Vulkan: size texture-vs-buffer choice by the bound, not the trace hint

[msluszniak]

Summary

filter_invalid_reprs() decides whether a tensor can live in a texture by comparing required_image_extents(tensor_val.shape) against the device limits. For a symbolic dimension that comparison resolves at the dim's hint -- the size of the example the model happened to be traced with -- rather than the maximum its exported range allows.

A model traced with a small example therefore picks a texture that is only valid at that size, and silently produces wrong results once it runs larger. Nothing raises.

This matters in practice because exporters conventionally trace with a small example input while declaring a large dynamic max, so the failing combination is the common one.

Reproduction

The Supertonic TTS text encoder (Supertone/supertonic-3), dynamic T in [8, 512], always executed at T=512 on a Galaxy S26 Ultra (Adreno 840). Only the trace-time example length differs, and the edge graphs are identical -- 552 call_function nodes, same ops, same literal args, same symbolic shapes, zero differing nodes:

traced at	cosine vs CPU	emitted .pte
T=64	0.537966	18147970 bytes
T=128	0.537966	18147970 bytes (identical file)
T=192	0.999414	18149890 bytes
T=512	0.999414	18149890 bytes (identical file)

The emitted binaries cluster exactly on the correctness boundary, which is what identifies this as a lowering decision rather than a bad kernel. With this change the T=64 export emits the 18149890-byte artifact and scores 0.999414.

Evaluating the bound also handles the unbounded case: a dim with no finite upper bound cannot be shown to fit any texture, so it falls back to buffer storage.

Test plan

Measured on Supertonic sub-models, against a CPU reference that the XNNPACK delegate reproduces at cosine 1.000000:

sub-model	before	after
text_encoder	0.406303	0.999414
vector_estimator	0.994432	0.999994
vocoder	0.016757	0.999977

Partitioner/lowering-side only, so no runtime rebuild is needed to reproduce the failure or the fix.

Note: the vocoder and vector_estimator numbers above also require #22399 (clamp with a symbolic bound), which is an instance of the same underlying theme -- a value derived from a dynamic dimension being read once at build time. This PR is independent of that one and touches a different file.

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22401

• 📄 Preview Python docs built from this PR

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

⚠️ 15 Awaiting Approval

As of commit 8255516 with merge base 5428092 ():

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

• Build documentation (gh)
• Build Presets (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.