[ExecuTorch][WebGPU] Add cat op + ValueList graph support (aten.cat.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.cat.default to the WebGPU delegate as an index-math scatter, and the ValueList graph value type it needs to read its tensor-list argument.

Composition (one dispatch per input):

• runtime/WebGPUGraph.{h,cpp} — adds ValueType::ValueList backed by std::vector<std::vector<int>> value_lists_ + get_value_list(int); build() deserializes vkgraph::GraphTypes::ValueList via value_as_ValueList()->items(); mirrors the existing scalar value plumbing.
• runtime/ops/cat/Cat.cpp — reads the input-tensor list via get_value_list, reads the static Int dim, validates each input rank + non-concat dims against the output; builds one shared out_meta TensorMeta uniform + a shared bind-group/pipeline layout; per input builds a fresh pipeline + bind group with in_meta + CatParams{concat_dim, off_k}, dispatches over compute_1d_workgroup_count(in.numel); releases per-input uniforms after each bind group, the shared out_meta after the loop.
• runtime/ops/cat/cat.wgsl — delinearizes the input index over the input strides, shifts the concat-dim coordinate by the host-computed off_k (running sum of prior input sizes), relinearizes over the output strides to scatter output[out] = input[in].
• The N disjoint-slab writes share one output buffer across separate dispatches and rely on the existing per-pass execute() ordering.
  @exported-using-ghexport

Differential Revision: D108793165

Differential Revision: D108793165

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20398

• 📄 Preview Python docs built from this PR

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

❌ 3 New Failures, 3 Unrelated Failures

As of commit 7e09096 with merge base e03f777 ():

NEW FAILURES - The following jobs have failed:

• pull / test-multimodal-linux (gemma3-4b) / linux-job (gh)
  RuntimeError: Command docker exec -t c052a3b6f530ef89ee162e6e382a5ea9afbf4ae23a1c58d984b77d438fd65e0e /exec failed with exit code 139
• pull / test-qnn-python-imports-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 5ec75bbddbef8bdea100ad4c48cfb43bcbfb457a62953c6a1308a82b65ef99a8 /exec failed with exit code 92
• pull / unittest-buck / linux / linux-job (gh)
  RuntimeError: Command docker exec -t be5b3d8c2563da1979d5beccaf2129c3a78128294623e026edf5cf75cbac324b /exec failed with exit code 3

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / test-qnn-testsuite-linux / test-backend-linux (qnn, models) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

BROKEN TRUNK - The following jobs failed but was present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh) (trunk failure)
• pull / unittest-buck / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 3

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 7m 26s —— View job

---

Claude Code Review

• Read changed files
• Analyze correctness, safety, performance
• Post review

I reviewed the diff against origin/gh/JulianCloudNTH/47/base. The implementation is careful and idiomatic — it closely mirrors Slice.cpp/Permute.cpp, the leak-safety ordering is thoughtful (all validation + compute_1d_workgroup_count happen before any GPU allocation), and the resource ownership comments are accurate. A few things worth addressing:

---

1. No test coverage, and aten.cat.default isn't in the partitioner allowlist (main gap)

Every other op in this backend (permute, slice, select, …) has a test/ops/<op>/test_<op>.py module and an entry in WEBGPU_SUPPORTED_OPS in backends/webgpu/test/tester.py:21-32. cat has neither:

# tester.py — cat is missing
WEBGPU_SUPPORTED_OPS = [
    ...
    exir_ops.edge.aten.permute_copy.default,
    # exir_ops.edge.aten.cat.default,   <-- not here
]

Without the allowlist entry the VulkanPartitioner won't delegate cat, so the new kernel is effectively unreachable through the standard op-test path and gets zero CI coverage in this PR. If the test lands later in the stack, a one-line note here would help; otherwise consider adding aten.cat.default to the allowlist + a test/ops/cat/test_cat.py (multi-input, negative dim, concat on dims 0/1/2/3, 1-input edge case). Fix this →

2. A fresh WGPUComputePipeline is created per input even though it's identical (Cat.cpp:135-148)

The shader, pipeline layout, and wg_size constant are invariant across inputs — only the bind group changes. The per-input pipeline is correct given the destructor releases d.pipeline for each dispatch (WebGPUGraph.cpp:175-181), so a shared handle would double-free — and your // fresh pipeline+bind group per input (no double-free) comment shows you know this. But you can avoid the redundant compiles by creating the pipeline once and calling wgpuComputePipelineAddRef for each extra dispatch (matching the destructor's release count). This is build-time only (not the runtime hot path), so it's low priority — fine to leave for a follow-up, but worth a comment if so.

3. Minor: boilerplate duplication with Permute.cpp

The bind-group-layout + pipeline-layout setup (5 entries: in / out-rw / out_meta / in_meta / params) is byte-for-byte identical to Permute.cpp:119-145. Not blocking given the repo's "no trivial helpers" guidance, but this is substantial (~30 LOC) shared boilerplate across a growing set of in→out + 3 uniforms ops; a single make_io_meta_params_layout(device) helper would pay off as more ops land.

---

Things I checked that look correct

• WGSL index math: delinearize over in_meta.strides, shift concat-dim coord by off_k, relinearize over out_meta.strides — matches the gather-style pattern in slice.wgsl, just inverted to a scatter. ✓
• Division-by-zero in the WGSL loop: rem / in_meta.strides[d] only runs for d < in_meta.ndim, where strides are nonzero for non-empty tensors; empty inputs (numel==0) hit the early return and get a 0-count dispatch, so the divide never executes. ✓
• Uniform sizing: Params is 8 bytes in WGSL but bound with sizeof(CatParams)==16 — a larger binding than the struct is legal, and the static_assert locks the C++ side. ✓
• Disjoint-slab writes: N dispatches write non-overlapping output regions and never read each other, so the reliance on execute() ordering is sound; no barrier needed. ✓
• Validation: negative dim normalization, rank match, non-concat dim-size match, and concat_sum == out.dims[dim] are all checked before dispatch. Nice defensive set. ✓
• get_value_list / ValueList deserialization mirrors the existing IntList plumbing correctly. ✓

One thing I couldn't run in this environment: python backends/webgpu/scripts/gen_wgsl_headers.py --check to confirm cat_wgsl.h matches cat.wgsl (the committed wgsl-sha256 looks present and CI should gate this) — just make sure that check is green.
· branch gh/JulianCloudNTH/47/head