIndex = struct { - pub const Map = std.StringArrayHashMapUnmanaged(std.ArrayList([]const u8)); - - arena: std.heap.ArenaAllocator, - - map: Map, - metadata: Metadatas, - - pub fn init(allocator: std.mem.Allocator) SafetensorsIndex { - return .{ - .arena = std.heap.ArenaAllocator.init(allocator), - .map = .empty, - .metadata = .empty, - }; - } - - pub fn iterator(self: *SafetensorsIndex) Map.Iterator { - return self.map.iterator(); - } - - pub fn deinit(self: *SafetensorsIndex) void { - const allocator = self.arena.allocator(); - self.map.deinit(allocator); - self.metadata.deinit(allocator); - self.arena.deinit(); - } -}; - -fn parseSafetensorsIndex( - allocator: std.mem.Allocator, - reader: *std.Io.Reader, -) !SafetensorsIndex { - var safetensors_index: SafetensorsIndex = .init(allocator); - errdefer safetensors_index.deinit(); - - const arena_allocator = safetensors_index.arena.allocator(); - var json_reader: std.json.Reader = .init(arena_allocator, reader); - - const index = try std.json.parseFromTokenSourceLeaky( - std.json.Value, - arena_allocator, - &json_reader, - .{ .allocate = .alloc_if_needed }, - ); - - const weight_map = index.object.get("weight_map"); - - if (weight_map) |wm| { - var it = wm.object.iterator(); - - while (it.next()) |entry| { - const weight_name = entry.key_ptr.*; - const filename = entry.value_ptr.string; - - const map_entry = try safetensors_index.map.getOrPut(arena_allocator, filename); - - if (!map_entry.found_existing) { - map_entry.value_ptr.* = .empty; - } - - try map_entry.value_ptr.append(arena_allocator, try arena_allocator.dupe(u8, weight_name)); - } - } else { - log.warn("No weight_map attribute found in index", .{}); - } - - if (index.object.get("metadata")) |metadata_val| { - safetensors_index.metadata = try parseMetadata(arena_allocator, metadata_val); - } - - return safetensors_index; -} - -fn parseSafetensorsIndexFiles( - allocator: std.mem.Allocator, - io: std.Io, - registry: *TensorRegistry, - safetensors_index: SafetensorsIndex, - repo: std.Io.Dir, -) !void { - const filenames = safetensors_index.map.keys(); - for (filenames) |filename| { - const file = try repo.openFile(io, filename, .{}); - defer file.close(io); - try parseSafetensors(allocator, io, registry, file); - } -} - -pub fn resolveFiletype(path: []const u8) FileType { - if (std.mem.endsWith(u8, path, ".safetensors.index.json")) { - return .index; - } else if (std.mem.endsWith(u8, path, ".safetensors")) { - return .safetensors; - } else { - return .unknown; - } -} - pub const ModelPathResolutionError = error{ FileNotFound, InvalidPath, } || std.mem.Allocator.Error; -pub fn resolveModelRepo( - io: std.Io, - path: []const u8, -) ModelPathResolutionError!std.Io.Dir { - if (std.mem.endsWith(u8, path, ".safetensors.index.json") or - std.mem.endsWith(u8, path, ".safetensors")) - { - const dir_path = std.fs.path.dirname(path) orelse return ModelPathResolutionError.InvalidPath; +pub fn resolveModelRepo(io: std.Io, path: []const u8) ModelPathResolutionError!std.Io.Dir { + switch (FileType.resolve(path)) { + .full_index, .index, .safetensors => { + // The path is a file, let's open it's folder + const dir_path = std.fs.path.dirname(path) orelse return ModelPathResolutionError.InvalidPath; - const dir = std.Io.Dir.openDir(.cwd(), io, dir_path, .{}) catch |e| { - return switch (e) { - error.FileNotFound => ModelPathResolutionError.FileNotFound, - else => ModelPathResolutionError.InvalidPath, + const dir = std.Io.Dir.openDir(.cwd(), io, dir_path, .{}) catch |e| { + return switch (e) { + error.FileNotFound => ModelPathResolutionError.FileNotFound, + else => ModelPathResolutionError.InvalidPath, + }; }; - }; - return dir; + return dir; + }, + .unknown => { + // The path is probably a directory, let's try to open it as such + const dir = std.Io.Dir.openDir(.cwd(), io, path, .{}) catch |e| { + return switch (e) { + error.FileNotFound => ModelPathResolutionError.FileNotFound, + else => ModelPathResolutionError.InvalidPath, + }; + }; + return dir; + }, } - - const dir = std.Io.Dir.openDir(.cwd(), io, path, .{}) catch |e| { - return switch (e) { - error.FileNotFound => ModelPathResolutionError.FileNotFound, - else => ModelPathResolutionError.InvalidPath, - }; - }; - - return dir; } -pub fn resolveModelEntrypoint(io: std.Io, repo: std.Io.Dir) ModelPathResolutionError!struct { std.Io.File, []const u8 } { - { - const index_name = "model.safetensors.index.json"; - if (repo.openFile(io, index_name, .{ .mode = .read_only })) |index_file| { - return .{ index_file, "model.safetensors.index.json" }; - } else |_| {} - } - - { - const model_name = "model.safetensors"; - if (repo.openFile(io, model_name, .{ .mode = .read_only })) |model_file| { - return .{ model_file, "model.safetensors" }; - } else |_| {} - } - - return ModelPathResolutionError.FileNotFound; -} - -const Metadata = union(enum) { - null: void, - int: i64, - float: f64, - bool: bool, - string: []const u8, - - array_bool: []const bool, - array_int: []const i64, - array_float: []const f64, - array_string: []const []const u8, - - pub const ItemType = enum { - int, - float, - bool, - string, - - pub fn toZigType(comptime kind: ItemType) type { - return switch (kind) { - .int => i64, - .float => f64, - .bool => bool, - .string => []const u8, - }; - } +pub fn resolveModelEntrypoint(io: std.Io, repo: std.Io.Dir) ModelPathResolutionError!struct { FileType, std.Io.File } { + const default_entrypoints: []const struct { FileType, []const u8 } = &.{ + .{ .full_index, "model.safetensors.full_index.json" }, + .{ .full_index, "consolidated.safetensors.full_index.json" }, + .{ .index, "model.safetensors.index.json" }, + .{ .index, "consolidated.safetensors.index.json" }, + .{ .safetensors, "model.safetensors" }, + .{ .safetensors, "consolidated.safetensors" }, }; - pub fn wrap(x: anytype) Metadata { - return switch (@TypeOf(x)) { - inline u8, i8, u16, i16, u32, i32, u64, i64 => .{ .int = @intCast(x) }, - inline f16, f32, f64 => .{ .float = @floatCast(x) }, - bool => .{ .bool = x }, - []const u8 => .{ .string = x }, - else => @panic("Unsupported type for Value: " ++ @typeName(@TypeOf(x))), - }; + for (default_entrypoints) |entrypoint| { + const file_type, const file_name = entrypoint; + const file = repo.openFile(io, file_name, .{ .mode = .read_only }) catch continue; + + return .{ file_type, file }; } - pub fn copySlice(allocator: std.mem.Allocator, any_slice: anytype) !Metadata { - return switch (@TypeOf(any_slice[0])) { - inline u8, i8, u16, i16, u32, i32, u64, i64 => { - const res = try allocator.alloc(i64, any_slice.len); - for (res, any_slice) |*r, val| r.* = @intCast(val); - return .{ .array_int = res }; - }, - inline f16, f32, f64 => { - const res = try allocator.alloc(f64, any_slice.len); - for (res, any_slice) |*r, val| r.* = @floatCast(val); - return .{ .array_float = res }; - }, - bool => .{ .array_bool = try allocator.dupe(bool, any_slice) }, - []const u8 => .{ .array_string = try allocator.dupe([]const u8, @alignCast(any_slice)) }, - else => @panic("Unsupported type for Value: " ++ @typeName(@TypeOf(any_slice))), - }; - } + return error.FileNotFound; +} - pub fn format( - self: Metadata, - writer: *std.Io.Writer, - ) !void { - switch (self) { - .null => _ = try writer.write("null"), - .string => |s| try writer.print("{s}", .{s}), - .bool => |b| try writer.print("{}", .{b}), - .int => |i| try writer.print("{d}", .{i}), - .float => |f| try writer.print("{d}", .{f}), - .array_bool => |arr| { - try writer.writeByte('['); - for (arr, 0..) |v, i| { - if (i > 0) try writer.writeAll(", "); - try writer.print("{}", .{v}); - } - try writer.writeByte(']'); - }, - .array_int => |arr| { - try writer.writeByte('['); - for (arr, 0..) |v, i| { - if (i > 0) try writer.writeAll(", "); - try writer.print("{d}", .{v}); - } - try writer.writeByte(']'); - }, - .array_float => |arr| { - try writer.writeByte('['); - for (arr, 0..) |v, i| { - if (i > 0) try writer.writeAll(", "); - try writer.print("{d}", .{v}); - } - try writer.writeByte(']'); - }, - .array_string => |arr| { - try writer.writeByte('['); - for (arr, 0..) |v, i| { - if (i > 0) try writer.writeAll(", "); - try writer.print("\"{s}\"", .{v}); - } - try writer.writeByte(']'); - }, - } - } +const safetensors_dtype_to_zml_dtype = std.StaticStringMap(DataType).initComptime(.{ + .{ "F64", .f64 }, + .{ "F32", .f32 }, + .{ "F16", .f16 }, + .{ "BF16", .bf16 }, + .{ "F8_E8M0", .f8e8m0 }, + .{ "F8_E4M3", .f8e4m3fn }, + .{ "F4_E2M1", .f4e2m1 }, + .{ "I64", .i64 }, + .{ "I32", .i32 }, + .{ "I16", .i16 }, + .{ "I8", .i8 }, + .{ "U64", .u64 }, + .{ "U32", .u32 }, + .{ "U16", .u16 }, + .{ "U8", .u8 }, + .{ "BOOL", .bool }, +}); - pub fn deinit(self: Metadata, allocator: std.mem.Allocator) void { - switch (self) { - .string => |s| allocator.free(s), - .array_bool => |s| allocator.free(s), - .array_int => |s| allocator.free(s), - .array_float => |s| allocator.free(s), - .array_string => |s| { - for (s) |str| allocator.free(str); - allocator.free(s); - }, - else => {}, - } - } - - pub fn clone(self: Metadata, allocator: std.mem.Allocator) !Metadata { - return switch (self) { - .null => .null, - .int => |v| .{ .int = v }, - .float => |v| .{ .float = v }, - .bool => |v| .{ .bool = v }, - .string => |s| .{ .string = try allocator.dupe(u8, s) }, - .array_bool => |s| .{ .array_bool = try allocator.dupe(bool, s) }, - .array_int => |s| .{ .array_int = try allocator.dupe(i64, s) }, - .array_float => |s| .{ .array_float = try allocator.dupe(f64, s) }, - .array_string => |s| blk: { - const new_slice = try allocator.alloc([]const u8, s.len); - errdefer allocator.free(new_slice); - for (s, 0..) |str, i| { - new_slice[i] = try allocator.dupe(u8, str); - } - break :blk .{ .array_string = new_slice }; - }, - }; - } -}; - -fn stringToDtype(safetensor_type: []const u8) !DataType { - const map = std.StaticStringMap(DataType).initComptime(.{ - .{ "F64", .f64 }, - .{ "F32", .f32 }, - .{ "F16", .f16 }, - .{ "BF16", .bf16 }, - .{ "F8_E8M0", .f8e8m0 }, - .{ "F8_E4M3", .f8e4m3fn }, - .{ "F4_E2M1", .f4e2m1 }, - .{ "I64", .i64 }, - .{ "I32", .i32 }, - .{ "I16", .i16 }, - .{ "I8", .i8 }, - .{ "U64", .u64 }, - .{ "U32", .u32 }, - .{ "U16", .u16 }, - .{ "U8", .u8 }, - .{ "BOOL", .bool }, - }); - - return map.get(safetensor_type) orelse { - log.err("Unsupported safetensor data type: {s}", .{safetensor_type}); +fn stringToDtype(safetensor_dtype: []const u8) !DataType { + return safetensors_dtype_to_zml_dtype.get(safetensor_dtype) orelse { + log.err("Unsupported safetensor data type: {s}", .{safetensor_dtype}); return error.UnsupportedDataType; }; } - -fn parseMetadata(allocator: std.mem.Allocator, val: std.json.Value) !Metadatas { - var metadatas: Metadatas = .{}; - var prefix_buf: [1024]u8 = undefined; - var prefix = StringBuilder.initBuffer(&prefix_buf); - - try populateMetadata(allocator, &prefix, val, &metadatas); - - return metadatas; -} - -fn populateMetadata(allocator: std.mem.Allocator, prefix: *StringBuilder, val: std.json.Value, metadatas: *Metadatas) !void { - const key = prefix.items; - return switch (val) { - .null => try metadatas.put(allocator, try allocator.dupe(u8, key), .null), - .bool => |v| try metadatas.put(allocator, try allocator.dupe(u8, key), .{ .bool = v }), - .integer => |v| try metadatas.put(allocator, try allocator.dupe(u8, key), .{ .int = v }), - .float => |v| try metadatas.put(allocator, try allocator.dupe(u8, key), .{ .float = v }), - .number_string, .string => |v| try metadatas.put(allocator, try allocator.dupe(u8, key), .{ .string = try allocator.dupe(u8, v) }), - .array => |v| { - if (v.items.len == 0) return; - if (validSlice(v)) |item_type| { - const data: Metadata = switch (item_type) { - .bool => blk: { - const values = try allocator.alloc(bool, v.items.len); - for (v.items, 0..) |item, i| values[i] = item.bool; - break :blk .{ .array_bool = values }; - }, - .integer => blk: { - const values = try allocator.alloc(i64, v.items.len); - for (v.items, 0..) |item, i| values[i] = item.integer; - break :blk .{ .array_int = values }; - }, - .float => blk: { - const values = try allocator.alloc(f64, v.items.len); - for (v.items, 0..) |item, i| values[i] = item.float; - break :blk .{ .array_float = values }; - }, - inline .string, .number_string => |tag| blk: { - const values = try allocator.alloc([]const u8, v.items.len); - for (v.items, 0..) |item, i| { - values[i] = try allocator.dupe(u8, @field(item, @tagName(tag))); - } - break :blk .{ .array_string = values }; - }, - .null, .array, .object => unreachable, - }; - try metadatas.put(allocator, try allocator.dupe(u8, key), data); - } else { - for (v.items, 0..) |item, i| { - const old_len = prefix.items.len; - if (prefix.items.len > 0) { - prefix.appendAssumeCapacity('.'); - } - prefix.items.len += std.fmt.printInt(prefix.unusedCapacitySlice(), i, 10, .lower, .{}); - try populateMetadata(allocator, prefix, item, metadatas); - prefix.items.len = old_len; - } - } - }, - .object => |v| { - var obj_iter = v.iterator(); - while (obj_iter.next()) |entry| { - const old_len = prefix.items.len; - if (prefix.items.len > 0) { - prefix.appendAssumeCapacity('.'); - } - prefix.appendSliceAssumeCapacity(entry.key_ptr.*); - try populateMetadata(allocator, prefix, entry.value_ptr.*, metadatas); - prefix.items.len = old_len; - } - }, - }; -} - -/// We can only create a Zig slice out of json array, if all values -/// in the array have the same type. -fn validSlice(v: std.json.Array) ?std.meta.Tag(std.json.Value) { - if (v.items.len == 0) return null; - - const item_type: std.meta.Tag(std.json.Value) = v.items[0]; - switch (item_type) { - .null, .array, .object => return null, - else => {}, - } - - for (v.items[1..]) |item| { - if (item != item_type) - return null; - } - - return item_type; -} diff --git a/zml/shape.zig b/zml/shape.zig index b0863367..386b594e 100644 --- a/zml/shape.zig +++ b/zml/shape.zig @@ -117,14 +117,31 @@ pub const Shape = struct { stdx.debug.compileError("Wrong type, got {}. Expected .{{.a, .b}}", .{T}); } - pub fn parseTags(v: anytype) TagsArray { + pub fn parseTags(v: anytype) []const Shape.Tag { const T = @TypeOf(v); - stdx.debug.assertComptime(stdx.meta.isTupleOf(T, @EnumLiteral()), "Wrong type, got {}. Expected .{{ .a, .b }}", .{T}); - var tags_: TagsArray = .empty; - inline for (v) |field| { - tags_.appendAssumeCapacity(toTag(field)); + if (T == []const Shape.Tag) return v; + comptime validateTags(T); + return comptime frozen: { + var tagz: [v.len]Shape.Tag = undefined; + for (v, 0..) |field, i| tagz[i] = @tagName(field).ptr; + // Force the creation of a comptime constant, so that we can return a pointer to it. + const frozen_tags = tagz; + break :frozen &frozen_tags; + }; + } + + fn validateTags(T: type) void { + const message = "Expected an enum-literal tuple such as .{ .b, .d }, got " ++ @typeName(T); + switch (@typeInfo(T)) { + .@"struct" => |info| { + if (!info.is_tuple) @compileError(message); + for (info.field_types) |Field| { + if (Field != @EnumLiteral()) @compileError(message); + } + if (info.field_types.len > Shape.MAX_RANK) @compileError("Too many shape tags (maximum 8)"); + }, + else => @compileError(message), } - return tags_; } /// Create a shape from a struct literal, eg: @@ -808,6 +825,18 @@ pub const Shape = struct { } } + pub fn assertTags(self: Shape, tagz: anytype) void { + const T = @TypeOf(tagz); + if (T == Shape) { + return self.assertTags(tagz.tags()); + } else if (T != []const Shape.Tag) { + return self.assertTags(comptime parseTags(tagz)); + } + + stdx.debug.assert(tagz.len == self.rank(), "Not enough tags for shape {f}, got {}", .{ self, stdx.fmt.stringsZ(tagz) }); + stdx.debug.assert(std.mem.eql(Shape.Tag, tagz, self.tags()), "Expected tags {f}, got {f}", .{ stdx.fmt.stringsZ(tagz), self }); + } + pub fn withDtype(self: Shape, dt: DataType) Shape { var res = self; res._dtype = dt; @@ -827,6 +856,9 @@ pub const Shape = struct { .grid = undefined, }; + /// Returns shape metadata with the requested sharding on `mesh`. + /// Omitted dimensions in a mapping are replicated, including during compilation. + /// Use `Tensor.withPartitioning` to emit a constraint on a graph value. pub fn withPartitioning(self: Shape, mesh: *const Sharding.Mesh, partitioning: anytype) Shape { if (@TypeOf(partitioning) != Sharding.Partitioning) { return self.withPartitioning(mesh, Sharding.Partitioning.parse(self.tags(), mesh, .replicated, partitioning)); diff --git a/zml/tensor.zig b/zml/tensor.zig index 8c148867..ee084835 100644 --- a/zml/tensor.zig +++ b/zml/tensor.zig @@ -75,14 +75,11 @@ pub const Tensor = struct { return self._shape.byteSize(); } - /// Internal use - /// - /// Creates a tensor from a Shape and an mlir.Value. - /// Note: sharding information is discarded, the output tensor is considered open for replication. - /// Use _resultPropagateSharding when the shape sharding should be propagated. + /// Internal: creates a graph tensor with dimensions, dtype and tags from `sh`, + /// and open sharding for the partitioner to infer from graph constraints. pub fn _result(sh: Shape, val: *const mlir.Value) Tensor { var res: Tensor = .{ ._shape = sh, ._value = val, .id = nextTensorId() }; - res._shape._sharding.partition = .open(sh.rank()); + res._shape._sharding = .open; if (builtin.mode == .debug) { // Check that the MLIR value actually have the same shape. @@ -95,16 +92,16 @@ pub const Tensor = struct { return res; } + /// Internal: creates a constructor result and honors `sh`'s partitioning when + /// it specifies a mesh. pub fn _resultPropagateSharding(sh: Shape, val: *const mlir.Value) Tensor { - var res = _result(sh, val); - res._shape._sharding = sh._sharding; - return res; + const result = _result(sh, val); + const mesh = sh._sharding.mesh orelse return result; + return result.withPartitioning(mesh, sh._sharding.partition); } - /// Creates a Tensor from a mlir.Value - /// - /// The shape is derived from the type of the mlir.Value. - /// mesh is `unknown` + /// Wraps an MLIR tensor value, deriving dimensions and dtype from its type. + /// Dimension tags are unknown and sharding is open for partitioner inference. pub fn fromMlirValue(val: *const mlir.Value) Tensor { const ctx = Compiler.current(); const ranked_tensor = val.type_().isA(mlir.RankedTensorType).?; @@ -116,7 +113,7 @@ pub const Tensor = struct { ._dtype = ctx.dtype(ranked_tensor.elementType()), ._dims = .empty, ._tags = .{ .buffer = @splat(Shape.TagUnknown), .len = n }, - ._sharding = .replicated, + ._sharding = .open, }; for (0..n) |i| { sh._dims.appendAssumeCapacity(ranked_tensor.dimension(i)); @@ -166,12 +163,22 @@ pub const Tensor = struct { return res; } - /// Specify the mesh of the input tensor. - /// * mesh: *const zml.Mesh, but during compilation the `.meshe_name` syntax can used to get - /// a known mesh from the compilation options. - /// * partition spec: a struct where the field names match the axis of the given mesh + /// Declares this tensor's sharding on `mesh_`. + /// Pass a `*const Mesh`, or a mesh name such as `.tp` during compilation. + /// `partitioning` accepts a `Sharding.Partitioning` or a mapping from dimension + /// tags to logical mesh axes, `.replicated`, or `.open`. /// - /// eg `x.withPartitioning(tp, .{ .h = .model }))` or `x.withPartitioning(.tp, .{ .h = .model }))` + /// During compilation, emits a graph constraint. For example, + /// `x.withPartitioning(.tp, .{ .h = .model })` splits dimension `h` over `model` + /// and leaves omitted dimensions open for inference. Use `.replicated` to + /// keep a dimension whole on each shard. + /// + /// Outside compilation, updates the tensor's shape to describe an executable + /// input's sharding. Omitted dimensions are replicated and no graph constraint is + /// emitted. An explicit `Sharding.Partitioning` uses its supplied specs in both contexts. + /// + /// Executable inputs and explicitly annotated outputs must have no open + /// dimensions. An output with no declared sharding defaults to replication. pub fn withPartitioning(self: Tensor, mesh_: anytype, partitioning: anytype) Tensor { if (@TypeOf(mesh_) == @EnumLiteral()) { const compiler = Compiler.currentOrNull() orelse @panic("Out side of compilation, withPartitioning expects an explicit *const zml.Mesh object as input"); @@ -179,7 +186,8 @@ pub const Tensor = struct { } if (@TypeOf(partitioning) != Sharding.Partitioning) { - const parsed = Sharding.Partitioning.parse(self._shape.tags(), mesh_, .open, partitioning); + const default: Sharding.PartitionSpec = if (Compiler.currentOrNull() == null) .replicated else .open; + const parsed = Sharding.Partitioning.parse(self._shape.tags(), mesh_, default, partitioning); return self.withPartitioning(mesh_, parsed); } @@ -222,11 +230,14 @@ pub const Tensor = struct { }, }; - return _resultPropagateSharding(partitioned_shape, op_result); + var res = _result(partitioned_shape, op_result); + res._shape = partitioned_shape; + return res; } - /// Force the input tensor to be fully replicated. - /// This can be detrimental for performance. + /// Emits a graph constraint requiring a complete copy on every device. + /// Requires active compilation. For a replicated input or buffer description, + /// use the default sharding from `Tensor.init` or `Shape.init`. pub fn replicate(self: Tensor) Tensor { const compiler = Compiler.current(); const partitioning: Sharding.Partitioning = .repeat(.replicated, self.rank()); @@ -267,11 +278,17 @@ pub const Tensor = struct { try std.testing.expectEqual(Sharding.PartitionSpec.replicated, x2_replicated.shape().partition(.h)); } - /// Copy the given tensor to the specified memory. + /// Copy the given tensor to the specified memory, preserving its sharding. pub fn toMemory(self: Tensor, kind: Memory.Kind) Tensor { const ctx = Compiler.current(); switch (ctx.platform.target) { - .cpu, .neuron, .metal => return self, + .cpu, .neuron, .metal => { + // A single memory needs no copy, but the result must still be a distinct + // tensor: `reuseBuffer` donations are keyed by tensor id. + var res = _result(self._shape, self.value()); + res._shape._sharding = self._shape._sharding; + return res; + }, .cuda, .rocm, .tpu, .oneapi, .furiosa => {}, } @@ -298,7 +315,8 @@ pub const Tensor = struct { ctx.location, ).appendTo(currentBlock()); - const res = _result(self._shape, op.result(0)); + var res = _result(self._shape, op.result(0)); + res._shape._sharding = self._shape._sharding; ctx.currentScope().id_to_memory.putNoClobber(ctx.currentScope().arena.allocator(), res.id, kind) catch @panic("OOM"); return res; } @@ -845,7 +863,7 @@ pub const Tensor = struct { fn update(self: Rng, new_state: *const mlir.Value) Rng { return .{ - ._state = _resultPropagateSharding(self._state._shape, new_state).reuseBuffer(self._state), + ._state = _result(self._state._shape, new_state).reuseBuffer(self._state), .algorithm = self.algorithm, }; } @@ -886,7 +904,8 @@ pub const Tensor = struct { floats = floats.mul(scalar(opts.max - opts.min, dt)).addConstant(opts.min); // Convert back to integer if needed. - return .{ rng, _resultPropagateSharding(shape_, floats.convert(shape_.dtype()).value()) }; + const result = floats.convert(shape_.dtype()); + return .{ rng, _resultPropagateSharding(shape_, result.value()) }; } test uniform { @@ -1515,7 +1534,7 @@ pub const Tensor = struct { pub fn dotAxes(lhs: Tensor, rhs: Tensor, tags: anytype) DotAxes { const T = @TypeOf(tags); if (comptime T == Shape.Tag or T == @EnumLiteral()) { - return lhs.dotAxes(rhs, .{tags}); + return lhs.dotAxes(rhs, @as([]const Shape.Tag, &.{Shape.toTag(tags)})); } const lhs_contracting_axes, _ = lhs.shape().parseAxes(tags); @@ -2705,7 +2724,7 @@ pub const Tensor = struct { mlirx.Type.rankedTensor(mlirCtx(), res_shape), currentLoc(), ).appendTo(currentBlock()); - return _result(res_shape, op.result(0)); + return _resultPropagateSharding(res_shape, op.result(0)); } pub const LinspaceArgs = struct { @@ -2780,7 +2799,8 @@ pub const Tensor = struct { } pub fn zeroes(sh: Shape) Tensor { - return Tensor.constant(sh.dtype().zero()).broad(sh); + const result = Tensor.constant(sh.dtype().zero()).broad(sh); + return _resultPropagateSharding(sh, result.value()); } pub fn uninitialized(sh: Shape) Tensor { @@ -2795,7 +2815,7 @@ pub const Tensor = struct { .appendTo(currentBlock()); const tensor = stablehlo.unpin(ctx.mlir_ctx, buffer.result(0), ctx.location) .appendTo(currentBlock()); - return _result(sh, tensor.result(0)); + return _resultPropagateSharding(sh, tensor.result(0)); } /// Embeds a buffer with concrete values into an Mlir program. @@ -2803,7 +2823,7 @@ pub const Tensor = struct { const elem_type = mlirx.Type.fromDType(mlirCtx(), sh.dtype()); //const elem_type = mlirx.denseElementAttrType(val.dtype()) orelse std.debug.panic("constantTensor expects a dtype that can be serialized to MLIR, like f32 or i32, got {f}", .{val.shape()}); const constant_op = stablehlo.constant(mlirCtx(), sh.dims(), elem_type, bytes_, currentLoc()).appendTo(currentBlock()); - return _result(sh, constant_op.result(0)); + return _resultPropagateSharding(sh, constant_op.result(0)); } /// Returns a Tensor containing the result of the outer product between the input Tensors. @@ -2969,7 +2989,7 @@ pub const Tensor = struct { var res_shape = self._shape; const ones: [constants.MAX_RANK]i64 = @splat(1); res_shape._dims.insertSlice(ax, ones[0..tags_.len]) catch unreachable; - res_shape._tags.insertSlice(ax, tags_.constSlice()) catch unreachable; + res_shape._tags.insertSlice(ax, tags_) catch unreachable; return self.reshape(res_shape); } @@ -4806,7 +4826,7 @@ pub const Tensor = struct { fn body(body_ctx: @This(), _: void) void { ops.customCall("zml$print", body_ctx.input, {}, .{ .name = body_ctx.name }, .{ .has_side_effect = true }); } - }).body, .{ .input = input, .name = full_name }, {}); + }).body, .{ .input = input, .name = full_name }, {}, ctx.platform.replicated_mesh, .{ .in_specs = .{ .input = .replicated } }); }, .oneapi, .neuron, .furiosa => {}, } diff --git a/zml/zml.zig b/zml/zml.zig index 23d12e75..dafc3a0f 100644 --- a/zml/zml.zig +++ b/zml/zml.zig @@ -26,7 +26,7 @@ pub const Exe = exe.Exe; pub const FnExe = exe.FnExe; pub const floats = @import("floats.zig"); pub const io = @import("io.zig"); -pub const kernel = @import("kernel.zig"); +pub const kernel = @import("kernel/kernel.zig"); pub const logo = @import("logo.zig"); pub const mem = @import("mem.zig"); pub const Bufferized = mem.Bufferized; @@ -81,3 +81,9 @@ pub const KiB = 1024; pub const MiB = 1024 * KiB; pub const GiB = 1024 * MiB; pub const TiB = 1024 * GiB; + +comptime { + // Forces zig to compile `@zml//bazel` module even when users don't depend on it even indirectly. + // Otherwise exported functions declared in `bazel` aren't compiled. + std.testing.refAllDecls(bazel); +}