{{ message }}
WebGPU op-test framework + mul op manual merge#20389
Merged
Merged
Conversation
… .pte+golden -> gtest driver) Pull Request resolved: #20339 A manifest-driven op-test framework for the WebGPU backend, mirroring Vulkan's `op_tests/cases.py` (declarative per-op suites) but with a torch-computed golden loaded in C++, since the native test binary has no ATen. An op fits when it is stateless and expressible as one `module(inputs) -> golden` forward; stateful KV-cache ops (`sdpa`, `update_cache`) stay hand-written. Lands first as the shared foundation for the following op test-diffs — adding an op's test becomes one `cases.py` entry. Composition: - `test_suite.py` — schema (`WebGPUTestSuite`/`Case`/`InputSpec`) + a `register_op_test` decorator; per-case `required`/`heavy`/`golden_fn`, per-suite `golden_dtype`. - `cases.py` — the declarative suites; registers the landed `add` + `rms_norm`; later ops append one entry each. - `generate_op_tests.py` — per case: export via `VulkanPartitioner` to `.pte`, compute the fp64 torch golden (dual-oracle gate), serialize inputs+golden as fp32, emit `manifest.json`. - `op_test_driver.cpp` + `driver_util.{h,cpp}` — generic gtest driver: one test per manifest entry, runs forward on-device, abs/rel + shape + reconciliation checks. - `CMakeLists`/`ci.sh` — `webgpu_op_test` + device-free `webgpu_op_test_util_test`, wired into the Dawn(Tint)+SwiftShader CI. ghstack-source-id: 394712836 @exported-using-ghexport Differential Revision: [D108816389](https://our.internmc.facebook.com/intern/diff/D108816389/)
…-test framework Pull Request resolved: #20357 `add` and `rms_norm` already have declarative suites in the `cases.py` framework, so their standalone tests are redundant. Remove them, leaving the framework as the single home for these stateless single-forward ops (mirroring Vulkan). `update_cache` and `sdpa` stay hand-written — stateful KV-cache replay can't be a single `module -> golden` forward. Removed: - `test_single_add`/`test_chained_add` + their `main()` plumbing from `test_webgpu_native.cpp`. - the standalone `test/native/test_rms_norm.cpp` binary + its CMake target. - the add/rms_norm export+run wiring in `test_webgpu_native_ci.sh` + `test_build_webgpu.sh`. - `webgpu_native_test` re-gated on the executorch wheel being importable (was the add `.pte`); it still hosts the quantized_linear/SDPA/update_cache/symint sweeps. ghstack-source-id: 394625048 @exported-using-ghexport Differential Revision: [D108821384](https://our.internmc.facebook.com/intern/diff/D108821384/)
Pull Request resolved: #20358 Adds `aten.mul.Tensor` to the WebGPU delegate with full PyTorch broadcast, plus the shared `runtime/ops/TensorMeta.h` per-tensor uniform that broadcast ops reuse. Mul is on the Llama critical path — `F.silu` decomposes to `sigmoid` + `mul`, and SwiGLU multiplies two same-shape activations (the fast path). Composition (single dispatch): - `TensorMeta.h` (NEW) — 48-byte std140 `{ndim, numel, sizes[4], strides[4]}` UBO mirroring Vulkan's per-tensor `BufferMetadata`; `fill_tensor_meta_broadcast` right-aligns operand dims (rank>4 throws); `static_assert(sizeof==48)`. - `mul/BinaryOp.cpp` — builds 3 `TensorMeta` UBOs (out/in1/in2 at bindings 3/4/5), guards fp32 + rank≤4, 1D-dispatches over `compute_1d_workgroup_count(numel)`, releases all uniforms after the bind group. - `mul/binary_mul.wgsl` — same-shape fast path + a broadcast path (delinearize output index, clamp each input coord per-dim to size-1, relinearize on input strides). - `WebGPUUtils.h` — adds the shared `utils::make_uniform` helper (first use). ghstack-source-id: 394848336 @exported-using-ghexport Differential Revision: [D108793167](https://our.internmc.facebook.com/intern/diff/D108793167/)
ghost
requested review from
meta-cla
Bot
added
the
CLA Signed
SS-JIA
approved these changes
Jun 18, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Manual co-merge of the stacked ghstack PRs #20339, #20357, and #20358 — all reviewed and accepted, but not picked up by auto-merge. Combining them into one PR so they can be merged directly.
This contains three commits:
Op-test codegen framework ([ExecuTorch][WebGPU] Op-test codegen framework (cases.py -> generated .pte+golden -> gtest driver) #20339) — a declarative
cases.py→ generated.pte+ golden → gtest driver for WebGPU op tests, mirroring the Vulkanop_testssetup. Each op declares its shapes/configs incases.py; the generator exports a.pteper case and compares the on-GPU result against an fp64 torch golden on Dawn.Consolidate landed-op tests into the framework ([ExecuTorch][WebGPU] Consolidate landed-op tests into the cases.py op-test framework #20357) — migrates the existing
addandrms_normtests off the C++ monolith into thecases.pyframework.update_cache/sdpaare kept standalone (stateful —they don't fit the single-forward model).
Add mul op with full broadcast ([ExecuTorch][WebGPU] Add mul op with full broadcast (aten.mul.Tensor) #20358) —
aten.mul.Tensorfor the WebGPU delegate with full broadcasting, plus the sharedTensorMetabroadcast-uniform infrastructure. On the Llama critical path (SwiGLU).