diff --git a/MODULE.bazel.lock b/MODULE.bazel.lock index b23d9054..be285fcc 100644 --- a/MODULE.bazel.lock +++ b/MODULE.bazel.lock @@ -155,6 +155,7 @@ "https://bazel-registry.zml.ai/modules/stardoc/0.6.2/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/stardoc/0.7.0/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/stardoc/0.7.1/MODULE.bazel": "not found", + "https://bazel-registry.zml.ai/modules/toolchains_protoc/0.3.7/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/upb/0.0.0-20220923-a547704/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/upb/0.0.0-20230516-61a97ef/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/with_cfg.bzl/0.8.0/MODULE.bazel": "not found", @@ -354,6 +355,8 @@ "https://bcr.bazel.build/modules/stardoc/0.7.0/MODULE.bazel": "05e3d6d30c099b6770e97da986c53bd31844d7f13d41412480ea265ac9e8079c", "https://bcr.bazel.build/modules/stardoc/0.7.1/MODULE.bazel": "3548faea4ee5dda5580f9af150e79d0f6aea934fc60c1cc50f4efdd9420759e7", "https://bcr.bazel.build/modules/stardoc/0.7.1/source.json": "b6500ffcd7b48cd72c29bb67bcac781e12701cc0d6d55d266a652583cfcdab01", + "https://bcr.bazel.build/modules/toolchains_protoc/0.3.7/MODULE.bazel": "d2758e5af5838970175274c4fcb67edc8d2f4f8f348885da90c42fafc221236d", + "https://bcr.bazel.build/modules/toolchains_protoc/0.3.7/source.json": "07e0eab9be4aded2a8074504a80c55c5a48e811a9824061642af639baaabf985", "https://bcr.bazel.build/modules/upb/0.0.0-20220923-a547704/MODULE.bazel": "7298990c00040a0e2f121f6c32544bab27d4452f80d9ce51349b1a28f3005c43", "https://bcr.bazel.build/modules/upb/0.0.0-20230516-61a97ef/MODULE.bazel": "c0df5e35ad55e264160417fd0875932ee3c9dda63d9fccace35ac62f45e1b6f9", "https://bcr.bazel.build/modules/with_cfg.bzl/0.8.0/MODULE.bazel": "d836870543677303f1cb6f9649e955c772e9f15311c2926a2468c9a10eb214d9", diff --git a/examples/MODULE.bazel.lock b/examples/MODULE.bazel.lock index 0ba88287..e97baa93 100644 --- a/examples/MODULE.bazel.lock +++ b/examples/MODULE.bazel.lock @@ -158,6 +158,7 @@ "https://bazel-registry.zml.ai/modules/stardoc/0.6.2/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/stardoc/0.7.0/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/stardoc/0.7.1/MODULE.bazel": "not found", + "https://bazel-registry.zml.ai/modules/toolchains_protoc/0.3.7/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/upb/0.0.0-20220923-a547704/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/upb/0.0.0-20230516-61a97ef/MODULE.bazel": "not found", "https://bazel-registry.zml.ai/modules/with_cfg.bzl/0.8.0/MODULE.bazel": "not found", @@ -361,6 +362,8 @@ "https://bcr.bazel.build/modules/stardoc/0.7.0/MODULE.bazel": "05e3d6d30c099b6770e97da986c53bd31844d7f13d41412480ea265ac9e8079c", "https://bcr.bazel.build/modules/stardoc/0.7.1/MODULE.bazel": "3548faea4ee5dda5580f9af150e79d0f6aea934fc60c1cc50f4efdd9420759e7", "https://bcr.bazel.build/modules/stardoc/0.7.1/source.json": "b6500ffcd7b48cd72c29bb67bcac781e12701cc0d6d55d266a652583cfcdab01", + "https://bcr.bazel.build/modules/toolchains_protoc/0.3.7/MODULE.bazel": "d2758e5af5838970175274c4fcb67edc8d2f4f8f348885da90c42fafc221236d", + "https://bcr.bazel.build/modules/toolchains_protoc/0.3.7/source.json": "07e0eab9be4aded2a8074504a80c55c5a48e811a9824061642af639baaabf985", "https://bcr.bazel.build/modules/upb/0.0.0-20220923-a547704/MODULE.bazel": "7298990c00040a0e2f121f6c32544bab27d4452f80d9ce51349b1a28f3005c43", "https://bcr.bazel.build/modules/upb/0.0.0-20230516-61a97ef/MODULE.bazel": "c0df5e35ad55e264160417fd0875932ee3c9dda63d9fccace35ac62f45e1b6f9", "https://bcr.bazel.build/modules/with_cfg.bzl/0.8.0/MODULE.bazel": "d836870543677303f1cb6f9649e955c772e9f15311c2926a2468c9a10eb214d9", diff --git a/examples/llama/llama.zig b/examples/llama/llama.zig index a162b432..2eaf8a8b 100644 --- a/examples/llama/llama.zig +++ b/examples/llama/llama.zig @@ -87,16 +87,15 @@ pub const LlamaLM = struct { rng: Tensor.Rng, ) struct { Tensor, KvCache, Tensor.Rng } { stdx.debug.assert(tokens_.dtype() == .u32 and tokens_.rank() >= 1 and token_index.dtype() == .u32 and token_index.rank() <= 1, "Can't run Llama ! Expected >=1d tokens and 0d token_index, got: {} and {}", .{ tokens_, token_index }); - var tokens = tokens_.withPartialTags(.{.s}); + const tokens = tokens_.withPartialTags(.{.s}); const out, const updated_kv_cache = zml.call(self.model, .forward, .{ tokens, token_index, kv_cache }); - tokens, const new_rng = self.sampleTokens(self.lm_head, tokens, out, rng, self.gen_opts); - return .{ tokens, updated_kv_cache, new_rng }; + const new_tokens, const new_rng = self.sampleTokens(self.lm_head, out, rng, self.gen_opts); + return .{ new_tokens.convert(tokens.dtype()).reuseBuffer(tokens), updated_kv_cache, new_rng }; } pub fn sampleTokens( self: LlamaLM, lm_head_: ?zml.nn.Linear, - tokens_: Tensor, out_: Tensor, rng: Tensor.Rng, opts: zml.nn.SamplingStrategy, @@ -115,7 +114,7 @@ pub const LlamaLM = struct { logits = logits.rename(.{ .d = .voc }); const next_tokens, const new_rng = zml.nn.sampleTokens(logits, opts, rng); - return .{ next_tokens.convert(tokens_.dtype()).reuseBuffer(tokens_), new_rng }; + return .{ next_tokens, new_rng }; } pub fn increment(_: u8, token_index: Tensor) Tensor { diff --git a/examples/llama/main.zig b/examples/llama/main.zig index 87148371..9a17c404 100644 --- a/examples/llama/main.zig +++ b/examples/llama/main.zig @@ -22,11 +22,17 @@ pub const std_options = .{ .logFn = asynk.logFn(std.log.defaultLog), }; -pub fn tokenizePromptLlama3(allocator: std.mem.Allocator, tokenizer: zml.tokenizer.Tokenizer, config: LlamaLM.Config, prompt: []const u8) ![]u32 { +pub fn tokenizePrompt(allocator: std.mem.Allocator, tokenizer: zml.tokenizer.Tokenizer, config: LlamaLM.Config, prompt: []const u8, skip_llama3_encoding: bool) ![]u32 { var tokens = std.ArrayList(u32).init(allocator); var encoder = try tokenizer.encoder(); defer encoder.deinit(); + if (skip_llama3_encoding) { + // Copy to the arraylist so the ownership is the same in both branches. + try tokens.appendSlice(try encoder.encode(prompt)); + return tokens.toOwnedSlice(); + } + const start_header_id = tokenizer.tokenToId("<|start_header_id|>") orelse return error.NoSuchToken; const end_header_id = tokenizer.tokenToId("<|end_header_id|>") orelse return error.NoSuchToken; const eot_id = tokenizer.tokenToId("<|eot_id|>") orelse return error.NoSuchToken; @@ -59,84 +65,61 @@ pub fn generateText( prompt: []const u8, skip_llama3_encoding: bool, ) ![]const u8 { - var tokenizer_encoder = try tokenizer.encoder(); - defer tokenizer_encoder.deinit(); + const prompt_tok: []const u32 = try tokenizePrompt(allocator, tokenizer, config, prompt, skip_llama3_encoding); + defer allocator.free(prompt_tok); + var tokenizer_decoder = try tokenizer.decoder(); defer tokenizer_decoder.deinit(); - const prompt_tok: []const u32 = if (skip_llama3_encoding) try tokenizer_encoder.encode(prompt) else try tokenizePromptLlama3(allocator, tokenizer, config, prompt); - defer allocator.free(prompt_tok); - - const dims = llama_.model.shape(); - const max_seq_len = dims.s; - - // Prefill - // initialize a 0..max_seq_len buffer with the tokenized prompt - const prefill_buffer = try allocator.alloc(u32, @intCast(max_seq_len)); - @memset(prefill_buffer, 0); - for (0..prompt_tok.len) |i| { - prefill_buffer[i] = @intCast(prompt_tok[i]); - } - defer allocator.free(prefill_buffer); - const platform = mod_generate.platform(); + const max_seq_len = llama_.model.shape().s; - // prepare device buffers for the prefill tokens and the index - var prefill_tokens = try zml.Buffer.fromSlice(platform, .{max_seq_len}, prefill_buffer); - defer prefill_tokens.deinit(); - var prefill_token_index = try zml.Buffer.constant(platform, zml.Shape.init(.{}, .u32), 0); - - defer prefill_token_index.deinit(); - - // init RNG and prefill + // init RNG and buffers var rng = try zml.Tensor.Rng.init(platform, seed); - prefill_tokens, var kv_cache, rng = mod_prefill.call(.{ prefill_tokens, prefill_token_index, kv_cache_, rng }); - defer kv_cache.k.deinit(); - defer kv_cache.v.deinit(); - defer kv_cache.layer_index.deinit(); + var generated_token_buffer = [_]u32{undefined}; - // Prepare for token-by-token generation - var first_token_hostbuffer = [_]u32{prompt_tok[prompt_tok.len - 1]}; // start with the prompt's last token - var current_token = try zml.Buffer.fromSlice(platform, .{1}, &first_token_hostbuffer); + var kv_cache = prefill: { + // prepare device buffers for the prefill tokens and their positions + const prefill_buffer = try allocator.alloc(u32, max_seq_len); + @memcpy(prefill_buffer[0..prompt_tok.len], prompt_tok); + + var prefill_tokens = try zml.Buffer.fromSlice(platform, .{max_seq_len}, prefill_buffer); + defer prefill_tokens.deinit(); + var prefill_token_pos = try zml.Buffer.constant(platform, zml.Shape.init(.{}, .u32), 0); + defer prefill_token_pos.deinit(); + + const prefilled_tokens, const kv_cache, rng = mod_prefill.call(.{ prefill_tokens, prefill_token_pos, kv_cache_, rng }); + _ = try prefilled_tokens.toHost(std.mem.sliceAsBytes(prefill_buffer)); + generated_token_buffer[0] = prefill_buffer[prompt_tok.len - 1]; + break :prefill kv_cache; + }; + defer zml.aio.unloadBuffers(&kv_cache); + + // Prepare for token-by-token generation, + // start with the token generated based on the full prompt. + var current_token = try zml.Buffer.fromSlice(platform, .{1}, &generated_token_buffer); defer current_token.deinit(); - // Here we will copy the generated token from device - var generated_token_buffer = [_]u32{0}; - // Here we collect the generated text var output = std.ArrayList(u8).init(allocator); defer output.deinit(); - const tracer_buffer = try allocator.alloc(u8, @intCast(max_seq_len)); - defer allocator.free(tracer_buffer); - const tracer = zml.tools.Tracer.init("ai.zml.models.llama"); const output_tokens_len = max_seq_len - prompt_tok.len - 1; const start = std.time.microTimestamp(); - var num_tokens_generated: usize = 0; + // One token has alreadyh been generated by the prefill. + var num_tokens_generated: usize = 1; - generation: for (0..output_tokens_len) |i| { - const frame_id = tracer.frameStart(try std.fmt.bufPrintZ(tracer_buffer, "Generate token {}/{}", .{ i + 1, output_tokens_len })); - - // current token index needs to go into a zml.Buffer - const token_index_buffer = &[_]u32{@intCast(prompt_tok.len + i)}; - const token_index = try zml.Buffer.fromSlice(platform, .{}, token_index_buffer); - - defer token_index.deinit(); - - // call to generate the next token - current_token, kv_cache, rng = mod_generate.call(.{ current_token, token_index, kv_cache, rng }); - - tracer.frameEnd(frame_id, try std.fmt.bufPrintZ(tracer_buffer, "Generated token {}/{}", .{ i + 1, output_tokens_len })); - - // extract the generated token from the buffer - _ = try current_token.toHost(std.mem.sliceAsBytes(&generated_token_buffer)); + generation: for (0..output_tokens_len + 1) |i| { + // collect and print generated sequence + num_tokens_generated += 1; const generated_token = generated_token_buffer[0]; - // de-tokenize generated token into a string - const chunk = try tokenizer_decoder.next(@intCast(generated_token)) orelse unreachable; - num_tokens_generated = i; + const chunk = try tokenizer_decoder.next(generated_token) orelse unreachable; + try output.appendSlice(chunk); + std.debug.print("{s}", .{chunk}); // check for eos + if (i == output_tokens_len) break :generation; switch (config.eos_token_id.value) { .int => |eos| if (generated_token == @as(u32, @intCast(eos))) break :generation, .ints => |eos_list| { @@ -146,9 +129,16 @@ pub fn generateText( }, } - // collect and print generated sequence - try output.appendSlice(chunk); - std.debug.print("{s}", .{chunk}); + // current token pos needs to go into a zml.Buffer + const token_pos_buffer = &[_]u32{@intCast(prompt_tok.len + i)}; + const token_pos = try zml.Buffer.fromSlice(platform, .{}, token_pos_buffer); + defer token_pos.deinit(); + + // call to generate the next token + current_token, kv_cache, rng = mod_generate.call(.{ current_token, token_pos, kv_cache, rng }); + + // extract the generated token from the buffer + _ = try current_token.toHost(std.mem.sliceAsBytes(&generated_token_buffer)); } const end = std.time.microTimestamp(); const duration = stdx.math.divFloat(f64, end - start, std.time.us_per_s); diff --git a/examples/simple_layer/main.zig b/examples/simple_layer/main.zig index 6f61c0b0..75053c97 100644 --- a/examples/simple_layer/main.zig +++ b/examples/simple_layer/main.zig @@ -2,19 +2,17 @@ const std = @import("std"); const zml = @import("zml"); const asynk = @import("async"); -/// Model definition -const Layer = struct { - bias: ?zml.Tensor = null, - weight: zml.Tensor, +const Tensor = zml.Tensor; - pub fn forward(self: Layer, x: zml.Tensor) zml.Tensor { - var y = self.weight.mul(x); - if (self.bias) |bias| { - y = y.add(bias); - } - return y; - } -}; +pub fn conditionalForward(cache: Tensor, indices: Tensor, input: Tensor) Tensor { + const pages = cache.gatherValues(.{.page}, indices.appendAxes(.{.coord}), .{}); + var result = pages.dot(input, .{.hd}); + const zeros = zml.Tensor.constant(result.shape(), result.dtype().zero()); + const cutoff = zml.Tensor.constant(indices.shape(), indices.dtype().constant(cache.dim(.page))); + const mask = zml.Tensor.cmp(indices, .LT, cutoff).broad(result.shape()); + result = mask.select(result, zeros); + return result; +} pub fn main() !void { try asynk.AsyncThread.main(std.heap.c_allocator, asyncMain); @@ -30,69 +28,93 @@ pub fn asyncMain() !void { var arena_state = std.heap.ArenaAllocator.init(allocator); defer arena_state.deinit(); const arena = arena_state.allocator(); + _ = arena; // autofix var context = try zml.Context.init(); defer context.deinit(); - const platform = context.autoPlatform(.{}); - context.printAvailablePlatforms(platform); - - // Our weights and bias to use - var weights = [4]f16{ 2.0, 2.0, 2.0, 2.0 }; - var bias = [4]f16{ 1.0, 2.0, 3.0, 4.0 }; - const input_shape = zml.Shape.init(.{4}, .f16); - - // We manually produce a BufferStore. You would not normally do that. - // A BufferStore is usually created by loading model data from a file. - var buffers: zml.aio.BufferStore.Buffers = .{}; - try buffers.put(arena, "weight", zml.HostBuffer.fromArray(&weights)); - try buffers.put(arena, "bias", zml.HostBuffer.fromArray(&bias)); - - // the actual BufferStore - const buffer_store: zml.aio.BufferStore = .{ - .arena = arena_state, - .buffers = buffers, + const compilation_options = zml.CompilationOptions{ + .xla_dump_to = "/tmp/zml/simple_layer", }; - // A clone of our model, consisting of shapes. We only need shapes for compiling. - // We use the BufferStore to infer the shapes. - var model_shapes = try zml.aio.populateModel(Layer, allocator, buffer_store); - model_shapes.weight = model_shapes.weight.withSharding(.{-1}); - model_shapes.bias = model_shapes.bias.?.withSharding(.{-1}); + const platform = context.autoPlatform(.{}).withCompilationOptions(compilation_options); + context.printAvailablePlatforms(platform); - // Start compiling. This uses the inferred shapes from the BufferStore. - // The shape of the input tensor, we have to pass in manually. - var compilation = try asynk.asyncc(zml.compileModel, .{ allocator, Layer.forward, model_shapes, .{input_shape}, platform }); + var profiler = platform.getProfiler(null); + defer profiler.deinit(); - // Produce a bufferized weights struct from the fake BufferStore. - // This is like the inferred shapes, but with actual values. - // We will need to send those to the computation device later. - var model_weights = try zml.aio.loadModelBuffers(Layer, model_shapes, buffer_store, arena, platform); - defer zml.aio.unloadBuffers(&model_weights); // for good practice + profiler.start(); + defer { + profiler.stop(); + profiler.dumpAsJsonTo(allocator, std.fs.cwd(), "trace.json") catch unreachable; + } - // Wait for compilation to finish - const compiled = try compilation.awaitt(); + const dtype = .bf16; + const batch_size = 1; + const page_count = 8192; + const max_page_count = 2048; + const chunk_size = 16; + const num_heads = 8; + const head_dim = 8192; - // pass the model weights to the compiled module to create an executable module - var executable = compiled.prepare(model_weights); - defer executable.deinit(); + const cache_shape = zml.Shape.init(.{ .page = page_count, .k_chunk = chunk_size, .h = num_heads, .hd = head_dim }, dtype); + const indices_shape = zml.Shape.init(.{ .b = batch_size, .max_page = max_page_count }, .u32); + const input_shape = zml.Shape.init(.{ .b = batch_size, .q = 1, .h = num_heads, .hd = head_dim }, dtype); + const mod = try zml.compileFn(allocator, conditionalForward, .{ cache_shape, indices_shape, input_shape }, platform); + defer mod.deinit(); - // prepare an input buffer - // Here, we use zml.HostBuffer.fromSlice to show how you would create a HostBuffer - // with a specific shape from an array. - // For situations where e.g. you have an [4]f16 array but need a .{2, 2} input shape. - var input = [4]f16{ 5.0, 5.0, 5.0, 5.0 }; - var input_buffer = try zml.Buffer.from(platform, zml.HostBuffer.fromSlice(input_shape, &input)); - defer input_buffer.deinit(); + var generator = std.Random.DefaultPrng.init(0); + const random = generator.random(); + + var cache = try createRandomBuffer(allocator, platform, cache_shape, random); + defer cache.deinit(); + var input = try createRandomBuffer(allocator, platform, input_shape, random); + defer input.deinit(); + + const indices_data = try allocator.alloc(u32, max_page_count); + defer allocator.free(indices_data); + + //@memset(indices_data, std.math.maxInt(u32)); + @memset(indices_data, 1); + //indices_data[0] = 1; + //@memcpy(indices_data[0..16], &[_]u32{ 3, 5, 17, 546, 43, 68, 22, 23, 24, 25, 26, 27, 89, 90, 91, 92 }); + var indices = try zml.Buffer.fromSlice(platform, indices_shape, indices_data); + _ = try indices.awaitt(); // call our executable module - var result: zml.Buffer = executable.call(.{input_buffer}); + var result: zml.Buffer = mod.call(.{ cache, indices, input }); defer result.deinit(); +} - // fetch the result to CPU memory - const cpu_result = try result.toHostAlloc(arena); - std.debug.print( - "\nThe result of {d} * {d} + {d} = {d}\n", - .{ &weights, &input, &bias, cpu_result.items(f16) }, - ); +fn createRandomBuffer(allocator: std.mem.Allocator, platform: zml.Platform, shape: zml.Shape, random: std.Random) !zml.Buffer { + const data = try allocator.alloc(u8, shape.byteSize()); + defer allocator.free(data); + + switch (shape.dtype()) { + inline else => |v| { + const ZigType = v.toZigType(); + switch (comptime v.class()) { + .bool => unreachable, + .integer => { + for (std.mem.bytesAsSlice(ZigType, data)) |*e| e.* = random.int(ZigType); + }, + .float => { + const value = random.float(f64); + for (std.mem.bytesAsSlice(ZigType, data)) |*e| e.* = if (ZigType == f64) + value + else if (ZigType == f32) + @floatCast(value) + else if (ZigType == f16) + @floatCast(value) + else + @bitCast(random.int(std.meta.Int(.unsigned, @bitSizeOf(ZigType)))); + }, + .complex => unreachable, + } + }, + } + + var temp = try zml.Buffer.fromBytes(platform, shape, data); + _ = try temp.awaitt(); + return temp; } diff --git a/pjrt/pjrt.zig b/pjrt/pjrt.zig index 9cfa5fb5..8fd1b75e 100644 --- a/pjrt/pjrt.zig +++ b/pjrt/pjrt.zig @@ -6,7 +6,9 @@ const c = @import("c"); const log = std.log.scoped(.pjrt); -pub const Profiler = @import("profiler.zig").Profiler; +pub const profiler = @import("profiler.zig"); + +const Profiler = profiler.Profiler; test { std.testing.refAllDecls(@This()); @@ -732,7 +734,7 @@ pub const Buffer = opaque { return ret.on_device_size_in_bytes; } - pub fn copyToDevice(self: *const Buffer, api: *const Api, device: Device) ApiError!Buffer { + pub fn copyToDevice(self: *const Buffer, api: *const Api, device: Device) ApiError!*Buffer { const ret = try api.call(.PJRT_Buffer_CopyToDevice, .{ .buffer = self.inner(), .dst_device = device.inner, diff --git a/pjrt/profiler.zig b/pjrt/profiler.zig index 1b0d1a69..9a937382 100644 --- a/pjrt/profiler.zig +++ b/pjrt/profiler.zig @@ -3,7 +3,7 @@ const c = @import("c"); const tsl_proto = @import("//tsl:profiler_options_proto"); const log = std.log.scoped(.@"pjrt/profiler"); -const TraceContainer = @import("convert/trace_container.zig").TraceContainer; +pub const TraceContainer = @import("convert/trace_container.zig").TraceContainer; /// Pjrt Profiler extension pub const Profiler = struct { @@ -60,14 +60,14 @@ pub const Profiler = struct { } pub fn start(self: *Profiler) void { - self.transition("start", .ready, .started); + //self.transition("start", .ready, .started); if (self.api == null) return; var args: c.PLUGIN_Profiler_Start_Args = .{ .profiler = self.inner }; self.check(self.api.?.start.?(&args)) catch unreachable; } pub fn stop(self: *Profiler) void { - self.transition("stop", .started, .stopped); + //self.transition("stop", .started, .stopped); if (self.api == null) return; var args: c.PLUGIN_Profiler_Stop_Args = .{ .profiler = self.inner }; @@ -75,7 +75,7 @@ pub const Profiler = struct { } pub fn collectData(self: *Profiler, allocator: std.mem.Allocator) !ProfilingData { - self.transition("collect_data", .stopped, .done); + //self.transition("collect_data", .stopped, .done); if (self.api == null) return .{ .external = &.{} }; var args: c.PLUGIN_Profiler_CollectData_Args = .{ diff --git a/zml/aio.zig b/zml/aio.zig index 175ffb32..ea6f1f6a 100644 --- a/zml/aio.zig +++ b/zml/aio.zig @@ -564,8 +564,7 @@ pub fn loadModelBuffersWithPrefix( return res; } -/// Takes a bufferized version of a `model`, ie a mirror struct of the `model`, and deinit all the -/// Buffer found. +/// Deinit all buffers in the given struct pub fn unloadBuffers(model: anytype) void { zml.meta.visit((struct { fn cb(_: void, buffer: *zml.Buffer) void { @@ -574,6 +573,20 @@ pub fn unloadBuffers(model: anytype) void { }).cb, {}, model); } +/// Await all buffers in the given struct +pub fn awaitAll(buffers: anytype) !void { + var err: ?anyerror = null; + zml.meta.visit((struct { + fn cb(err_ptr: *?anyerror, buffer: *zml.Buffer) void { + _ = buffer.awaitt() catch |e| { + err_ptr.* = e; + }; + } + }).cb, &err, buffers); + + return err orelse {}; +} + fn visitStructAndLoadBuffer(allocator: std.mem.Allocator, prefix_builder: *PrefixBuilder, buffer_store: BufferStore, obj: anytype, platform: zml.Platform) !void { const err_msg = "visitStructAndLoadBuffer must be called with a pointer to type. Received "; const type_info, const T = switch (@typeInfo(@TypeOf(obj))) { @@ -593,6 +606,7 @@ fn visitStructAndLoadBuffer(allocator: std.mem.Allocator, prefix_builder: *Prefi stdx.debug.assert(host_buffer.shape().eql(obj._shape), "loadModelBuffers expects to find the same shapes in the model and in the buffer store, got {} and {} for tensor {s}", .{ obj._shape, host_buffer, prefix }); buf_with_metadata._shape = obj._shape; obj.* = try zml.Buffer.from(platform, buf_with_metadata); + _ = try obj.awaitt(); } else { return error.BufferNotFound; }; diff --git a/zml/buffer.zig b/zml/buffer.zig index a9fba67a..b5d7f01f 100644 --- a/zml/buffer.zig +++ b/zml/buffer.zig @@ -27,12 +27,30 @@ const log = std.log.scoped(.zml); /// * loading weights from disk directly to the `device zml.aio.loadBuffers` /// * can be created by calling `HostBuffer.toDevice(platform)`. pub const Buffer = struct { + pub const Shard = struct { + // TODO: we should use SoA instead of AoS because of the "ready" bool. + // the Buffer struct is already huge. + buffer: *pjrt.Buffer, + ready_event: ?*pjrt.Event = null, + ready: bool = false, + + pub fn awaitt(self: *Shard, api: *const pjrt.Api) pjrt.ApiError!void { + if (self.ready) { + return; + } + if (self.ready_event orelse self.buffer.getReadyEvent(api)) |ev| { + try ev.awaitt(api); + } + self.ready = true; + } + }; + _shape: Shape, _api: *const pjrt.Api, _shards: Shards, pub const MAX_NUM_SHARDS: u8 = Platform.MAX_NUM_DEVICES; - pub const Shards = std.BoundedArray(*pjrt.Buffer, MAX_NUM_SHARDS); + pub const Shards = std.BoundedArray(Shard, MAX_NUM_SHARDS); /// Copies the content of the given buffer from host memory to the accelerator memory. pub fn from(platform: Platform, host_buffer: HostBuffer) !Buffer { @@ -56,49 +74,35 @@ pub const Buffer = struct { const buffer_type = bufferTypeFromDtype(host_buffer.shape().dtype()); const byte_strides = host_buffer.strides() orelse host_buffer.shape().computeStrides().constSlice(); - var frames: std.BoundedArray(asynk.Frame(pjrt.Client.bufferFromHostBuffer), MAX_NUM_SHARDS) = .{}; const devices = platform.getDevices(); for (0..n_partitions) |i| { - // If no sharding if found, the given buffer is replicated on all devices. const buf = if (sharding_ax) |ax| buf: { const start: i64 = @as(i64, @intCast(i)) * chunk_size; break :buf host_buffer.slice1d(ax, .{ .start = start, .end = start + chunk_size }); } else host_buffer; - const frame = try asynk.asyncc(pjrt.Client.bufferFromHostBuffer, .{ - platform.pjrt_client, - platform.pjrt_api, - .{ - .data = buf.data, - .buffer_type = buffer_type, - .dims = buf.shape().dims(), - .byte_strides = byte_strides, - .device = devices[i], - .host_buffer_semantics = .ImmutableOnlyDuringCall, - }, + const pjrt_buffer, const event = try platform.pjrt_client.bufferFromHostBuffer(platform.pjrt_api, .{ + .data = buf.data, + .buffer_type = buffer_type, + .dims = buf.shape().dims(), + .byte_strides = byte_strides, + .device = devices[i], + .host_buffer_semantics = .ImmutableUntilTransferCompletes, + }); + res._shards.appendAssumeCapacity(.{ + .buffer = pjrt_buffer, + .ready_event = event, }); - - frames.appendAssumeCapacity(frame); } - for (frames.slice()) |*frame| { - const pjrt_buffer = try frame.awaitt(); - res._shards.appendAssumeCapacity(pjrt_buffer); - } return res; } - /// Wraps pre-exisiting `pjrt.Buffer` shards into one `zml.Buffer`. - pub fn fromPjrtBuffers(platform: Platform, shape_: Shape, pjrt_buffers: []const *pjrt.Buffer) Buffer { - stdx.debug.assert(pjrt_buffers.len <= MAX_NUM_SHARDS, "ZML doesn't support having more than {} shards. Received {} shards for one buffer.", .{ MAX_NUM_SHARDS, pjrt_buffers.len }); - stdx.debug.assert(pjrt_buffers.len > 0, "fromPjrtBuffers expects at least one buffer, got 0.", .{}); - var shards: Shards = .{}; - shards.appendSliceAssumeCapacity(pjrt_buffers); - return .{ - ._api = platform.pjrt_api, - ._shape = shape_, - ._shards = shards, - }; + pub fn awaitt(self: *Buffer) pjrt.ApiError!*Buffer { + for (self._shards.slice()) |*shard| { + try shard.awaitt(self._api); + } + return self; } /// Copies the given Zig slice to the accelerator memory and @@ -112,7 +116,31 @@ pub const Buffer = struct { /// return a Buffer using the array shape. pub fn fromArray(platform: Platform, arr: anytype) !Buffer { const host_buffer = HostBuffer.fromArray(&arr); - return try from(platform, host_buffer); + var device_buffer = try from(platform, host_buffer); + // Await cause we need to finish the copy while the array is still on the stack. + // Alternatively we could accept a pointer to the array but that will break all fromArray callers + _ = try device_buffer.awaitt(); + return device_buffer; + } + + test fromArray { + const zml = @import("zml.zig"); + const platform = zml.testing.env(); + + const values_h = [2][8]i32{ + .{ 0, -1, 2, -3, 4, -5, 6, -7 }, + .{ 8, -9, 10, -11, 12, -13, 14, -15 }, + }; + var values_d = try fromArray(platform, values_h); + defer values_d.deinit(); + _ = try values_d.awaitt(); + + try std.testing.expectEqual(-15, (try values_d.getValue([2][8]i32))[1][7]); + } + + /// Creates a buffer from the given shape and bytes. + pub fn fromBytes(platform: Platform, shape_: Shape, s: []const u8) !Buffer { + return from(platform, HostBuffer.fromBytes(shape_, s)); } /// Creates a Buffer with a single element. @@ -168,7 +196,9 @@ pub const Buffer = struct { else => unreachable, } const host_buffer: HostBuffer = .{ ._shape = shape_, ._strides = strides, .data = &bytes }; - return try from(platform, host_buffer); + var buf = try from(platform, host_buffer); + _ = try buf.awaitt(); + return buf; } test constant { @@ -193,6 +223,7 @@ pub const Buffer = struct { /// Creates a Buffer from a pointer into device memory. /// This allows to interface with other libraries producing buffers. + /// The returned buffer will be considered as ready to read. pub fn asViewOfDeviceBuffer(platform: Platform, shape_: Shape, stream: ?*const anyopaque, device_data: *anyopaque) !Buffer { const minor_to_major: [Shape.MAX_RANK]i64 = comptime blk: { var res: [Shape.MAX_RANK]i64 = undefined; @@ -220,7 +251,12 @@ pub const Buffer = struct { }); var shards: Shards = .{}; - shards.appendAssumeCapacity(pjrt_buffer); + shards.appendAssumeCapacity(.{ + .buffer = pjrt_buffer, + // We don't have event here so we assume it's ready + .ready_event = null, + .ready = true, + }); return .{ ._api = platform.pjrt_api, ._shape = shape_, @@ -233,9 +269,9 @@ pub const Buffer = struct { stdx.debug.assert(self._shape.byteSize() == @sizeOf(T), "Buffer {} has {d} bytes of data, can't load it to a {s} with {d} bytes", .{ self, self._shape.byteSize(), @typeName(T), @sizeOf(T) }); var res: T = undefined; stdx.debug.internalAssert(!self.hasShardedAxis(), "TODO: support sharded Buffer -> Host transfer", .{}); - const maybe_event = try self._shards.get(0).toHostBuffer(self._api, std.mem.asBytes(&res)); + const maybe_event = try self._shards.get(0).buffer.toHostBuffer(self._api, std.mem.asBytes(&res)); if (maybe_event) |event| { - try event.await_(self._api); + try event.awaitt(self._api); } return res; } @@ -245,22 +281,23 @@ pub const Buffer = struct { /// The returned `HostBuffer` doesn't own the memory. pub fn toHost(self: Buffer, output: []u8) !HostBuffer { stdx.debug.internalAssert(!self.hasShardedAxis(), "TODO: support sharded Buffer -> Host transfer", .{}); - const maybe_event = try self._shards.get(0).toHostBuffer(self._api, output); - if (maybe_event) |event| { - try event.await_(self._api); - } - return HostBuffer.fromBytes(self.shape(), output); + const maybe_event = try self._shards.get(0).buffer.toHostBuffer(self._api, output); + var hb = HostBuffer.fromBytes(self.shape(), output); + hb._event = maybe_event; + hb._api = self._api; + hb._ready = if (maybe_event) |ev| ev.isReady(self._api) else true; + return hb; } /// Copies the content of the Buffer to the host. /// The returned `HostBuffer` does own the memory. pub fn toHostAlloc(self: Buffer, allocator: std.mem.Allocator) !HostBuffer { - const output = try HostBuffer.empty(allocator, self.shape()); + var output = try HostBuffer.empty(allocator, self.shape()); stdx.debug.internalAssert(!self.hasShardedAxis(), "TODO: support sharded Buffer -> Host transfer", .{}); - const maybe_event = try self._shards.get(0).toHostBuffer(self._api, @constCast(output.data)); - if (maybe_event) |event| { - try event.await_(self._api); - } + const maybe_event = try self._shards.get(0).buffer.toHostBuffer(self._api, @constCast(output.data)); + output._event = maybe_event; + output._api = self._api; + output._ready = if (maybe_event) |ev| ev.isReady(self._api) else true; return output; } @@ -268,8 +305,8 @@ pub const Buffer = struct { /// Depending on the platform, the memory is typically not released to the OS /// but just marked as available in the memory pool. pub fn deinit(self: *const Buffer) void { - for (self._shards.constSlice()) |buffer| { - buffer.deinit(self._api); + for (self._shards.constSlice()) |shard| { + shard.buffer.deinit(self._api); } } @@ -386,3 +423,27 @@ test bufferTypeFromDtype { try std.testing.expectEqual(dt, bufferTypeFromDtype(dtypeFromBufferType(dt))); } } + +/// await all buffers in the given struct +pub fn awaitAll(buffers: anytype) pjrt.ApiError!void { + var err: ?pjrt.ApiError = null; + meta.visit((struct { + fn cb(err_ptr: *?pjrt.ApiError, buffer: *Buffer) void { + _ = buffer.awaitt() catch |e| { + err_ptr.* = e; + return; + }; + } + }).cb, &err, buffers); + + return err orelse {}; +} + +/// deinit all buffers in the given struct +pub fn deinitAll(buffers: anytype) void { + meta.visit((struct { + fn cb(_: void, buffer: *Buffer) void { + buffer.deinit(); + } + }).cb, {}, buffers); +} diff --git a/zml/exe.zig b/zml/exe.zig index b34d5e9b..db87ddf7 100644 --- a/zml/exe.zig +++ b/zml/exe.zig @@ -203,18 +203,13 @@ pub const BaseExe = struct { .arguments = self.input_per_device, .num_args = self.input_buffer_count, .results = self.output_per_device, + // Create a local inste .events = events[0..sharding.num_partitions], // this allows to tell a specific buffer shouldn't be donated, // even if it has been marked as "can be donated" during compilation. // TODO: expose it ? .non_donatable_input_indices = &.{}, }) catch unreachable; - - for (events[0..sharding.num_partitions]) |e| { - if (e) |ev| { - ev.await_(self.platform.pjrt_api) catch unreachable; - } - } } pub fn serialize(self: BaseExe, writer: anytype) !void { @@ -235,13 +230,27 @@ pub const BaseExe = struct { self.ready_buffer_count += n; } - pub fn getOutputBuffer(self: BaseExe, i: usize) Buffer { - var shards: Buffer.Shards = .{}; - for (self.output_per_device) |dev_out| { - shards.appendAssumeCapacity(dev_out[i]); - } + pub fn numOutputs(self: BaseExe) u32 { + return @intCast(self.result_shapes.len); + } - return Buffer.fromPjrtBuffers(self.platform, self.result_shapes[i], shards.constSlice()); + /// Takes pjrt.Buffer from each device and make it into a Buffer. + /// Set the event of the buffer to the event of the call. + pub fn getOutputBuffer(self: BaseExe, i: usize) Buffer { + std.debug.assert(i < self.numOutputs()); + var shards: Buffer.Shards = .{}; + for (0..self.num_devices) |dev| { + shards.appendAssumeCapacity(.{ + .buffer = self.output_per_device[dev][i], + .ready_event = null, + .ready = false, + }); + } + return .{ + ._shape = self.result_shapes[i], + ._api = self.platform.pjrt_api, + ._shards = shards, + }; } }; @@ -283,7 +292,7 @@ pub fn Exe(ArgsT: type, ReturnT: type) type { std.debug.assert(total_ready == self.inner.input_buffer_count); self.inner._unsafeCall(); var result: Bufferized(ReturnT) = undefined; - assignRawBuffers(&result, self.inner.platform, self.inner.output_per_device, self.inner.result_shapes); + assignRawBuffers(&result, self.inner); return result; } }; @@ -316,7 +325,7 @@ fn fillBuffers(v: anytype, buffers: []const [*]*pjrt.Buffer, start: u32) u32 { const model_sharding = ctx.buffers.len; stdx.debug.assert(buffer._shards.len == model_sharding, "Can't feed a {}-sharded tensor into a {}-sharded model", .{ buffer._shards.len, ctx.buffers.len }); for (buffer._shards.constSlice(), 0..) |shard, d| { - ctx.buffers[d][ctx.index] = shard; + ctx.buffers[d][ctx.index] = shard.buffer; } ctx.index += 1; } @@ -325,31 +334,20 @@ fn fillBuffers(v: anytype, buffers: []const [*]*pjrt.Buffer, start: u32) u32 { } /// Visit the given struct and override tensors by creating a new one using the provided PJRT buffers. -fn assignRawBuffers(v: anytype, platform: Platform, buffers: []const [*]*pjrt.Buffer, buffer_shapes: []Shape) void { +fn assignRawBuffers(v: anytype, exe: BaseExe) void { const LocalContext = struct { - index: u32, - platform: Platform, - buffers: []const [*]*pjrt.Buffer, - buffer_shapes: []Shape, - }; - var local_ctx: LocalContext = .{ - .index = 0, - .platform = platform, - .buffers = buffers, - .buffer_shapes = buffer_shapes, - }; - meta.visit((struct { - fn cb(ctx: *LocalContext, buffer: *Buffer) void { + index: u32 = 0, + exe: *const BaseExe, + + fn cb(ctx: *@This(), buffer: *Buffer) void { const i = ctx.index; ctx.index += 1; - if (i >= ctx.buffer_shapes.len) return; + if (i >= ctx.exe.result_shapes.len) return; - var shards: Buffer.Shards = .{}; - for (ctx.buffers) |buff| { - shards.appendAssumeCapacity(buff[i]); - } - buffer.* = Buffer.fromPjrtBuffers(ctx.platform, ctx.buffer_shapes[i], shards.constSlice()); + buffer.* = ctx.exe.getOutputBuffer(i); } - }).cb, &local_ctx, v); - stdx.debug.internalAssert(local_ctx.index == buffer_shapes.len, "Pjrt call returned {} tensors, but the return type {s}, contains {} Buffers. Note that modules need to have a comptime know number of returned tensors.", .{ buffers.len, @typeName(@TypeOf(v)), local_ctx.index }); + }; + var local_ctx: LocalContext = .{ .exe = &exe }; + meta.visit(LocalContext.cb, &local_ctx, v); + stdx.debug.internalAssert(local_ctx.index == exe.result_shapes.len, "Pjrt call returned {} tensors, but the return type {s}, contains {} Buffers. Note that modules need to have a comptime know number of returned tensors.", .{ exe.result_shapes.len, @typeName(@TypeOf(v)), local_ctx.index }); } diff --git a/zml/hostbuffer.zig b/zml/hostbuffer.zig index cd7e3a60..d101c55f 100644 --- a/zml/hostbuffer.zig +++ b/zml/hostbuffer.zig @@ -1,6 +1,8 @@ const std = @import("std"); const stdx = @import("stdx"); +const pjrt = @import("pjrtx.zig"); + const Buffer = @import("buffer.zig").Buffer; const Data = @import("dtype.zig").Data; const DataType = @import("dtype.zig").DataType; @@ -22,6 +24,9 @@ pub const HostBuffer = struct { managed: u5, unmanaged, } = .unmanaged, + _ready: bool = true, + _api: ?*const pjrt.Api = null, + _event: ?*pjrt.Event = null, /// Allocates a HostBuffer with the given shape. /// The memory is left undefined. @@ -98,6 +103,25 @@ pub const HostBuffer = struct { }; } + test fromArray { + const zml = @import("zml.zig"); + { + const arr: [2][4]i32 = undefined; + const buf = fromArray(&arr); + try zml.testing.expectEqualShapes(buf.shape(), Shape.init(.{ 2, 4 }, .i32)); + } + { + const arr: [2][4][8]u8 = undefined; + const buf = fromArray(&arr); + try zml.testing.expectEqualShapes(buf.shape(), Shape.init(.{ 2, 4, 8 }, .u8)); + } + { + const arr: [3][2][4]f32 = undefined; + const buf = fromArray(&arr); + try zml.testing.expectEqualShapes(buf.shape(), Shape.init(.{ 3, 2, 4 }, .f32)); + } + } + pub const ArangeArgs = struct { start: i64 = 0, end: i64, @@ -255,6 +279,21 @@ pub const HostBuffer = struct { }; } + pub fn awaitt(self: *HostBuffer) !*HostBuffer { + if (self._ready) { + return self; + } + if (self._event) |ev| { + if (self._api) |api| { + try ev.awaitt(api); + } else { + stdx.debug.panic("Got an event but no API. This should not happen.", .{}); + } + } + self._ready = true; + return self; + } + pub fn format( self: HostBuffer, comptime fmt: []const u8, diff --git a/zml/module.zig b/zml/module.zig index 0c1bc19f..b9d0c5e6 100644 --- a/zml/module.zig +++ b/zml/module.zig @@ -891,7 +891,7 @@ fn compileModuleToPjrtExecutable(arena: std.mem.Allocator, platform: Platform, m try options.env_option_overrides.ensureUnusedCapacity(arena, 16); if (xla_dump_to_ orelse platform.compilation_options.xla_dump_to) |xla_dump_to| { setFlag(&options, "xla_dump_to", xla_dump_to); - setFlag(&options, "xla_dump_hlo_as_dot", true); + setFlag(&options, "xla_dump_hlo_as_proto", true); if (platform.compilation_options.xla_dump_fusion_visualization) { setFlag(&options, "xla_dump_fusion_visualization", true); } @@ -905,8 +905,8 @@ fn compileModuleToPjrtExecutable(arena: std.mem.Allocator, platform: Platform, m // https://github.com/NVIDIA/JAX-Toolbox?tab=readme-ov-file#environment-variables setFlag(&options, "xla_gpu_enable_triton_gemm", false); setFlag(&options, "xla_gpu_enable_latency_hiding_scheduler", true); - setFlag(&options, "xla_gpu_enable_llvm_module_compilation_parallelism", true); - setFlag(&options, "xla_gpu_enable_libnvptxcompiler", true); + // setFlag(&options, "xla_gpu_enable_llvm_module_compilation_parallelism", true); + // setFlag(&options, "xla_gpu_enable_libnvptxcompiler", true); // setFlag(&options, "xla_gpu_enable_cudnn_fmha", true); // setFlag(&options, "xla_gpu_fused_attention_use_cudnn_rng", true); // setFlag(&options, "xla_gpu_enable_cudnn_layer_norm", true); diff --git a/zml/nn.zig b/zml/nn.zig index 6f756e99..7d10a3a5 100644 --- a/zml/nn.zig +++ b/zml/nn.zig @@ -934,7 +934,7 @@ pub const PartialSoftmax = struct { /// Returns intermediary results to allow aggregating later. pub fn partialSoftmax(self: Tensor, axis: anytype) PartialSoftmax { const a = self.axis(axis); - const max_val = self.max(a); + const max_val = self.max(a).maximum(Tensor.scalar(-1e16, self.dtype())); const out = self.sub(max_val.broad(self.shape())).exp(); return .{ .values = out, diff --git a/zml/pjrtx.zig b/zml/pjrtx.zig index 7408e074..fb38c24f 100644 --- a/zml/pjrtx.zig +++ b/zml/pjrtx.zig @@ -14,10 +14,9 @@ const Target = @import("platform.zig").Target; const log = std.log.scoped(.zml); -pub const Profiler = pjrt.Profiler; +pub const profiler = pjrt.profiler; pub const ApiError = pjrt.ApiError; pub const ErrorCode = pjrt.ErrorCode; -pub const Buffer = pjrt.Buffer; pub const BufferType = pjrt.BufferType; pub const Device = pjrt.Device; pub const DeviceDescription = pjrt.DeviceDescription; @@ -31,6 +30,8 @@ pub const SerializeResult = pjrt.SerializeResult; pub const Executable = pjrt.Executable; pub const ExecuteError = ApiError; +const Profiler = profiler.Profiler; + fn InnerMixin(comptime innerT: type) type { return struct { inline fn inner(self: anytype) if (@typeInfo(@TypeOf(self)).Pointer.is_const) *const innerT else *innerT { @@ -63,13 +64,9 @@ pub const Client = opaque { } pub const BufferFromHostBufferArgs = pjrt.Client.BufferFromHostBufferArgs; - pub fn bufferFromHostBuffer(self: *const Client, api: *const Api, args: BufferFromHostBufferArgs) !*Buffer { - const buffer, const event_ = try asynk.callBlocking(pjrt.Client.bufferFromHostBuffer, .{ self.inner(), api, args }); - if (event_) |event__| { - const event: *Event = @ptrCast(event__); - try event.await_(api); - } - return buffer; + pub fn bufferFromHostBuffer(self: *const Client, api: *const Api, args: BufferFromHostBufferArgs) ApiError!struct { *Buffer, ?*Event } { + const buffer, const event_ = try self.inner().bufferFromHostBuffer(api, args); + return .{ @ptrCast(buffer), @ptrCast(event_) }; } pub fn deserializeAndLoad(self: *const Client, api: *const Api, bytes: []const u8) ApiError!*LoadedExecutable { @@ -125,11 +122,67 @@ pub const Client = opaque { /// Returns the Profiler for this API. /// Not all platform have a profiling api, for those the profiler object will do nothing. /// Platforms with known profiler extensions: cuda, xpu - pub fn getProfiler(self: *const Client, api: *const Api, options: pjrt.Profiler.Options) pjrt.Profiler { + pub fn getProfiler(self: *const Client, api: *const Api, options: Profiler.Options) Profiler { return self.inner().getProfiler(api, options); } }; +pub const Buffer = opaque { + pub const inner = InnerMixin(pjrt.Buffer).inner; + + pub fn deinit(self: *Buffer, api: *const Api) void { + self.inner().deinit(api); + } + + pub fn getDevice(self: *const Buffer, api: *const Api) ApiError!*Device { + return try self.inner().getDevice(api); + } + + pub fn delete(self: *Buffer, api: *const Api) void { + self.inner().delete(api); + } + + pub fn isDeleted(self: *const Buffer, api: *const Api) bool { + return self.inner().isDeleted(api); + } + + pub fn isOnCpu(self: *const Buffer, api: *const Api) bool { + return self.inner().isOnCpu(api); + } + + pub fn toHostBuffer(self: *const Buffer, api: *const Api, dst: []u8) ApiError!?*Event { + return @ptrCast(try self.inner().toHostBuffer(api, dst)); + } + + pub fn getElementType(self: *const Buffer, api: *const Api) BufferType { + return self.inner().getElementType(api); + } + + pub fn getDimensions(self: *const Buffer, api: *const Api) []const i64 { + return self.inner().getDimensions(api); + } + + pub fn getUnpaddedDimensions(self: *const Buffer, api: *const Api) ApiError![]const i64 { + return try self.inner().getUnpaddedDimensions(api); + } + + pub fn getOnDeviceSizeInBytes(self: *const Buffer, api: *const Api) ApiError!usize { + return try self.inner().getOnDeviceSizeInBytes(api); + } + + pub fn copyToDevice(self: *const Buffer, api: *const Api, device: Device) ApiError!*Buffer { + return @ptrCast(self.inner().copyToDevice(api, device)); + } + + pub fn getReadyEvent(self: *const Buffer, api: *const Api) ?*Event { + return @ptrCast(self.inner().getReadyEvent(api)); + } + + pub fn getOpaqueDeviceMemoryDataPointer(self: *const Buffer, api: *const Api) ApiError!*anyopaque { + return try self.inner().getOpaqueDeviceMemoryDataPointer(api); + } +}; + pub const Event = opaque { pub const inner = InnerMixin(pjrt.Event).inner; @@ -145,7 +198,8 @@ pub const Event = opaque { return self.inner().getEventError(api); } - pub fn await_(self: *Event, api: *const Api) !void { + pub const await_ = awaitt; + pub fn awaitt(self: *Event, api: *const Api) ApiError!void { defer self.deinit(api); if (self.isReady(api)) { @@ -201,13 +255,13 @@ pub const LoadedExecutable = opaque { }; pub fn execute(self: *const LoadedExecutable, api: *const Api, args: ExecuteArgs) ExecuteError!void { - try asynk.callBlocking(pjrt.LoadedExecutable.execute, .{ self.inner(), api, pjrt.LoadedExecutable.ExecuteArgs{ + try self.inner().execute(api, pjrt.LoadedExecutable.ExecuteArgs{ .num_args = args.num_args, .arguments = @ptrCast(args.arguments), .results = @ptrCast(args.results), .events = @ptrCast(args.events), .non_donatable_input_indices = args.non_donatable_input_indices, - } }); + }); } pub fn getExecutable(self: *LoadedExecutable, api: *const Api) ApiError!*Executable { diff --git a/zml/platform.zig b/zml/platform.zig index ec4ae4c6..76acb159 100644 --- a/zml/platform.zig +++ b/zml/platform.zig @@ -82,8 +82,8 @@ pub const Platform = struct { /// Returns the Profiler for this API. /// Not all platform have a profiling api, for those the profiler object will do nothing. /// Platforms with known profiler extensions: cuda, xpu - pub fn getProfiler(self: Platform, options: ?pjrt.Profiler.Options) pjrt.Profiler { - return self.pjrt_client.getProfiler(self.pjrt_api, options orelse pjrt.Profiler.default_options); + pub fn getProfiler(self: Platform, options: ?pjrt.profiler.Profiler.Options) pjrt.profiler.Profiler { + return self.pjrt_client.getProfiler(self.pjrt_api, options orelse pjrt.profiler.Profiler.default_options); } }; diff --git a/zml/tensor.zig b/zml/tensor.zig index 23ab003b..42ffa5ca 100644 --- a/zml/tensor.zig +++ b/zml/tensor.zig @@ -2419,12 +2419,17 @@ pub const Tensor = struct { /// of the operator which is backend specific. update_fn: *const fn (Tensor, Tensor) Tensor = increment, + pub fn maximum(old_value: Tensor, new_value: Tensor) Tensor { + return old_value.maximum(new_value.convert(old_value.dtype())); + } + pub fn increment(old_value: Tensor, new_value: Tensor) Tensor { - return old_value.add(new_value.convert(old_value.dtype())); + return old_value.add(new_value); } pub fn override(old_value: Tensor, new_value: Tensor) Tensor { - return new_value.convert(old_value.dtype()); + _ = old_value; + return new_value; } }; diff --git a/zml/testing.zig b/zml/testing.zig index e686b603..4ae4140c 100644 --- a/zml/testing.zig +++ b/zml/testing.zig @@ -41,11 +41,13 @@ pub fn expectClose(left_: anytype, right_: anytype, tolerance: f32) !void { .{ try left_.toHostAlloc(allocator), true } else .{ left_, false }; + _ = try left.awaitt(); var right: zml.HostBuffer, const should_free_right = if (@TypeOf(right_) == zml.Buffer) .{ try right_.toHostAlloc(allocator), true } else .{ right_, false }; + _ = try right.awaitt(); defer { if (should_free_left) left.deinit(allocator); diff --git a/zml/zml.zig b/zml/zml.zig index 088ad687..8c075bcf 100644 --- a/zml/zml.zig +++ b/zml/zml.zig @@ -18,6 +18,7 @@ pub const Tensor = @import("tensor.zig").Tensor; // Namespaces pub const context = @import("context.zig"); +pub const buffer = @import("buffer.zig"); pub const exe = @import("exe.zig"); pub const floats = @import("floats.zig"); pub const helpers = @import("helpers.zig");