From 8cfbdb14ad61070b7f153b291dc6ac6b35cef600 Mon Sep 17 00:00:00 2001 From: Corentin Godeau Date: Fri, 13 Nov 2026 10:20:41 +0000 Subject: [PATCH 1/4] zml: introduce `zml/meta` and `zml/constants` modules --- zml/BUILD.bazel | 19 +++++++++++++++++-- zml/Compiler.zig | 2 +- zml/buffer.zig | 4 ++-- zml/exe.zig | 2 +- zml/io.zig | 2 +- zml/mem.zig | 2 +- zml/nn.zig | 4 ++-- zml/ops.zig | 4 ++-- zml/ops/manual_computation.zig | 2 +- zml/pjrtx.zig | 2 +- zml/platform.zig | 2 +- zml/shape.zig | 2 +- zml/slice.zig | 4 ++-- zml/tensor.zig | 4 ++-- zml/zml.zig | 8 ++++---- 15 files changed, 39 insertions(+), 24 deletions(-) diff --git a/zml/BUILD.bazel b/zml/BUILD.bazel index f52ecdd5..5f8b93f1 100644 --- a/zml/BUILD.bazel +++ b/zml/BUILD.bazel @@ -73,6 +73,21 @@ zig_library( visibility = ["//visibility:public"], ) +zig_library( + name = "meta", + import_name = "zml/meta", + main = "meta.zig", + visibility = ["//visibility:public"], + deps = ["//stdx"], +) + +zig_library( + name = "constants", + import_name = "zml/constants", + main = "constants.zig", + visibility = ["//visibility:public"], +) + zig_library( name = "zml", srcs = [ @@ -82,7 +97,6 @@ zig_library( "attention/attention.zig", "buffer.zig", "buffer/HostAccessible.zig", - "constants.zig", "dtype.zig", "exe.zig", "floats.zig", @@ -97,7 +111,6 @@ zig_library( "kernels/paged_attention/paged_attention.zig", "logo.zig", "mem.zig", - "meta.zig", "mlirx.zig", "nn.zig", "ops.zig", @@ -146,6 +159,8 @@ zig_library( "-freference-trace=20", ], deps = [ + ":constants", + ":meta", ":posix", ":profiler_options", ":profiling_traceme_lib", diff --git a/zml/Compiler.zig b/zml/Compiler.zig index 93eafa8d..73b5057a 100644 --- a/zml/Compiler.zig +++ b/zml/Compiler.zig @@ -2,6 +2,7 @@ const std = @import("std"); const c = @import("c"); const func = @import("mlir/dialects/func"); +const meta = @import("zml/meta"); const mlir = @import("mlir"); const pjrt = @import("pjrt"); const shardy = @import("mlir/dialects/shardy"); @@ -14,7 +15,6 @@ const Buffer = @import("buffer.zig").Buffer; const DataType = @import("dtype.zig").DataType; const Exe = @import("exe.zig").Exe; const Memory = @import("platform.zig").Memory; -const meta = @import("meta.zig"); const mlirx = @import("mlirx.zig"); const ops = @import("ops.zig"); const pjrtx = @import("pjrtx.zig"); diff --git a/zml/buffer.zig b/zml/buffer.zig index 8714d80f..70be7c83 100644 --- a/zml/buffer.zig +++ b/zml/buffer.zig @@ -1,14 +1,14 @@ const std = @import("std"); +const constants = @import("zml/constants"); +const meta = @import("zml/meta"); const pjrt = @import("pjrt"); const platforms = @import("platforms"); const stdx = @import("stdx"); -const constants = @import("constants.zig"); const DataType = @import("dtype.zig").DataType; const mem = @import("mem.zig"); const Memory = @import("platform.zig").Memory; -const meta = @import("meta.zig"); const pjrtx = @import("pjrtx.zig"); const Platform = @import("platform.zig").Platform; const Shape = @import("shape.zig").Shape; diff --git a/zml/exe.zig b/zml/exe.zig index 32c0ac48..93895b85 100644 --- a/zml/exe.zig +++ b/zml/exe.zig @@ -1,13 +1,13 @@ const std = @import("std"); const builtin = @import("builtin"); +const meta = @import("zml/meta"); const pjrt = @import("pjrt"); const stdx = @import("stdx"); const Buffer = @import("buffer.zig").Buffer; const Compiler = @import("Compiler.zig"); const mem = @import("mem.zig"); -const meta = @import("meta.zig"); const Platform = @import("platform.zig").Platform; const tracer = @import("profiling/tracer.zig"); const Shape = @import("shape.zig").Shape; diff --git a/zml/io.zig b/zml/io.zig index ab8d8e6e..35501213 100644 --- a/zml/io.zig +++ b/zml/io.zig @@ -1,5 +1,6 @@ const std = @import("std"); +const meta = @import("zml/meta"); const pjrt = @import("pjrt"); const stdx = @import("stdx"); pub const VFS = @import("vfs"); @@ -10,7 +11,6 @@ const CreateOptions = @import("platform.zig").CreateOptions; const Exe = @import("exe.zig").Exe; const mem = @import("mem.zig"); const Memory = @import("platform.zig").Memory; -const meta = @import("meta.zig"); const pjrtx = @import("pjrtx.zig"); const Platform = @import("platform.zig").Platform; const tracer = @import("profiling/tracer.zig"); diff --git a/zml/mem.zig b/zml/mem.zig index a64ae968..3007044c 100644 --- a/zml/mem.zig +++ b/zml/mem.zig @@ -3,12 +3,12 @@ const Allocator = std.mem.Allocator; const Alignment = std.mem.Alignment; const assert = std.debug.assert; +const meta = @import("zml/meta"); const pjrt = @import("pjrt"); const Buffer = @import("buffer.zig").Buffer; const Device = @import("platform.zig").Device; const Memory = @import("platform.zig").Memory; -const meta = @import("meta.zig"); const Platform = @import("platform.zig").Platform; const Tensor = @import("tensor.zig").Tensor; diff --git a/zml/nn.zig b/zml/nn.zig index eceb316b..71474302 100644 --- a/zml/nn.zig +++ b/zml/nn.zig @@ -1,13 +1,13 @@ //! Common layer definition and functions for Neural Networks (NN) const std = @import("std"); +const constants = @import("zml/constants"); +const meta = @import("zml/meta"); 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"); diff --git a/zml/ops.zig b/zml/ops.zig index f8fdba7e..f1b0a8bd 100644 --- a/zml/ops.zig +++ b/zml/ops.zig @@ -1,7 +1,9 @@ const std = @import("std"); const builtin = @import("builtin"); +const constants = @import("zml/constants"); const func = @import("mlir/dialects/func"); +const meta = @import("zml/meta"); const mlir = @import("mlir"); const pjrt = @import("pjrt"); const platforms = @import("platforms"); @@ -11,10 +13,8 @@ const stdx = @import("stdx"); const Buffer = @import("buffer.zig").Buffer; const Bufferized = @import("mem.zig").Bufferized; const Compiler = @import("Compiler.zig"); -const constants = @import("constants.zig"); 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; diff --git a/zml/ops/manual_computation.zig b/zml/ops/manual_computation.zig index a0ec0037..d39ef844 100644 --- a/zml/ops/manual_computation.zig +++ b/zml/ops/manual_computation.zig @@ -6,7 +6,7 @@ const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); const Compiler = @import("../Compiler.zig"); -const meta = @import("../meta.zig"); +const meta = @import("zml/meta"); const mlirx = @import("../mlirx.zig"); const Shape = @import("../shape.zig").Shape; const Sharding = @import("../Sharding.zig"); diff --git a/zml/pjrtx.zig b/zml/pjrtx.zig index 824d2bd1..3185b608 100644 --- a/zml/pjrtx.zig +++ b/zml/pjrtx.zig @@ -1,11 +1,11 @@ const std = @import("std"); +const meta = @import("zml/meta"); const mlir = @import("mlir"); const pjrt = @import("pjrt"); const stablehlo = @import("mlir/dialects/stablehlo"); pub const DataType = @import("dtype.zig").DataType; -const meta = @import("meta.zig"); const Shape = @import("shape.zig").Shape; const Tensor = @import("tensor.zig").Tensor; diff --git a/zml/platform.zig b/zml/platform.zig index 54fb90e4..03acd864 100644 --- a/zml/platform.zig +++ b/zml/platform.zig @@ -2,11 +2,11 @@ const std = @import("std"); const builtin = @import("builtin"); const c = @import("c"); +const constants = @import("zml/constants"); const pjrt = @import("pjrt"); const stdx = @import("stdx"); 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"); diff --git a/zml/shape.zig b/zml/shape.zig index 386b594e..dddcdf63 100644 --- a/zml/shape.zig +++ b/zml/shape.zig @@ -2,9 +2,9 @@ const std = @import("std"); const testing = std.testing; const builtin = @import("builtin"); +const constants = @import("zml/constants"); const stdx = @import("stdx"); -const constants = @import("constants.zig"); const DataType = @import("dtype.zig").DataType; const Sharding = @import("Sharding.zig"); diff --git a/zml/slice.zig b/zml/slice.zig index 9451165b..8b593bd2 100644 --- a/zml/slice.zig +++ b/zml/slice.zig @@ -1,12 +1,12 @@ const std = @import("std"); +const builtin = @import("builtin"); +const constants = @import("zml/constants"); const stdx = @import("stdx"); -const constants = @import("constants.zig"); const DataType = @import("dtype.zig").DataType; const floats = @import("floats.zig"); const Shape = @import("shape.zig").Shape; -const builtin = @import("builtin"); pub fn isBytes(comptime T: type) bool { const type_info = @typeInfo(T); diff --git a/zml/tensor.zig b/zml/tensor.zig index ee084835..53c4ffe7 100644 --- a/zml/tensor.zig +++ b/zml/tensor.zig @@ -1,17 +1,17 @@ const std = @import("std"); const builtin = @import("builtin"); +const constants = @import("zml/constants"); +const meta = @import("zml/meta"); const mlir = @import("mlir"); const shardy = @import("mlir/dialects/shardy"); const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); const Compiler = @import("Compiler.zig"); -const constants = @import("constants.zig"); const DataType = @import("dtype.zig").DataType; const mem = @import("mem.zig"); const Memory = @import("platform.zig").Memory; -const meta = @import("meta.zig"); const mlirx = @import("mlirx.zig"); const ops = @import("ops.zig"); const Platform = @import("platform.zig").Platform; diff --git a/zml/zml.zig b/zml/zml.zig index dafc3a0f..c6a18fa3 100644 --- a/zml/zml.zig +++ b/zml/zml.zig @@ -8,16 +8,16 @@ const std = @import("std"); pub const bazel = @import("bazel"); const c = @import("c"); +pub const constants = @import("zml/constants"); +pub const meta = @import("zml/meta"); pub const pjrt = @import("pjrt"); const runfiles = @import("runfiles"); pub const stdx = @import("stdx"); pub const tokenizer = @import("zml/tokenizer"); pub const attention = @import("attention/attention.zig"); -pub const paged_attention = @import("kernels/paged_attention/paged_attention.zig"); pub const Buffer = @import("buffer.zig").Buffer; pub const Compiler = @import("Compiler.zig"); -pub const constants = @import("constants.zig"); pub const dtype = @import("dtype.zig"); pub const Data = dtype.Data; pub const DataType = dtype.DataType; @@ -27,12 +27,12 @@ pub const FnExe = exe.FnExe; pub const floats = @import("floats.zig"); pub const io = @import("io.zig"); pub const kernel = @import("kernel/kernel.zig"); +pub const moe = @import("kernels/moe/moe.zig"); +pub const paged_attention = @import("kernels/paged_attention/paged_attention.zig"); pub const logo = @import("logo.zig"); pub const mem = @import("mem.zig"); pub const Bufferized = mem.Bufferized; -pub const meta = @import("meta.zig"); pub const mlir = @import("mlirx.zig"); -pub const moe = @import("kernels/moe/moe.zig"); pub const nn = @import("nn.zig"); pub const ops = @import("ops.zig"); pub const pjrtx = @import("pjrtx.zig"); From 7281f75e82caa72eeb09e0b99b47dc9e5c99f172 Mon Sep 17 00:00:00 2001 From: Corentin Godeau Date: Fri, 13 Nov 2026 10:28:09 +0000 Subject: [PATCH 2/4] zml: introduce `zml/profiling` module --- zml/BUILD.bazel | 29 ++++++++++++++++++++++++----- zml/Compiler.zig | 2 +- zml/exe.zig | 2 +- zml/io.zig | 2 +- zml/platform.zig | 2 +- zml/profiling/profiling.zig | 2 ++ zml/zml.zig | 2 +- 7 files changed, 31 insertions(+), 10 deletions(-) create mode 100644 zml/profiling/profiling.zig diff --git a/zml/BUILD.bazel b/zml/BUILD.bazel index 5f8b93f1..b3d9a006 100644 --- a/zml/BUILD.bazel +++ b/zml/BUILD.bazel @@ -88,6 +88,29 @@ zig_library( visibility = ["//visibility:public"], ) +zig_library( + name = "profiling", + srcs = [ + "profiling/profiler.zig", + "profiling/tracer.zig", + ], + import_name = "zml/profiling", + main = "profiling/profiling.zig", + tags = ["manual"], + visibility = ["//visibility:public"], + zigopts = ["-lc"], + deps = [ + ":profiler_options", + ":profiling_traceme_lib", + "//ffi:zig", + "//pjrt", + "//platforms", + "//stdx", + "//tools/xspace_to_perfetto:xspace_to_perfetto_lib", + "//upb", + ], +) + zig_library( name = "zml", srcs = [ @@ -117,8 +140,6 @@ zig_library( "ops/manual_computation.zig", "pjrtx.zig", "platform.zig", - "profiling/profiler.zig", - "profiling/tracer.zig", "quantization.zig", "safetensors.zig", "shape.zig", @@ -162,8 +183,7 @@ zig_library( ":constants", ":meta", ":posix", - ":profiler_options", - ":profiling_traceme_lib", + ":profiling", ":xla_compile_options_upb", ":xla_data_upb", "//bazel", @@ -188,7 +208,6 @@ zig_library( "//platforms", "//platforms/tpu:ragged_paged", "//stdx", - "//tools/xspace_to_perfetto:xspace_to_perfetto_lib", "//upb", "//vfs", "//zml/tokenizer", diff --git a/zml/Compiler.zig b/zml/Compiler.zig index 73b5057a..1bf994cc 100644 --- a/zml/Compiler.zig +++ b/zml/Compiler.zig @@ -8,6 +8,7 @@ const pjrt = @import("pjrt"); const shardy = @import("mlir/dialects/shardy"); const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); +const tracer = @import("zml/profiling").tracer; const upb = @import("upb"); const zio = @import("zio"); @@ -19,7 +20,6 @@ const mlirx = @import("mlirx.zig"); const ops = @import("ops.zig"); const pjrtx = @import("pjrtx.zig"); const Platform = @import("platform.zig").Platform; -const tracer = @import("profiling/tracer.zig"); const Shape = @import("shape.zig").Shape; const Sharding = @import("Sharding.zig"); const Tensor = @import("tensor.zig").Tensor; diff --git a/zml/exe.zig b/zml/exe.zig index 93895b85..e370c708 100644 --- a/zml/exe.zig +++ b/zml/exe.zig @@ -4,12 +4,12 @@ const builtin = @import("builtin"); const meta = @import("zml/meta"); const pjrt = @import("pjrt"); const stdx = @import("stdx"); +const tracer = @import("zml/profiling").tracer; const Buffer = @import("buffer.zig").Buffer; const Compiler = @import("Compiler.zig"); const mem = @import("mem.zig"); const Platform = @import("platform.zig").Platform; -const tracer = @import("profiling/tracer.zig"); const Shape = @import("shape.zig").Shape; const Sharding = @import("Sharding.zig"); const Tensor = @import("tensor.zig").Tensor; diff --git a/zml/io.zig b/zml/io.zig index 35501213..15e6c21d 100644 --- a/zml/io.zig +++ b/zml/io.zig @@ -3,6 +3,7 @@ const std = @import("std"); const meta = @import("zml/meta"); const pjrt = @import("pjrt"); const stdx = @import("stdx"); +const tracer = @import("zml/profiling").tracer; pub const VFS = @import("vfs"); const Buffer = @import("buffer.zig").Buffer; @@ -13,7 +14,6 @@ const mem = @import("mem.zig"); const Memory = @import("platform.zig").Memory; const pjrtx = @import("pjrtx.zig"); const Platform = @import("platform.zig").Platform; -const tracer = @import("profiling/tracer.zig"); const safetensors = @import("safetensors.zig"); const Shape = @import("shape.zig").Shape; const Sharding = @import("Sharding.zig"); diff --git a/zml/platform.zig b/zml/platform.zig index 03acd864..7fb5c3d2 100644 --- a/zml/platform.zig +++ b/zml/platform.zig @@ -4,6 +4,7 @@ const builtin = @import("builtin"); const c = @import("c"); const constants = @import("zml/constants"); const pjrt = @import("pjrt"); +const profiler_ = @import("zml/profiling").profiler; const stdx = @import("stdx"); pub const Target = @import("platforms").Platform; @@ -11,7 +12,6 @@ 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"); const zml = @import("zml.zig"); diff --git a/zml/profiling/profiling.zig b/zml/profiling/profiling.zig new file mode 100644 index 00000000..018f0e0c --- /dev/null +++ b/zml/profiling/profiling.zig @@ -0,0 +1,2 @@ +pub const profiler = @import("profiler.zig"); +pub const tracer = @import("tracer.zig"); diff --git a/zml/zml.zig b/zml/zml.zig index c6a18fa3..89c6e5f5 100644 --- a/zml/zml.zig +++ b/zml/zml.zig @@ -14,6 +14,7 @@ pub const pjrt = @import("pjrt"); const runfiles = @import("runfiles"); pub const stdx = @import("stdx"); pub const tokenizer = @import("zml/tokenizer"); +pub const tracer = @import("zml/profiling").tracer; pub const attention = @import("attention/attention.zig"); pub const Buffer = @import("buffer.zig").Buffer; @@ -40,7 +41,6 @@ pub const platform = @import("platform.zig"); pub const Memory = platform.Memory; pub const Platform = platform.Platform; pub const Target = platform.Target; -pub const tracer = @import("profiling/tracer.zig"); pub const quantization = @import("quantization.zig"); pub const Quantization = quantization.Quantization; pub const safetensors = @import("safetensors.zig"); From 3b5d66913a3b02262a3b206939c6af9781917824 Mon Sep 17 00:00:00 2001 From: Corentin Godeau Date: Fri, 13 Nov 2026 10:37:22 +0000 Subject: [PATCH 3/4] zml: introduce `zml/core` module --- zml/BUILD.bazel | 104 ++++++++++------ zml/Sharding.zig | 2 +- zml/buffer.zig | 2 +- zml/buffer/HostAccessible.zig | 6 +- zml/core.zig | 82 +++++++++++++ zml/io.zig | 2 +- zml/kernel/fly/test.zig | 2 +- zml/kernel/metal/test.zig | 2 +- zml/kernel/tcl/test.zig | 2 +- .../flash_attention/metal/flash_attn.zig | 2 +- zml/kernels/moe/api.zig | 2 +- .../moe/flashinfer_cutlass/BUILD.bazel | 6 +- .../flashinfer_cutlass/flashinfer_cutlass.zig | 4 +- zml/kernels/moe/weights.zig | 4 +- .../paged_attention/metal/paged_attn.zig | 2 +- zml/kernels/paged_attention/options.zig | 2 +- zml/nn.zig | 2 +- zml/ops.zig | 22 ++-- zml/ops/manual_computation.zig | 6 +- zml/platform.zig | 2 +- zml/quantization.zig | 2 +- zml/tensor.zig | 72 +++++------ zml/testing.zig | 2 +- zml/zml.zig | 115 +++++++----------- 24 files changed, 272 insertions(+), 177 deletions(-) create mode 100644 zml/core.zig diff --git a/zml/BUILD.bazel b/zml/BUILD.bazel index b3d9a006..f5e8c79c 100644 --- a/zml/BUILD.bazel +++ b/zml/BUILD.bazel @@ -112,27 +112,17 @@ zig_library( ) zig_library( - name = "zml", + name = "core", srcs = [ "Compiler.zig", "Sharding.zig", - "attention/Options.zig", - "attention/attention.zig", "buffer.zig", "buffer/HostAccessible.zig", "dtype.zig", "exe.zig", "floats.zig", "io.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", + "//zml/kernel", "mem.zig", "mlirx.zig", "nn.zig", @@ -146,8 +136,71 @@ zig_library( "slice.zig", "tensor.zig", "testing.zig", + "//zml/kernels/moe/flashinfer_cutlass:runtime_srcs", + ], + import_name = "zml/core", + main = "core.zig", + tags = ["manual"], + visibility = ["//visibility:public"], + zigopts = [ + "-lc", + "-freference-trace=20", + ], + deps = [ + ":constants", + ":meta", + ":posix", + ":profiling", + ":xla_compile_options_upb", + ":xla_data_upb", + "//bazel", + "//dsl/cuda_tile:builder", + "//dsl/cute:builder", + "//dsl/fly:builder", + "//dsl/metal:builder", + "//dsl/mosaic_tpu:builder", + "//dsl/tcl:builder", + "//dsl/triton:builder", + "//ffi:zig", + "//mlir", + "//mlir/dialects/func", + "//mlir/dialects/mosaic_tpu", + "//mlir/dialects/shardy", + "//mlir/dialects/stablehlo", + "//pjrt", + "//platforms", + "//stdx", + "//upb", + "//vfs", + "@zio", + ] + select({ + "//platforms:neuron.enabled": [ + "//platforms/neuron:nki_kernel", + "//platforms/neuron:topology", + ], + "//conditions:default": [], + }) + select({ + "//platforms:cuda.enabled": [ + "//platforms/cuda/flashinfer_cutlass_moe", + ], + "//conditions:default": [], + }), +) + +zig_library( + name = "zml", + srcs = [ + "attention/Options.zig", + "attention/attention.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", "zml.zig", - "//zml/kernel", "//zml/kernels/flash_attention/cute", "//zml/kernels/flash_attention/metal", "//zml/kernels/flash_attention/nki", @@ -180,43 +233,22 @@ zig_library( "-freference-trace=20", ], deps = [ - ":constants", - ":meta", - ":posix", - ":profiling", - ":xla_compile_options_upb", - ":xla_data_upb", - "//bazel", - "//dsl/cuda_tile:builder", - "//dsl/cute:builder", - "//dsl/fly:builder", - "//dsl/metal:builder", + ":core", + ":logo", "//dsl/mosaic_tpu:builder", "//dsl/tcl:builder", - "//dsl/triton:builder", - "//ffi:zig", "//mlir", - "//mlir/dialects/func", "//mlir/dialects/llvm", - "//mlir/dialects/mosaic_tpu", "//mlir/dialects/nvvm", - "//mlir/dialects/shardy", - "//mlir/dialects/stablehlo", "//mlir/dialects/ttir", "//mlir/dialects/vector", - "//pjrt", "//platforms", "//platforms/tpu:ragged_paged", "//stdx", - "//upb", - "//vfs", "//zml/tokenizer", - "@rules_zig//zig/runfiles", - "@zio", ] + select({ "//platforms:neuron.enabled": [ "//platforms/neuron:nki_kernel", - "//platforms/neuron:topology", ], "//conditions:default": [], }) + select({ diff --git a/zml/Sharding.zig b/zml/Sharding.zig index 040216bc..1f05b4df 100644 --- a/zml/Sharding.zig +++ b/zml/Sharding.zig @@ -3249,7 +3249,7 @@ test "GSPMD attributes preserve folded logical axis order" { test "GSPMD executes a reduction across a folded logical axis" { if (comptime !Target.cpu.isEnabled()) return error.SkipZigTest; - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const io = std.testing.io; const allocator = std.testing.allocator; diff --git a/zml/buffer.zig b/zml/buffer.zig index 70be7c83..81112de9 100644 --- a/zml/buffer.zig +++ b/zml/buffer.zig @@ -376,7 +376,7 @@ pub const Buffer = struct { }; test "device round-trip" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const io = std.testing.io; const allocator = std.testing.allocator; const platform = zml.testing.env(); diff --git a/zml/buffer/HostAccessible.zig b/zml/buffer/HostAccessible.zig index b3e82a9c..db17ae7b 100644 --- a/zml/buffer/HostAccessible.zig +++ b/zml/buffer/HostAccessible.zig @@ -1878,7 +1878,7 @@ fn expectDevicesReadShards(platform: *const Platform, shape_: Shape) !void { } test "HostAccessible executions read shared inputs on every device" { - const zml = @import("../zml.zig"); + const zml = @import("../core.zig"); const allocator = std.testing.allocator; const io = std.testing.io; const platform = testing.env(); @@ -2107,7 +2107,7 @@ test "HostAccessible.View copies across layouts, shards and replicas" { } test "HostAccessible await after executable donation permits the next host update" { - const zml = @import("../zml.zig"); + const zml = @import("../core.zig"); const allocator = std.testing.allocator; const io = std.testing.io; const platform = testing.env(); @@ -2159,7 +2159,7 @@ test "HostAccessible await after executable donation permits the next host updat } test "HostAccessible await follows an output that doesn't reuse the pinned allocation" { - const zml = @import("../zml.zig"); + const zml = @import("../core.zig"); const allocator = std.testing.allocator; const io = std.testing.io; const platform = testing.env(); diff --git a/zml/core.zig b/zml/core.zig new file mode 100644 index 00000000..65154db0 --- /dev/null +++ b/zml/core.zig @@ -0,0 +1,82 @@ +const std = @import("std"); + +pub const bazel = @import("bazel"); +pub const constants = @import("zml/constants"); +pub const meta = @import("zml/meta"); +pub const pjrt = @import("pjrt"); +pub const stdx = @import("stdx"); +pub const tracer = @import("zml/profiling").tracer; + +pub const Buffer = @import("buffer.zig").Buffer; +pub const Compiler = @import("Compiler.zig"); +pub const dtype = @import("dtype.zig"); +pub const Data = dtype.Data; +pub const DataType = dtype.DataType; +pub const exe = @import("exe.zig"); +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/kernel.zig"); +pub const mem = @import("mem.zig"); +pub const Bufferized = mem.Bufferized; +pub const mlir = @import("mlirx.zig"); +pub const nn = @import("nn.zig"); +pub const ops = @import("ops.zig"); +pub const pjrtx = @import("pjrtx.zig"); +pub const platform = @import("platform.zig"); +pub const Memory = platform.Memory; +pub const Platform = platform.Platform; +pub const Target = platform.Target; +pub const quantization = @import("quantization.zig"); +pub const Quantization = quantization.Quantization; +pub const safetensors = @import("safetensors.zig"); +pub const shape = @import("shape.zig"); +pub const Shape = shape.Shape; +pub const Sharding = @import("Sharding.zig"); +pub const Mesh = Sharding.Mesh; +pub const slice = @import("slice.zig"); +pub const Slice = slice.Slice; +pub const tensor = @import("tensor.zig"); +pub const Tensor = tensor.Tensor; +pub const testing = @import("testing.zig"); + +// Platform owns registration and runner lifetime until these are separated from core. +pub const flashinfer_cutlass = struct { + pub const contract = @import("kernels/moe/flashinfer_cutlass/contract.zig"); + pub const runtime = @import("kernels/moe/flashinfer_cutlass/runtime.zig"); +}; + +/// Deprecated, use `zml.Compiler` instead +pub const module = struct { + pub const CompilationContext = Compiler; + pub const CompileError = Compiler.Error; + pub const CompilationOptions = Compiler.Options; + + /// Deprecated, use `platform.compileFn` or `zml.Compiler.compile` + pub fn compile( + allocator: std.mem.Allocator, + io_: std.Io, + comptime func: anytype, + args: std.meta.ArgsTuple(@TypeOf(func)), + platform_: *const Platform, + opts: Compiler.Options, + ) Compiler.Error!Exe { + return Compiler.compile(allocator, io_, platform_, func, args, opts); + } +}; + +test "core" { + std.testing.refAllDecls(@This()); +} + +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); +} diff --git a/zml/io.zig b/zml/io.zig index 15e6c21d..93829541 100644 --- a/zml/io.zig +++ b/zml/io.zig @@ -7,7 +7,7 @@ const tracer = @import("zml/profiling").tracer; pub const VFS = @import("vfs"); const Buffer = @import("buffer.zig").Buffer; -const Bufferized = @import("zml.zig").Bufferized; +const Bufferized = @import("core.zig").Bufferized; const CreateOptions = @import("platform.zig").CreateOptions; const Exe = @import("exe.zig").Exe; const mem = @import("mem.zig"); diff --git a/zml/kernel/fly/test.zig b/zml/kernel/fly/test.zig index ae788d96..6b77ad3e 100644 --- a/zml/kernel/fly/test.zig +++ b/zml/kernel/fly/test.zig @@ -3,7 +3,7 @@ const std = @import("std"); const fly_builder = @import("dsl/fly/builder"); const fly = @import("fly.zig"); -const zml = @import("../../zml.zig"); +const zml = @import("../../core.zig"); /// C = A + B over (M, N) f32, 128-bit vectorized and predicated on the /// ragged edge. diff --git a/zml/kernel/metal/test.zig b/zml/kernel/metal/test.zig index 0a342af0..a184869c 100644 --- a/zml/kernel/metal/test.zig +++ b/zml/kernel/metal/test.zig @@ -1,7 +1,7 @@ const std = @import("std"); const metal = @import("metal.zig"); -const zml = @import("../../zml.zig"); +const zml = @import("../../core.zig"); /// `out[i] = a[i] * scale + b[i]` over `n` floats. const MetalAxpy = struct { diff --git a/zml/kernel/tcl/test.zig b/zml/kernel/tcl/test.zig index b9ae1875..20804b8b 100644 --- a/zml/kernel/tcl/test.zig +++ b/zml/kernel/tcl/test.zig @@ -1,6 +1,6 @@ const std = @import("std"); -const zml = @import("../../zml.zig"); +const zml = @import("../../core.zig"); const tcl = @import("tcl.zig"); /// `furiosa.tcl.examples.softmax` over f32. diff --git a/zml/kernels/flash_attention/metal/flash_attn.zig b/zml/kernels/flash_attention/metal/flash_attn.zig index de4b68da..4d584dab 100644 --- a/zml/kernels/flash_attention/metal/flash_attn.zig +++ b/zml/kernels/flash_attention/metal/flash_attn.zig @@ -4,7 +4,7 @@ const std = @import("std"); -const metal = @import("../../../kernel/kernel.zig").metal; +const metal = @import("zml/core").kernel.metal; const ArgSpec = metal.ArgSpec; const Builder = metal.Builder; diff --git a/zml/kernels/moe/api.zig b/zml/kernels/moe/api.zig index b66593c1..b11ac1d5 100644 --- a/zml/kernels/moe/api.zig +++ b/zml/kernels/moe/api.zig @@ -2,7 +2,7 @@ const std = @import("std"); const stdx = @import("stdx"); -const Tensor = @import("../../tensor.zig").Tensor; +const Tensor = @import("zml/core").Tensor; const log = std.log.scoped(.@"zml/moe"); diff --git a/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel b/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel index c9c5d9a6..d94ff51b 100644 --- a/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel +++ b/zml/kernels/moe/flashinfer_cutlass/BUILD.bazel @@ -2,9 +2,13 @@ package(default_visibility = ["//zml:__pkg__"]) filegroup( name = "flashinfer_cutlass", + srcs = ["flashinfer_cutlass.zig"], +) + +filegroup( + name = "runtime_srcs", srcs = [ "contract.zig", - "flashinfer_cutlass.zig", "runtime.zig", ], ) diff --git a/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig b/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig index 92317c7c..715198f9 100644 --- a/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig +++ b/zml/kernels/moe/flashinfer_cutlass/flashinfer_cutlass.zig @@ -9,9 +9,9 @@ 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 contract = @import("zml/core").flashinfer_cutlass.contract; const Attributes = contract.Attributes; -const runtime = @import("runtime.zig"); +const runtime = @import("zml/core").flashinfer_cutlass.runtime; const log = std.log.scoped(.@"zml/moe/flashinfer_cutlass"); diff --git a/zml/kernels/moe/weights.zig b/zml/kernels/moe/weights.zig index 0e288856..69036523 100644 --- a/zml/kernels/moe/weights.zig +++ b/zml/kernels/moe/weights.zig @@ -1,5 +1,5 @@ -const nn = @import("../../nn.zig"); -const Tensor = @import("../../tensor.zig").Tensor; +const nn = @import("zml/core").nn; +const Tensor = @import("zml/core").Tensor; pub fn unpackedWeight(linear: nn.Linear) Tensor { const quantization = linear.quantization orelse return linear.weight; diff --git a/zml/kernels/paged_attention/metal/paged_attn.zig b/zml/kernels/paged_attention/metal/paged_attn.zig index 04f4c1ac..10a088e5 100644 --- a/zml/kernels/paged_attention/metal/paged_attn.zig +++ b/zml/kernels/paged_attention/metal/paged_attn.zig @@ -3,7 +3,7 @@ const std = @import("std"); -const metal = @import("../../../kernel/kernel.zig").metal; +const metal = @import("zml/core").kernel.metal; const ArgSpec = metal.ArgSpec; const Builder = metal.Builder; diff --git a/zml/kernels/paged_attention/options.zig b/zml/kernels/paged_attention/options.zig index 35ad4049..159da9b2 100644 --- a/zml/kernels/paged_attention/options.zig +++ b/zml/kernels/paged_attention/options.zig @@ -1,4 +1,4 @@ -const Tensor = @import("../../tensor.zig").Tensor; +const Tensor = @import("zml/core").Tensor; pub const AttentionMask = union(enum) { standard: struct { diff --git a/zml/nn.zig b/zml/nn.zig index 71474302..592efa2e 100644 --- a/zml/nn.zig +++ b/zml/nn.zig @@ -15,7 +15,7 @@ pub const Quantization = quantization.Quantization; const Shape = @import("shape.zig").Shape; const Slice = @import("slice.zig").Slice; const Tensor = @import("tensor.zig").Tensor; -const zml = @import("zml.zig"); +const zml = @import("core.zig"); const log = std.log.scoped(.@"zml/nn"); diff --git a/zml/ops.zig b/zml/ops.zig index f1b0a8bd..59f5130a 100644 --- a/zml/ops.zig +++ b/zml/ops.zig @@ -667,7 +667,7 @@ pub fn @"while"( pub const while_ = @"while"; test @"while" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const While = struct { @@ -812,7 +812,7 @@ pub fn @"if"( } test "if" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const allocator = std.testing.allocator; @@ -893,7 +893,7 @@ pub fn if2( } test if2 { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const allocator = std.testing.allocator; @@ -1107,7 +1107,7 @@ pub fn fly(inputs: anytype, outputs: anytype, opts: FlyOps) [outputs.len]Tensor } test "fly custom call, both entry ABIs" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); if (platform.target != .rocm) return error.SkipZigTest; @@ -1298,7 +1298,7 @@ pub fn neuronNki(inputs: anytype, outputs: anytype, opts: NeuronNkiOps) [outputs } test "triton" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); if (platform.target != .cuda and platform.target != .rocm and platform.target != .oneapi) return error.SkipZigTest; @@ -1617,7 +1617,7 @@ pub fn cute(inputs: anytype, outputs: anytype, opts: CuteOps) [outputs.len]Tenso } test "cute" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); if (platform.target != .cuda) return error.SkipZigTest; @@ -1682,7 +1682,7 @@ test "cute" { } test "cuda_tile" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); if (platform.target != .cuda) return error.SkipZigTest; @@ -1741,7 +1741,7 @@ test "cuda_tile" { } test "cuda_tile grid" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); if (platform.target != .cuda) return error.SkipZigTest; @@ -1808,7 +1808,7 @@ test "cuda_tile grid" { } test "cuda_tile zeroed_outputs" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); if (platform.target != .cuda) return error.SkipZigTest; @@ -2107,7 +2107,7 @@ fn scatterConfig( } test scatterConfig { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); var comp: zml.Compiler = .init(std.testing.allocator, std.testing.io, platform, .{}); @@ -2984,7 +2984,7 @@ fn customCallArgsFromPjrtCallFrame(I: type, O: type, A: type, call_frame: *pjrt. } test customCall { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const tp = zml.testing.tpMesh(); diff --git a/zml/ops/manual_computation.zig b/zml/ops/manual_computation.zig index d39ef844..68dfaccf 100644 --- a/zml/ops/manual_computation.zig +++ b/zml/ops/manual_computation.zig @@ -376,7 +376,7 @@ const gspmd = struct { }; test manualComputation { - const zml = @import("../zml.zig"); + const zml = @import("../core.zig"); const allocator = std.testing.allocator; const io = std.testing.io; const platform = zml.testing.env(); @@ -436,7 +436,7 @@ test manualComputation { } test "manualComputation composes across logical and physical meshes" { - const zml = @import("../zml.zig"); + const zml = @import("../core.zig"); if (comptime !zml.Target.cpu.isEnabled()) return error.SkipZigTest; const allocator = std.testing.allocator; const io = std.testing.io; @@ -505,7 +505,7 @@ test "manualComputation composes across logical and physical meshes" { } test "partitionSpecs preserves optional tensors and active cache variants" { - const zml = @import("../zml.zig"); + const zml = @import("../core.zig"); const Logical = Sharding.Partitioning.Logical; const Inputs = struct { layers: []const Tensor, diff --git a/zml/platform.zig b/zml/platform.zig index 7fb5c3d2..f4547988 100644 --- a/zml/platform.zig +++ b/zml/platform.zig @@ -13,7 +13,7 @@ const flashinfer_cutlass_contract = @import("kernels/moe/flashinfer_cutlass/cont const flashinfer_cutlass_runtime = @import("kernels/moe/flashinfer_cutlass/runtime.zig"); const pjrtx = @import("pjrtx.zig"); const Sharding = @import("Sharding.zig"); -const zml = @import("zml.zig"); +const zml = @import("core.zig"); const log = std.log.scoped(.zml); diff --git a/zml/quantization.zig b/zml/quantization.zig index db20619a..5c616080 100644 --- a/zml/quantization.zig +++ b/zml/quantization.zig @@ -3,7 +3,7 @@ const std = @import("std"); const stdx = @import("stdx"); -const zml = @import("zml.zig"); +const zml = @import("core.zig"); const DataType = @import("dtype.zig").DataType; const platform_mod = @import("platform.zig"); diff --git a/zml/tensor.zig b/zml/tensor.zig index 53c4ffe7..686f9683 100644 --- a/zml/tensor.zig +++ b/zml/tensor.zig @@ -322,7 +322,7 @@ pub const Tensor = struct { } test toMemory { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const io = std.testing.io; @@ -479,7 +479,7 @@ pub const Tensor = struct { } test reuseBuffer { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const io = std.testing.io; @@ -617,7 +617,7 @@ pub const Tensor = struct { } test fmod { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const inputs: [2][6]f32 = .{ .{ -3.0, -2, -1, 1, 2, 3 }, .{ 1, 2, 3, 4, 5, -5 } }; @@ -799,7 +799,7 @@ pub const Tensor = struct { } test "Tensor.fft IRFFT restores RFFT last dimension" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const rfft_input_len: i64 = 10; @@ -909,7 +909,7 @@ pub const Tensor = struct { } test uniform { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const Stats = struct { const Stats = @This(); @@ -1007,7 +1007,7 @@ pub const Tensor = struct { } test gumbel { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const Stats = struct { const Stats = @This(); @@ -1382,7 +1382,7 @@ pub const Tensor = struct { test "convert f32 -> f4e2m1" { const floats = @import("floats.zig"); - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x_f32 = [_]f32{ 0.0, 0.5, 1, 1.5, 2, 3, 4, 6, -0.0, -0.5, -1, -1.5, -2, -3, -4, -6 }; var x_f4: [x_f32.len / 2]floats.Float4E2M1.Packed = undefined; @@ -1415,7 +1415,7 @@ pub const Tensor = struct { } test "convert f32 -> f8e3m4" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x = [_]f32{ 1.1 / 4.0, 1.1 / 8.0, 1.1 / 16.0, 1.1 / 32.0, 1.1 / 64.0, 1.1 / 128.0 }; @@ -1446,7 +1446,7 @@ pub const Tensor = struct { } test "convert u2 -> u8" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x_u2: [2]stdx.Packed(u2) = .{ .pack(0, 1, 2, 3), .pack(3, 2, 1, 0) }; comptime std.debug.assert(@sizeOf(@TypeOf(x_u2)) == 2); @@ -1476,7 +1476,7 @@ pub const Tensor = struct { test "convert f4e2m1 -> f32" { const floats = @import("floats.zig"); - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x_f4_packed = [_]floats.Float4E2M1.Packed{ .fromF32(0, 0.5), .fromF32(1, 1.5), .fromF32(2, 3), .fromF32(4, 6) }; @@ -1570,7 +1570,7 @@ pub const Tensor = struct { } test dot { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); inline for (.{ @@ -1719,7 +1719,7 @@ pub const Tensor = struct { } test leakyReLU { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const input: Tensor = .init(.{2}, .f32); @@ -1757,7 +1757,7 @@ pub const Tensor = struct { } test swiglu { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const input: Tensor = .init(.{ .d = 4 }, .f32); @@ -2030,7 +2030,7 @@ pub const Tensor = struct { } test cumulativeSum { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -2338,7 +2338,7 @@ pub const Tensor = struct { } test slice { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x: Tensor = .init(.{ 2, 5 }, .f32); @@ -2378,7 +2378,7 @@ pub const Tensor = struct { } test "dynamic slices" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -2498,7 +2498,7 @@ pub const Tensor = struct { } test repeat1d { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -2561,7 +2561,7 @@ pub const Tensor = struct { } test stutter1d { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -2684,7 +2684,7 @@ pub const Tensor = struct { } test arange { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const exe = try zml.module.compile( @@ -2770,7 +2770,7 @@ pub const Tensor = struct { } test scalar { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -3063,7 +3063,7 @@ pub const Tensor = struct { } test gather { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -3217,7 +3217,7 @@ pub const Tensor = struct { } test gatherSlices { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -3426,7 +3426,7 @@ pub const Tensor = struct { } test scatterSlices { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -3648,7 +3648,7 @@ pub const Tensor = struct { } test argMax { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const allocator = std.testing.allocator; const ArgMaxTest = struct { @@ -3720,7 +3720,7 @@ pub const Tensor = struct { } test argsort { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -3991,7 +3991,7 @@ pub const Tensor = struct { } test chunkExact { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); // Only test shapes @@ -4044,7 +4044,7 @@ pub const Tensor = struct { } test chunkAllowTrailing { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); // Only test shapes @@ -4102,7 +4102,7 @@ pub const Tensor = struct { } test split { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -4236,7 +4236,7 @@ pub const Tensor = struct { } test dynamicUpdateSlice { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); { @@ -4432,7 +4432,7 @@ pub const Tensor = struct { } test toDiagonal { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -4485,7 +4485,7 @@ pub const Tensor = struct { } test triangular { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { @@ -4663,7 +4663,7 @@ pub const Tensor = struct { } test cartesianProduct { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x: Tensor = .init(.{6}, .i32); @@ -4741,7 +4741,7 @@ pub const Tensor = struct { } test cartesianProductStacked { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const x: Tensor = .init(.{6}, .i32); const y: Tensor = .init(.{4}, .i32); @@ -4887,7 +4887,7 @@ fn getComparisonType(ctx: *mlir.Context, dtype: DataType) *const stablehlo.Compa } test "Tensor.maxPool1d" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const MaxPool = struct { @@ -4929,7 +4929,7 @@ test "Tensor.maxPool1d" { } test "Tensor.maxPool2d" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const MaxPool = struct { @@ -5038,7 +5038,7 @@ test transposeIsJustAReshape { } test "unused tensor" { - const zml = @import("zml.zig"); + const zml = @import("core.zig"); const platform = zml.testing.env(); const Local = struct { diff --git a/zml/testing.zig b/zml/testing.zig index 5e3d6333..b2671935 100644 --- a/zml/testing.zig +++ b/zml/testing.zig @@ -6,7 +6,7 @@ const stdx = @import("stdx"); const Platform = @import("platform.zig").Platform; const Sharding = @import("Sharding.zig"); const Slice = @import("slice.zig").Slice; -const zml = @import("zml.zig"); +const zml = @import("core.zig"); const log = std.log.scoped(.@"zml/testing"); diff --git a/zml/zml.zig b/zml/zml.zig index 89c6e5f5..2649eb31 100644 --- a/zml/zml.zig +++ b/zml/zml.zig @@ -5,85 +5,62 @@ // Namespaces const std = @import("std"); +const core = @import("zml/core"); -pub const bazel = @import("bazel"); -const c = @import("c"); -pub const constants = @import("zml/constants"); -pub const meta = @import("zml/meta"); -pub const pjrt = @import("pjrt"); -const runfiles = @import("runfiles"); -pub const stdx = @import("stdx"); +pub const bazel = core.bazel; +pub const constants = core.constants; +pub const meta = core.meta; +pub const pjrt = core.pjrt; +pub const stdx = core.stdx; pub const tokenizer = @import("zml/tokenizer"); -pub const tracer = @import("zml/profiling").tracer; +pub const tracer = core.tracer; pub const attention = @import("attention/attention.zig"); -pub const Buffer = @import("buffer.zig").Buffer; -pub const Compiler = @import("Compiler.zig"); -pub const dtype = @import("dtype.zig"); -pub const Data = dtype.Data; -pub const DataType = dtype.DataType; -pub const exe = @import("exe.zig"); -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/kernel.zig"); +pub const Buffer = core.Buffer; +pub const Compiler = core.Compiler; +pub const dtype = core.dtype; +pub const Data = core.Data; +pub const DataType = core.DataType; +pub const exe = core.exe; +pub const Exe = core.Exe; +pub const FnExe = core.FnExe; +pub const floats = core.floats; +pub const io = core.io; +pub const kernel = core.kernel; pub const moe = @import("kernels/moe/moe.zig"); pub const paged_attention = @import("kernels/paged_attention/paged_attention.zig"); -pub const logo = @import("logo.zig"); -pub const mem = @import("mem.zig"); -pub const Bufferized = mem.Bufferized; -pub const mlir = @import("mlirx.zig"); -pub const nn = @import("nn.zig"); -pub const ops = @import("ops.zig"); -pub const pjrtx = @import("pjrtx.zig"); -pub const platform = @import("platform.zig"); -pub const Memory = platform.Memory; -pub const Platform = platform.Platform; -pub const Target = platform.Target; -pub const quantization = @import("quantization.zig"); -pub const Quantization = quantization.Quantization; -pub const safetensors = @import("safetensors.zig"); -pub const shape = @import("shape.zig"); -pub const Shape = shape.Shape; -pub const Sharding = @import("Sharding.zig"); -pub const Mesh = Sharding.Mesh; -pub const slice = @import("slice.zig"); -pub const Slice = slice.Slice; -pub const tensor = @import("tensor.zig"); -pub const Tensor = tensor.Tensor; -pub const testing = @import("testing.zig"); +pub const logo = @import("zml/logo"); +pub const mem = core.mem; +pub const Bufferized = core.Bufferized; +pub const mlir = core.mlir; +pub const nn = core.nn; +pub const ops = core.ops; +pub const pjrtx = core.pjrtx; +pub const platform = core.platform; +pub const Memory = core.Memory; +pub const Platform = core.Platform; +pub const Target = core.Target; +pub const quantization = core.quantization; +pub const Quantization = core.Quantization; +pub const safetensors = core.safetensors; +pub const shape = core.shape; +pub const Shape = core.Shape; +pub const Sharding = core.Sharding; +pub const Mesh = core.Mesh; +pub const slice = core.slice; +pub const Slice = core.Slice; +pub const tensor = core.tensor; +pub const Tensor = core.Tensor; +pub const testing = core.testing; /// Deprecated, use `zml.Compiler` instead -pub const module = struct { - pub const CompilationContext = Compiler; - pub const CompileError = Compiler.Error; - pub const CompilationOptions = Compiler.Options; - - /// Deprecated, use `platform.compileFn` or `zml.Compiler.compile` - pub fn compile( - allocator: std.mem.Allocator, - io_: std.Io, - comptime func: anytype, - args: std.meta.ArgsTuple(@TypeOf(func)), - platform_: *const Platform, - opts: Compiler.Options, - ) Compiler.Error!Exe { - return Compiler.compile(allocator, io_, platform_, func, args, opts); - } -}; +pub const module = core.module; test "zml" { std.testing.refAllDecls(@This()); } -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); -} +pub const KiB = core.KiB; +pub const MiB = core.MiB; +pub const GiB = core.GiB; +pub const TiB = core.TiB; From ac2cb2fd740dbe86478c1d0826b04de0ccbf6a4c Mon Sep 17 00:00:00 2001 From: Corentin Godeau Date: Fri, 13 Nov 2026 10:42:33 +0000 Subject: [PATCH 4/4] zml: introduce `zml/nn` and `zml/quantization` modules --- zml/BUILD.bazel | 33 +++++++++++++++++++++++++++++++-- zml/core.zig | 3 --- zml/kernels/moe/weights.zig | 2 +- zml/nn.zig | 23 ++++++++++++----------- zml/quantization.zig | 10 +++++----- zml/zml.zig | 6 +++--- 6 files changed, 52 insertions(+), 25 deletions(-) diff --git a/zml/BUILD.bazel b/zml/BUILD.bazel index f5e8c79c..25323615 100644 --- a/zml/BUILD.bazel +++ b/zml/BUILD.bazel @@ -125,12 +125,10 @@ zig_library( "//zml/kernel", "mem.zig", "mlirx.zig", - "nn.zig", "ops.zig", "ops/manual_computation.zig", "pjrtx.zig", "platform.zig", - "quantization.zig", "safetensors.zig", "shape.zig", "slice.zig", @@ -187,6 +185,35 @@ zig_library( }), ) +zig_library( + name = "quantization", + import_name = "zml/quantization", + main = "quantization.zig", + tags = ["manual"], + visibility = ["//visibility:public"], + deps = [ + ":core", + "//stdx", + ], +) + +zig_library( + name = "nn", + import_name = "zml/nn", + main = "nn.zig", + tags = ["manual"], + visibility = ["//visibility:public"], + deps = [ + ":constants", + ":core", + ":meta", + ":quantization", + "//mlir", + "//mlir/dialects/stablehlo", + "//stdx", + ], +) + zig_library( name = "zml", srcs = [ @@ -235,6 +262,8 @@ zig_library( deps = [ ":core", ":logo", + ":nn", + ":quantization", "//dsl/mosaic_tpu:builder", "//dsl/tcl:builder", "//mlir", diff --git a/zml/core.zig b/zml/core.zig index 65154db0..521e4c1b 100644 --- a/zml/core.zig +++ b/zml/core.zig @@ -21,15 +21,12 @@ pub const kernel = @import("kernel/kernel.zig"); pub const mem = @import("mem.zig"); pub const Bufferized = mem.Bufferized; pub const mlir = @import("mlirx.zig"); -pub const nn = @import("nn.zig"); pub const ops = @import("ops.zig"); pub const pjrtx = @import("pjrtx.zig"); pub const platform = @import("platform.zig"); pub const Memory = platform.Memory; pub const Platform = platform.Platform; pub const Target = platform.Target; -pub const quantization = @import("quantization.zig"); -pub const Quantization = quantization.Quantization; pub const safetensors = @import("safetensors.zig"); pub const shape = @import("shape.zig"); pub const Shape = shape.Shape; diff --git a/zml/kernels/moe/weights.zig b/zml/kernels/moe/weights.zig index 69036523..fd3c9e82 100644 --- a/zml/kernels/moe/weights.zig +++ b/zml/kernels/moe/weights.zig @@ -1,4 +1,4 @@ -const nn = @import("zml/core").nn; +const nn = @import("zml/nn"); const Tensor = @import("zml/core").Tensor; pub fn unpackedWeight(linear: nn.Linear) Tensor { diff --git a/zml/nn.zig b/zml/nn.zig index 592efa2e..94347552 100644 --- a/zml/nn.zig +++ b/zml/nn.zig @@ -7,15 +7,16 @@ const mlir = @import("mlir"); const stablehlo = @import("mlir/dialects/stablehlo"); const stdx = @import("stdx"); -const DataType = @import("dtype.zig").DataType; -const mlirx = @import("mlirx.zig"); -const ops = @import("ops.zig"); -const quantization = @import("quantization.zig"); +const zml = @import("zml/core"); +const quantization = @import("zml/quantization"); + +const DataType = zml.DataType; +const mlirx = zml.mlir; +const ops = zml.ops; pub const Quantization = quantization.Quantization; -const Shape = @import("shape.zig").Shape; -const Slice = @import("slice.zig").Slice; -const Tensor = @import("tensor.zig").Tensor; -const zml = @import("core.zig"); +const Shape = zml.Shape; +const Slice = zml.Slice; +const Tensor = zml.Tensor; const log = std.log.scoped(.@"zml/nn"); @@ -229,7 +230,7 @@ test "block128 scaled dot layouts" { 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 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 dtype: DataType = if (platform.target == .rocm and zml.platform.rocm.computeCapability(platform) == .gfx942) .f8e4m3fnuz else .f8e4m3fn; const Local = struct { const Outputs = struct { actual: Tensor, expected: Tensor, linear: Tensor, linear_expected: Tensor }; @@ -631,7 +632,7 @@ pub fn rope(x: Tensor, pos_idx: ?Tensor, opts: RopeOpts) Tensor { const rotary_dim: u32 = opts.scaling.partialRotaryDim(head_dim); stdx.debug.assert(rotary_dim > 0 and @mod(rotary_dim, 2) == 0, "partial rope expects a even head dim (.hd), got {d}", .{rotary_dim}); - const x_real, const x_imag, const x_pass = zml.nn.splitRealImgPass(x, opts.layout, rotary_dim); + const x_real, const x_imag, const x_pass = splitRealImgPass(x, opts.layout, rotary_dim); const inv_freq = invFreq(head_dim, opts).withTags(.{.hd}); // compute sin and cos in f32 before downcasting to x type. @@ -644,7 +645,7 @@ pub fn rope(x: Tensor, pos_idx: ?Tensor, opts: RopeOpts) Tensor { const y_real = x_real.mul(cos).sub(x_imag.mul(sin)); const y_imag = x_real.mul(sin).add(x_imag.mul(cos)); - return zml.nn.mergeRealImgPass(y_real, y_imag, x_pass, opts.layout); + return mergeRealImgPass(y_real, y_imag, x_pass, opts.layout); } pub fn splitRealImg(x: Tensor, layout: RopeOpts.Layout) [2]Tensor { diff --git a/zml/quantization.zig b/zml/quantization.zig index 5c616080..70576db5 100644 --- a/zml/quantization.zig +++ b/zml/quantization.zig @@ -3,13 +3,13 @@ const std = @import("std"); const stdx = @import("stdx"); -const zml = @import("core.zig"); +const zml = @import("zml/core"); -const DataType = @import("dtype.zig").DataType; -const platform_mod = @import("platform.zig"); +const DataType = zml.DataType; +const platform_mod = zml.platform; const Platform = platform_mod.Platform; -const Shape = @import("shape.zig").Shape; -const Tensor = @import("tensor.zig").Tensor; +const Shape = zml.Shape; +const Tensor = zml.Tensor; pub const nvfp4_block_size = 16; pub const mx_block_size = 32; diff --git a/zml/zml.zig b/zml/zml.zig index 2649eb31..2363705d 100644 --- a/zml/zml.zig +++ b/zml/zml.zig @@ -33,15 +33,15 @@ pub const logo = @import("zml/logo"); pub const mem = core.mem; pub const Bufferized = core.Bufferized; pub const mlir = core.mlir; -pub const nn = core.nn; +pub const nn = @import("zml/nn"); pub const ops = core.ops; pub const pjrtx = core.pjrtx; pub const platform = core.platform; pub const Memory = core.Memory; pub const Platform = core.Platform; pub const Target = core.Target; -pub const quantization = core.quantization; -pub const Quantization = core.Quantization; +pub const quantization = @import("zml/quantization"); +pub const Quantization = quantization.Quantization; pub const safetensors = core.safetensors; pub const shape = core.shape; pub const Shape = core.Shape;