Vulkan: do not partition ops whose symbolic arguments the runtime reads at build time

[msluszniak]

Summary

Several Vulkan op implementations read a scalar or list argument with extract_scalar() / get_int_list() when the node is built, and bake the result into the dispatch. When that argument is symbolic (derived from a dynamic dimension) it is never refreshed. Depending on the op this either aborts at execute time or, worse, silently computes against a stale value.

This PR declines the affected nodes via are_node_inputs_supported_fn (following the existing _check_copy_is_noop precedent) so they fall back instead. Making these ops genuinely support symbolic arguments means plumbing the values through as symints and refreshing them on resize, which is left for a follow-up.

1. clamp with a symbolic bound — silent wrong results. get_val_or_inf() in UnaryOp.cpp:

float get_val_or_inf(ComputeGraph& graph, const ValueRef& val, bool max) {
  if (!graph.val_is_none(val)) {
    return graph.extract_scalar<float>(val);   // read once, at build time

Nothing raises. This is most damaging when clamp is applied to index tensors, where a wrong limit silently reorders or drops data downstream. hardtanh, hardshrink and leaky_relu use the same path and take the same guard.

Minimal repro — clamp(idx, 0, L-1) feeding index_select, exported with L dynamic and executed at L's upper bound, so the runtime shapes are identical to the static export:

	result
static export	cosine 1.000000 vs CPU
dynamic export	all 2048 output values are zero
with this fix	cosine 1.000000

2. constant_pad_nd with a symbolic pad — abort. A symbolic pad list is serialized as a VALUELIST, and Pad.cpp reads it with get_int_list():

Exception raised from toIntList at .../graph/containers/Value.h:272:
(isIntList()) is false! Expected value to have type IntList, got VALUELIST instead.

Most size-taking ops (View.cpp, Expand.cpp, Repeat.cpp, Squeeze.cpp, ...) already use the symint-aware extract_int_or_symint_list(). Switching Pad.cpp to it would not be sufficient on its own: add_constant_pad_nd_node() also bakes the amounts into a params buffer at build time, so the op would trade an abort for a wrong result.

3. _native_batch_norm_legit_no_training on a non-4d input — abort.

Exception raised from add_native_batch_norm_node at .../ops/impl/BatchNorm.cpp:65:
(in_sizes.size() == 4) is false! BatchNorm only support 4d tensor

The registration carries no rank constraint, so any conv1d model (rank-3 activations) partitions a batch norm the runtime refuses. This affects 1-D audio models generally.

Test plan

Reproduced with the Supertonic TTS model (Supertone/supertonic-3) on a Galaxy S26 Ultra (Adreno 840), lowering with VulkanPartitioner and running executor_runner. Reference outputs are from CPU and are reproduced by the XNNPACK delegate at cosine 1.000000, confirming the inputs and references are sound.

sub-model	before	after
vocoder	SIGABRT (BatchNorm only support 4d), then cosine 0.016757	0.999977
vector_estimator	cosine 0.994432	0.999994
text_encoder	SIGABRT (VALUELIST)	runs

These changes are partitioner-side only, so no runtime rebuild is needed to reproduce either the failures or the fixes.

Note: text_encoder still returns wrong output after these fixes (cosine 0.538, improved from 0.406). Its statistics match the reference closely (matching value multiset, std ratio 1.009) so it looks like a further placement/layout issue rather than bad arithmetic. I have ruled out slice_copy, arange and index.Tensor as the cause by op-exclusion bisection and am still narrowing it; that will be reported separately.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22399

• 📄 Preview Python docs built from this PR

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

⚠️ 16 Awaiting Approval

As of commit 6da6af3 with merge base 5428092 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

• Build documentation (gh)
• Build Presets (gh)
• Cadence Build & Test (gh)
• Lint (gh)
• periodic (gh)
• pull (gh)
• Test ARM Backend (gh)
• Test CoreML Backend (gh)
• Test Cortex-M Backend (gh)
• Test NXP Backend (gh)
• Test QNN Backend (gh)
• Test QNN Windows MSVC build (gh)
• Test Vulkan Backend (gh)
• Test Vulkan Backend (specialized runners) (gh)
• Test WebGPU Backend (gh)
• Test XNNPACK Backend (gh)

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.