[ExecuTorch][WebGPU] Add sigmoid op (aten.sigmoid.default)

[JulianCloudNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] aten.index.Tensor test suite (export + native golden) #20465
• [ExecuTorch][WebGPU] Add aten.index.Tensor (1D-self gather) #20464
• [ExecuTorch][WebGPU] Add clone op (aten.clone.default) #20463
• [ExecuTorch][WebGPU] Flatten landed-op test dirs to test/ops/test_<op>.py #20435
• [ExecuTorch][WebGPU] cat op test suite (cases.py op-test framework) #20399
• [ExecuTorch][WebGPU] Add cat op + ValueList graph support (aten.cat.default) #20398
• [ExecuTorch][WebGPU] permute_copy op test suite (cases.py op-test framework) #20397
• [ExecuTorch][WebGPU] Add permute_copy + IntList graph support (aten.permute_copy.default) #20396
• [ExecuTorch][WebGPU] slice_copy op test suite (cases.py op-test framework) #20395
• [ExecuTorch][WebGPU] Add slice_copy op (aten.slice_copy.Tensor) #20394
• [ExecuTorch][WebGPU] squeeze_copy + unsqueeze_copy test suites (cases.py op-test framework) #20393
• [ExecuTorch][WebGPU] Add squeeze_copy + unsqueeze_copy (flat copies) #20392
• [ExecuTorch][WebGPU] sigmoid op test suite (cases.py op-test framework) #20391
• -> [ExecuTorch][WebGPU] Add sigmoid op (aten.sigmoid.default) #20390
• [ExecuTorch][WebGPU] select_copy op test suite (cases.py op-test framework) #20363
• [ExecuTorch][WebGPU] Add select_copy op (aten.select_copy.int) #20362
• [ExecuTorch][WebGPU] view_copy op test suite (cases.py op-test framework) #20361
• [ExecuTorch][WebGPU] Add view_copy op (aten.view_copy.default) #20360
• [ExecuTorch][WebGPU] mul op test suite (cases.py op-test framework) #20359

Adds aten.sigmoid.default to the WebGPU delegate: element-wise 1/(1+exp(-x)) over a flat fp32 buffer. On the Llama critical path (F.silu -> sigmoid + mul).

Composition (single dispatch):

• sigmoid/UnaryOp.cpp — binds input (storage, read-only) + output (storage) + a Params{num_elements} uniform, 1D-dispatches over num_elements with override wg_size (clamped to the device limit); mirrors the add op (uniform mapped-at-creation, released after the bind group).
• sigmoid/sigmoid.wgsl — guards idx >= num_elements and writes the logistic of each element.
  @exported-using-ghexport

Differential Revision: D108793157

Differential Revision: D108793157

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20390

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI runner label rename: rebase PRs using old linux.rocm.gpu.gfx950.* labels

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[JulianCloudNTH]

@claude review

[claude]

Claude finished @JulianCloudNTH's task in 2m 3s —— View job

---

Code Review: WebGPU sigmoid op

• Read PR diff
• Compare against sibling ops (add/mul)
• Review correctness, resource management, edge cases
• Post review

Clean, well-scoped op that faithfully mirrors the add/mul pattern. WGSL is correct (1/(1+exp(-x)) with the idx >= num_elements guard), the generated header's wgsl-sha256 matches sigmoid.wgsl, and resource lifetimes match the established convention (shader/bgl/pipeline-layout released after the bind group is built; uniform buffer dropped after the bind group takes a ref). Nice cleanup using the shared utils::make_uniform helper instead of inlining the mapped-at-creation dance like add does.

Main issue — op won't actually be delegated / no test

aten.sigmoid.default is registered in the runtime via WEBGPU_REGISTER_OP, but it is not added to WEBGPU_SUPPORTED_OPS in backends/webgpu/test/tester.py:21, and there's no test/ops/sigmoid/test_sigmoid.py. Every other op in the runtime (add, mul, view_copy, select_copy, rms_norm) is present in that allowlist and has a matching test. As-is, nothing exercises this code path.

If a follow-up PR in the ghstack adds the partitioner allowlist entry + test, this is fine — please confirm. Otherwise this op is dead code in this PR. Note also the allowlist gates the Vulkan partitioner (WebGPU consumes the VK00 serialization), so delegation also depends on the Vulkan partitioner lowering sigmoid into the subgraph.

Add sigmoid test + allowlist entry →

Minor

• Dtype guard is approximate. The nbytes % sizeof(float) != 0 check (UnaryOp.cpp:35) rejects misaligned buffers but doesn't truly verify fp32 — any 4-byte-divisible dtype (e.g. int32) slips through and gets logistic'd as float. add doesn't validate at all, so this is already stricter than the baseline; just flagging that the error message ("expected fp32") overstates what's checked. Fine to leave.
• Workgroup size differs from add (64 vs 256). Not wrong — clamp_workgroup_size handles device limits either way — but if there's a rationale for the smaller size on the unary path it'd be worth a one-line comment; otherwise consider matching add for consistency.
• SigmoidParams carries a single num_elements + _pad[3]. Correct for the 16-byte uniform alignment requirement and matches the WGSL Params struct. No change needed.

Nits

• UnaryOp.cpp uses the modern nested namespace executorch::backends::webgpu while add/BinaryOp.cpp still uses the old three-line form — your version is the better style, just inconsistent across the directory.
• Directory is sigmoid/ but the file is UnaryOp.cpp (generic). Reasonable if more unary ops will share this scaffold later; otherwise Sigmoid.cpp would match the select/Select.cpp convention.

No correctness or memory-management bugs found. The one thing to resolve before merge is whether the partitioner allowlist + test land here or later in the stack.