[ExecuTorch][WebGPU] Dynamic resize hook for slice_copy (dynamic start/end)

[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 slice_copy support a dynamic gather range so the RoPE-freqs slice [input_pos : input_pos + S] works under one dynamic graph.

Problem: the static slice handler read start via a scalar reader that throws on a SymInt and ignored end (output length baked AOT). The RoPE-freqs slice uses a SymInt input_pos for start and a live S for the range, so the static op could neither build nor resize for it.

Solution: read start/end as possibly-dynamic SymInts and add a resize hook that recomputes the gather offset and live output length each step.

• Before: start is a static scalar (SymInt throws); end ignored; output length fixed at the serialized max.
• After: start/end read via a SymInt-aware reader; a hook recomputes out[dim] = (end - start + step - 1) / step, rewrites out_meta/in_meta/params UBOs + the dispatch count, and sets the output cur_dims.

Implementation:

• Hook registered on the start/end value-ids when they are SymInts and on the input tensor always (inert until resized, so a static slice is byte-identical).
• Output/input TensorMeta rebuilt from live dims; dim/step stay static.
• Keep the uniforms alive via own_uniform_buffer so the hook can rewrite them.
• Mirrors Vulkan resize_slice_copy_node.

Constraints: fp32-only; dim/step static; numerics + layout unchanged; inert on a static graph.

NOTE (stacking): this diff sits on top of the in-review slice_copy op (D108793168); rebase onto it once that op lands on master.

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

Differential Revision: D109906092

Differential Revision: D109906092

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20581

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure

As of commit 11a2c7b with merge base 124b677 ():

NEW FAILURE - The following job has failed:

• pull / test-binary-size-linux-gcc / linux-job (gh)
  RuntimeError: Command docker exec -t 23c7fae4486743141129fd5e091b63d6c170e816e88cbf44877d6a905b0be2f4 /exec failed with exit code 1

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 32s —— View job

---

Code Review: Dynamic resize hook for slice_copy

• Read PR diff and full Slice.cpp
• Check WebGPUGraph resize API (read_symint, add_resize_hook, add_tensor_resize_hook, own_uniform_buffer, propagate_resize)
• Compare against Vulkan resize_slice_copy_node
• Identify correctness issues
• Identify modularization opportunities

Overall this is clean and matches the established WebGPU dynamic-shape pattern well (SymInt hooks → tensor hooks, idempotent recompute, own_uniform_buffer to keep UBOs writable). The hook is correctly inert on a static graph. A few items below — one I'd consider a real (latent) correctness bug, the rest are quality/modularization.

---

1. Output non-slice dims should be derived from live input dims (divergence from Vulkan)

Slice.cpp:224-226

std::vector<int64_t> od = g.cur_dims(out_id);   // <- seeds from output's *prior* dims
od[dim] = len;
g.set_cur_dims(out_id, od);

The Vulkan reference this claims to mirror derives the output shape from the input sizes each resize:

std::vector<int64_t> new_out_sizes = in_sizes;        // backends/vulkan/.../Slice.cpp:63-64
new_out_sizes.at(dim) = (end - start + step - 1) / step;

For slice_copy, every non-dim axis of the output is by definition equal to the input's. By seeding od from the output's prior cur_dims and only rewriting od[dim], this hook leaves the other axes stale. The in_meta it writes does track in_dims, so if any non-slice axis of the input is also dynamic, out_meta and in_meta disagree on those axes — a silent shape mismatch.

For the RoPE-freqs case only dim is dynamic, so it happens to work today, but it's fragile and diverges from the reference. Suggest mirroring Vulkan:

std::vector<int64_t> od = in_dims;   // already have `const auto& in_dims = g.cur_dims(in_id);`
od[dim] = len;

Fix this →

---

2. Redundant hook executions (minor; idempotent so not a bug)

The same recompute closure is registered up to three times — on start_id, end_id, and in_id. When more than one trigger is dirty in the same step (e.g. input_pos SymInt + an input resize), recompute runs 2–3× per propagate_resize. It's fully idempotent (recomputes from live state), so output is correct, just wasted wgpuQueueWriteBuffer calls. Worth a one-line comment noting the intentional redundancy, or skip if the duplicate writes are negligible.

---

3. Modularization opportunities (the explicit ask)

a. Extract shared slice geometry — highest value. The build path (Slice.cpp:100-116) and the hook (Slice.cpp:217-242) independently compute start/clamp and fill SliceParams (dim/start/step). A small helper would remove the duplication and make the build path and resize path provably consistent:

struct SliceGeom { std::vector<int64_t> out_dims; SliceParams params; };
SliceGeom compute_slice_geom(const std::vector<int64_t>& in_dims,
                             int64_t dim, int64_t start, int64_t end, int64_t step);

Both sites call it; the hook additionally writes the UBOs/dispatch. This also naturally fixes finding #1, since the helper would build out_dims from in_dims.

b. read_index / is_symint / norm_clamp are reusable across transfer ops. slice, and soon select/narrow/index_select, all need "read a possibly-dynamic index" + "normalize+clamp into [0, size]". Vulkan already centralizes this as normalize_idx. Consider hoisting these three free functions into a shared WebGPU helper header (e.g. alongside TensorMeta.h or a new transfer helper) rather than re-declaring them per op. Even within this file, read_index and read_scalar (Slice.cpp:34-61) share the Int → INT64_MAX→default branch — read_scalar could be expressed as the strict variant of read_index.

c. Bind-group / pipeline boilerplate (out of scope here). Slice.cpp:122-200 is ~80 lines of BGL + pipeline + bind-group setup that is near-identical to other WebGPU ops. Not for this PR, but a make_compute_pipeline(...) / make_bind_group(...) helper would shrink every op file substantially. Flagging as a follow-up direction.

---

Minor

• Slice.cpp:76 comment still reads end unread (out shape is AOT) — now stale, end_id is read in the hook. Update or drop.
• len == 0 path → om.numel == 0 → compute_1d_workgroup_count(..., 0, ...). Confirm a 0-element dispatch is a safe no-op (Vulkan tolerates it); a quick assertion/comment would document the intent.

No blocking issues beyond #1, which I'd recommend addressing before this lands given it directly contradicts the "Mirrors Vulkan resize_slice_copy_node" claim in the description.
• Base: origin/gh/JulianCloudNTH/73/base

[SS-JIA]

Review automatically exported from Phabricator review in Meta.