Repository navigation
Add dynamic shapes to ExecuTorch static cache export tests #40260
Description
Feature request
Add dynamic shapes to the ExecuTorch text_export_static_cache tests on causalLM models such as Qwen.
Then modify the test generation function to prefill all at once instead of sequentially.
Motivation
Tests the export more accurately since production use cases will almost always export with dynamic cache to prefill all at once during inference.
Your contribution
I submit the PR @jackzhxng
Activity
ThankaBharathi commentedon Aug 28, 2025
Hi maintainers 👋, I’d like to work on this feature request.
Could you please assign it to me?
My approach
Add dynamic shapes to
ExecuTorch text_export_static_cachetests for causal LM models (e.g., Qwen).Update the test generation to prefill all at once instead of sequentially.
Add/extend tests to validate correctness on dynamic cache + batch prefill.
Verify on Qwen configs and ensure CI passes.
Thanks!
cc @jackzhxng
jackzhxng commentedon Aug 28, 2025
Feel free to take it!
ArthurZucker commentedon Oct 20, 2026
On
main, export goes throughtransformers.exporters(https://huggingface.co/docs/transformers/main/en/exporters) (#41992), which supports dynamic shapes; extending it is documented at https://huggingface.co/docs/transformers/main/en/exporters_extend. The legacy helpers inintegrations/executorch.pyare not where new export work should go.