[ExecuTorch][WebGPU] Dynamic resize hooks for add and mul

[ghost]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] Convert remaining native tests to GTest #20706
• [ExecuTorch][WebGPU] Use requiredFeatures instance API on native + emscripten Dawn #20652
• [ExecuTorch][WebGPU] 2D-fold mul + permute dispatch (lift 65535 1D cap) #20651
• [ExecuTorch][WebGPU] 2D compute dispatch tests — prefill golden + fold unit test #20584
• [ExecuTorch][WebGPU] 2D compute dispatch — lift the 65535 per-dim cap (prefill path) #20583
• [ExecuTorch][WebGPU] Dynamic-shape integration test (allocate-at-max + per-op resize) #20582
• [ExecuTorch][WebGPU] Dynamic resize hook for slice_copy (dynamic start/end) #20581
• [ExecuTorch][WebGPU] Dynamic resize hook for SDPA (live seq-len S) #20580
• [ExecuTorch][WebGPU] Dynamic resize hook for view_copy #20579
• [ExecuTorch][WebGPU] Dynamic resize hooks for sigmoid and select_copy #20578
• -> [ExecuTorch][WebGPU] Dynamic resize hooks for add and mul #20577
• [ExecuTorch][WebGPU] Dynamic resize hook for linear_q4gsw #20576
• [ExecuTorch][WebGPU] Dynamic resize hooks for rms_norm, embedding, rope #20575
• [ExecuTorch][WebGPU] Dynamic tensor-shape resize engine core #20574
• [ExecuTorch][WebGPU] SymInt arithmetic ops (add/sub/mul/floordiv) for dynamic shapes #20573

Make the elementwise add and mul ops serve any live shape from one graph.

Problem: aten.add.Tensor and aten.mul.Tensor baked their element count + param UBO(s) + output shape at build() for the max shape. On a dynamic-shape graph at a smaller live shape they would over-dispatch and leave the output sized at the max.

Solution:

• Before: one fixed dispatch sized for the build-time shape.
• After: each registers a resize hook on BOTH operands (the dynamic one may be either operand by arg order). The hook recomputes the live element count, rewrites the param UBO(s), updates the dispatch workgroup_count_x, and sets the output cur_dims. Inert until an operand is resized.

Implementation:

• add: out follows the larger operand (robust when one input is a static residual and the other is the dynamic-S tensor); rewrites AddParams.
• mul: recomputes the broadcast output shape and rebuilds all three TensorMeta UBOs via fill_tensor_meta_broadcast.
• Each keeps its uniform buffer(s) alive via own_uniform_buffer instead of releasing at build.
• Mirrors Vulkan per-op resize_*_node (recompute sizes + dispatch each execute).

Constraints: Behavior-neutral on static graphs (the hook fires only when an operand's live shape differs from the max). No kernel/WGSL/numerics change.

Co-authored-with: Claude Code.
@exported-using-ghexport

Differential Revision: D109906093

Differential Revision: D109906093

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20577

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 1 Unrelated Failure

As of commit e85fad2 with merge base 124b677 ():

NEW FAILURES - The following jobs have failed:

• pull / test-arm-backend-no-driver (test_pytest_ops_tosa) / linux-job (gh)
  RuntimeError: Command docker exec -t e59c7216df639173da38821b24cd9bad10c0e1147ad33a6798d4898ce56ad89b /exec failed with exit code 1
• pull / test-binary-size-linux-gcc / linux-job (gh)
  RuntimeError: Command docker exec -t 9e6f8bd0ebc186487ebbcf1e7c9169abdf1fe3a06eee98afa65336c87f601bb6 /exec failed with exit code 1

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)

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.

[ghost]

@claude review and check for any areas or opportunities for modularization

[claude]

Claude finished @JulianCloudNTH's task in 2m 16s —— View job

---

Review: Dynamic resize hooks for add and mul

• Read both changed files in full
• Inspect WebGPUGraph resize-hook API + shared helpers (TensorMeta.h, WebGPUUtils.h)
• Compare add vs mul resize logic for correctness
• Identify modularization opportunities

The change is correct and well-scoped. The hooks key off cur_dims and set_cur_dims only re-dirties on an actual change, so the static path stays inert exactly as advertised, and propagate_resize converges. Ownership handoff (own_uniform_buffer instead of wgpuBufferRelease) is right — the dtor frees them, and the bind group still holds its own ref. The pattern matches the existing rms_norm resize hook, so it reads consistently with the rest of the backend.

A few correctness notes and the modularization opportunities you asked about.

Correctness / behavior notes

1. Both operands dirty → hook fires twice in one pass. Both add and mul register the same closure on in1_id and in2_id. In propagate_resize (WebGPUGraph.cpp:205), a single pass iterates all hooks against the processing set, so if both operands are resized in the same pass the closure runs twice (two redundant wgpuQueueWriteBuffer calls + workgroup recompute). It's idempotent so results are correct — just wasted work. In practice usually only one operand is dynamic, so this rarely triggers, but worth a one-line comment acknowledging it.

2. add output = larger-by-numel operand is a heuristic, not broadcasting. The add kernel is pure elementwise over num_elements with no per-tensor stride/broadcast metadata, so add cannot broadcast mismatched shapes (e.g. [3,1] + [1,4]) — only the "one operand is the full shape, the other is equal or a scalar-ish subset" case works. Picking n2 > n1 ? d2 : d1 preserves the static build's behavior, so this PR doesn't regress anything, but it inherits that latent limitation. Fine to leave; flagging so it's a conscious choice.

3. Uninitialized WebGPUTensor temporaries in the mul hook (mul/BinaryOp.cpp:187-190). ta/tb/to are default-constructed and only .dims is set; fill_tensor_meta_broadcast reads only .dims, so this is safe today — but it's an implicit coupling to that function's internals. See the overload suggestion below, which removes the temporaries entirely.

Modularization opportunities

These two *_impl functions now share three reusable primitives that don't currently exist as helpers:

(a) Broadcast output-shape computation. The right-alignment loop in mul/BinaryOp.cpp:178-184 is a standalone PyTorch-broadcast rule with no dependency on WebGPU types. It belongs next to fill_tensor_meta_broadcast in TensorMeta.h:

// TensorMeta.h
inline std::vector<int64_t> broadcast_out_dims(
    const std::vector<int64_t>& a, const std::vector<int64_t>& b) {
  const size_t r = std::max(a.size(), b.size());
  std::vector<int64_t> out(r, 1);
  for (size_t i = 0; i < r; i++) {
    const int64_t av = (i + a.size() < r) ? 1 : a[i - (r - a.size())];
    const int64_t bv = (i + b.size() < r) ? 1 : b[i - (r - b.size())];
    out[i] = av > bv ? av : bv;
  }
  return out;
}

Fix this →

(b) A fill_tensor_meta_broadcast overload taking const std::vector<int64_t>& so the mul hook stops fabricating WebGPUTensor temporaries:

inline void fill_tensor_meta_broadcast(
    const std::vector<int64_t>& dims, uint32_t out_ndim, TensorMeta* m);
// existing WebGPUTensor overload forwards to it via t.dims

This removes ta/tb/to and note (3) above. Fix this →

(c) "Register the same hook on both operands." Both ops do the identical two-line dance:

graph.add_tensor_resize_hook(in1_id, fn);
graph.add_tensor_resize_hook(in2_id, fn);

A small overload add_tensor_resize_hook(std::initializer_list<int> ids, fn) on WebGPUGraph would document intent and also let propagate_resize dedup so the hook fires once even when both operands are dirty (fixes note 1). Lower priority than (a)/(b).

On a full binary-op base class: I'd not go further than (a)–(b) right now. add (single AddParams UBO, numel-max, no broadcast metadata) and mul (three TensorMeta UBOs, full broadcast) differ enough in their bodies that a shared binary_op_impl template would need so many hooks it'd be harder to read than the two explicit functions — which cuts against the repo's "choose simpler, more concise" guidance. Extracting the genuinely shared data helpers (a/b) is the right altitude; collapsing the control flow is not.

Summary

No blocking issues — correct and behavior-neutral on static graphs. Recommended: (a) and (b) to remove the duplicated broadcast math and the fabricated tensor temporaries; (c) optional. Notes 1–2 are worth a comment but need no code change.
· branch gh/JulianCloudNTH/69/head

[SS-JIA]

Review automatically exported from Phabricator review in Meta.