[ExecuTorch][WebGPU] Generate typed to_copy shader variants

[JCNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] Preserve Linear and Q4 embedding params across dynamic resize #21483
• [ExecuTorch][WebGPU] Generate logical and arithmetic binary shader variants #21451
• [ExecuTorch][WebGPU] Generate extrema and unary shader variants #21450
• -> [ExecuTorch][WebGPU] Generate typed to_copy shader variants #21449
• [ExecuTorch][WebGPU] Make WGSL generation rollback-safe #21448
• [ExecuTorch][WebGPU] Add Qwen3 K16 attention and fp16 KV cache #21136
• [ExecuTorch][WebGPU] Add HuggingFace rotate-half RoPE #21135
• [ExecuTorch][WebGPU] Repair dynamic SDPA routing, add attestation #21134
• [ExecuTorch][WebGPU] Select emdawnwebgpu for Emscripten builds #21133
• [ExecuTorch][WebGPU] Add Llama K16 online causal attention #21132
• [ExecuTorch][WebGPU] Replace broad QKV fusion with BK64 kernel #21131
• [ExecuTorch][WebGPU] Route exact prefill shapes to the BK64 GEMM #21130
• [ExecuTorch][WebGPU] Fuse the SwiGLU elementwise chain #21129
• [ExecuTorch][WebGPU] Coordinate dynamic dispatch routes #21128
• [ExecuTorch][WebGPU] Add certified scoped output suppression #21127

Generate typed to_copy variants from one byte-identical template

The two conversion directions duplicated the same WGSL structure and could drift independently. This moves fp32-to-int32 and int32-to-fp32 into one typed family while preserving both expanded payloads. Mirrors Vulkan backends/vulkan/runtime/graph/ops/glsl/view_convert_buffer.{glsl,yaml}.

Key changes:

• to_copy_convert.wgsl and YAML — declare the two typed variants.
• Generated headers and codegen locks — preserve symbols, workgroups, registry entries, and payload hashes.
• Structural/native tests — lock both directions and the round trip.

Host routing and dispatch remain unchanged; tests expand fixture coverage.

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

Differential Revision: D113979712

Differential Revision: D113979712

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21449

• 📄 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 f9f2cdc with merge base 8200b9e ():

NEW FAILURE - The following job has failed:

• pull / unittest / macos / macos-job (gh)
  examples/models/test/test_export.py::ExportTest::test_efficient_sam_export_to_executorch

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.

[JCNTH]

@claude Double check for need of changes, modulariziation opportunity, etc

[claude]

Claude finished @JCNTH's task in 1m 52s —— View job

---

Review: WebGPU typed to_copy shader variants

• Read the full diff
• Review WGSL template + YAML
• Review generated headers / codegen locks
• Review tests
• Assess modularization opportunities

Verdict: solid, ship-ready. The refactor collapses two byte-identical shaders into one typed template and mirrors the established binary_op/boolean_op multi-variant convention almost exactly. The generated headers keep their distinct symbols, workgroup constants, and payload hashes unchanged, so host routing in ToCopy.cpp needs no edits. Nothing blocking; findings below are minor / optional.

Correctness — looks right

• The template collapse hinges on output[idx] = ${OUT_TYPE}(input[idx]). Both directions are pure scalar casts (i32(f32), f32(i32)), so a single expression covers both. ✅
• The wgsl-sha256 in each header is unchanged from the pre-refactor files — the expansion is byte-identical, which is exactly what the new test_to_copy_convert_template_roundtrip_byte_identical locks. Good regression guard.
• The float fixture (-8.75, -3.0, -1.5, -0.25, ...) deliberately includes fractional magnitudes, so the round-trip exercises truncation-toward-zero through int32 rather than trivially-invertible integers. Nice — this is the kind of value that would catch a wrong cast op. ✅

Minor: YAML license-header inconsistency

to_copy_convert.yaml:1-6 adds a BSD license header, but every sibling template YAML (binary_op.yaml, boolean_op.yaml, rms_norm.yaml, softmax.yaml) has none — they start directly with the template name. Adding headers is arguably better, but as-is it's the odd one out. Either drop it to match convention or (better) treat it as the start of a convention — just be consistent within the PR's scope.

Observation: float→int is only numerically exercised via the round-trip

In cases.py, _to_copy_factory maps only int_to_float and float_roundtrip (the suite's golden_dtype="float32" can't hold a pure int output). So the standalone float→int shader is only validated numerically through the round-trip module, not in isolation. Delegation of the float→int direction is covered directly in test_to_copy.py::test_float_to_int_delegates, so this is coverage-shaping, not a gap — just worth being aware that a float→int-only numeric regression would surface via the round-trip case rather than a dedicated one.

Modularization: _delegated_cast_dtypes is the one intricate spot

test_to_copy.py:88-102 does two-phase logic — return [] if an undelegated _to_dim_order_copy sits at the top level, otherwise harvest dtype kwargs from inside lowered submodules — using "_to_dim_order_copy" in str(getattr(node, "target", "")) string matching in both places. It works and is pragmatic, but:

• The substring match on str(target) is a bit fragile if the op's qualified name ever shifts. A structural check (comparing against exir_ops.edge.dim_order_ops._to_dim_order_copy.default, or the torch.ops.aten.to.dtype target already used in _prepartition_cast_dtypes) would be sturdier.
• The two branches repeat the same node-scan predicate; a small _has_dim_order_copy(nodes) helper would remove the duplication and make the "delegated vs not" intent explicit.

Both are readability nits, not correctness issues — reasonable to leave given the surrounding test style.

Nice touches

• Splitting _export into _lower (returns ep, edge) + _export (calls to_executorch) is clean and lets test_roundtrip_keeps_both_casts_in_delegate assert on the pre-partition graph, delegated graph, and executorch program independently.
• Workgroup-boundary sizes (63, 64, 65, 257) around wg_size=64 are the right stress points.

I did not run the test suite (write ops are disabled on this ghstack branch and pytest needs approval here), but the roundtrip/registry lock tests are self-checking against the committed hashes, so CI on the PR will confirm.
· branch gh/JCNTH/199/head