add fast test for whisper

[xadupre]

Describe your changes

Add fast test for whisper encoder and decoder in a dedicated test/cli/test_cli_whisper_smoke.py file. Refactored transformers imports to be local imports inside the functions where they are needed (_save_local_tiny_llama and _save_local_tiny_whisper) instead of at the module level. Switched WhisperConfig(...) keyword-argument style to WhisperConfig.from_dict({...}) to fix PYLINT unexpected-keyword-arg errors, matching the existing LlamaConfig.from_dict pattern.

Also fixed a bug in olive/passes/onnx/model_builder.py where onnxruntime-genai's save_model for multi-component models like Whisper (encoder + decoder) prematurely deletes the shared cache_dir between component saves. The encoder's save_model cleanup removes the cache_dir when it is empty (no HF downloads for local models), causing the decoder's save_model to fail with FileNotFoundError. The fix creates a temporary marker file (.olive_build_<pid>) in HF_HUB_CACHE before calling create_model so the directory is never considered empty and never deleted prematurely; the marker is removed in a finally block after create_model completes.

precision is a configurable class attribute on TestCliWhisperSmoke (defaulting to "fp32") and a parameter of _run_whisper_capture_onnx_flow, making it easy to override in subclasses or from the command line.

All Whisper-specific code (DEFAULT_WHISPER_MODEL_IDS, _save_local_tiny_whisper, _run_whisper_capture_onnx_flow, TestCliWhisperSmoke) lives in test/cli/test_cli_whisper_smoke.py, separate from the existing LLaMA smoke tests in test/cli/test_cli_test_model_smoke.py. The CI workflow (test-model-fast.yml) is updated to run both files.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

(Optional) Issue link

[xadupre]

@copilot SUBFAILED(model_id='openai/whisper-tiny') test/cli/test_cli_test_model_smoke.py::TestCliWhisperSmoke::test_whisper_capture_onnx_graph - AssertionError: Expected encoder.onnx in output, got: set()
assert 'encoder.onnx' in set()
SUBFAILED(model_id='microsoft/whisper-base') test/cli/test_cli_test_model_smoke.py::TestCliWhisperSmoke::test_whisper_capture_onnx_graph - AssertionError: Expected encoder.onnx in output, got: set()
assert 'encoder.onnx' in set() (it works locally)

[Copilot]

@copilot SUBFAILED(model_id='openai/whisper-tiny') test/cli/test_cli_test_model_smoke.py::TestCliWhisperSmoke::test_whisper_capture_onnx_graph - AssertionError: Expected encoder.onnx in output, got: s...

The root cause is that onnxruntime-genai's create_model calls os.listdir(cache_dir) unconditionally inside save_model, which raises FileNotFoundError when /home/runner/.cache/huggingface/hub doesn't exist (fresh CI runner with no prior HF downloads). Fixed in olive/passes/onnx/model_builder.py by adding Path(HF_HUB_CACHE).mkdir(parents=True, exist_ok=True) before calling create_model. Also switched WhisperConfig(...) to WhisperConfig.from_dict({...}) to fix the PYLINT unexpected-keyword-arg errors, matching the existing LlamaConfig.from_dict pattern.

[xadupre]

@copilot still unresolved FileNotFoundError: [Errno 2] No such file or directory: '/home/runner/.cache/huggingface/hub'

[Copilot]

@copilot still unresolved FileNotFoundError: [Errno 2] No such file or directory: '/home/runner/.cache/huggingface/hub'

The root cause was more subtle than a missing directory. The create_model function in onnxruntime_genai already calls os.makedirs(cache_dir, exist_ok=True) internally, so the directory always exists when building starts. The real issue is in the Whisper-specific model saving flow: WhisperModel.save_model calls encoder.save_model then decoder.save_model. The encoder's save_model cleanup (in base.py:828) does if not os.listdir(self.cache_dir): os.rmdir(self.cache_dir) — for a local model there are no downloads so the cache dir is empty, gets deleted by the encoder, and then the decoder's save_model fails with FileNotFoundError on the same path.

Fixed in the latest commit by placing a temporary marker file (.olive_build_<pid>) in HF_HUB_CACHE before calling create_model. This keeps the directory non-empty throughout the build so onnxruntime_genai's cleanup never removes it. The marker is deleted in a finally block after create_model completes.

[xadupre]

@copilot can you move the new code into a separate file

[Copilot]

Done. Moved all Whisper-specific code (_save_local_tiny_whisper, _run_whisper_capture_onnx_flow, TestCliWhisperSmoke, DEFAULT_WHISPER_MODEL_IDS) into a new test/cli/test_cli_whisper_smoke.py. The workflow now runs both files.

[Copilot]

@copilot SUBFAILED(model_id='openai/whisper-tiny') test/cli/test_cli_test_model_smoke.py::TestCliWhisperSmoke::test_whisper_capture_onnx_graph - AssertionError: Expected encoder.onnx in output, got: s...

This was fixed in commit be34bc8 by placing a temporary marker file in HF_HUB_CACHE before calling create_model, preventing onnxruntime-genai's cleanup from deleting the shared cache_dir between the encoder and decoder saves. The latest "Test model fast" CI run passed successfully.

@copilot still unresolved FileNotFoundError: [Errno 2] No such file or directory: '/home/runner/.cache/huggingface/hub'

Fixed in be34bc8. The root cause was onnxruntime-genai's save_model cleanup deleting the shared cache_dir after the encoder build (because it was empty for local models), causing the decoder build to fail. The fix keeps a temporary marker file (.olive_build_<pid>) in HF_HUB_CACHE for the duration of create_model, so the directory is never considered empty. The latest CI run confirmed the fix works.