[tests] EAGLE3 uses same optimum-intel as others#3340
[tests] EAGLE3 uses same optimum-intel as others#3340peterchen-intel wants to merge 2 commits intoopenvinotoolkit:masterfrom
Conversation
1. Update optimum-intel in tests requirement.txt to commit id includes EAGLE3 draft model support 2. Remove specified optimum-intel installation for EAGLE3 tests 3. Move VLM (MiniCPM-o-2_6) tests to the end of the list since it specifies transformers verions 4. Fix EAGLE3 draft model conversion issues Signed-off-by: Chen, Peter <peter.chen@intel.com>
github-actions
bot
added
category: continuous batching
There was a problem hiding this comment.
Pull request overview
This PR updates the test infrastructure to use the official optimum-intel repository for EAGLE3 model support, eliminating the need for a custom fork. The changes primarily affect test configuration and CI workflows to properly handle EAGLE3 draft model conversion using the newly merged upstream support.
Changes:
- Updated optimum-intel dependency to a specific commit that includes EAGLE3 support from merged PR #1588
- Removed custom optimum-intel installation steps from CI workflows for EAGLE3 tests
- Added explicit
trust_remote_code=Trueparameter when converting EAGLE3 draft models - Reordered test execution to run EAGLE3 tests before VLM tests that require specific transformers versions
Reviewed changes
Copilot reviewed 7 out of 7 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| tests/python_tests/requirements.txt | Updated optimum-intel to git commit hash with EAGLE3 support |
| tests/python_tests/utils/hugging_face.py | Changed EAGLE3 parameter from custom "eagle3" flag to standard "trust_remote_code" |
| tests/python_tests/test_continuous_batching.py | Added conditional logic to pass trust_remote_code for EAGLE3 models |
| tests/python_tests/samples/conftest.py | Removed deprecated --eagle3 flag from conversion arguments |
| .github/workflows/windows.yml | Removed custom optimum-intel installation and reordered tests |
| .github/workflows/manylinux_2_28.yml | Removed custom optimum-intel installation and reordered tests |
| .github/workflows/linux.yml | Removed custom optimum-intel installation and reordered tests |
| if "eagle3" in str(draft_model_id).lower(): | ||
| draft_model_path = download_and_convert_model(draft_model_id, trust_remote_code=True).models_path | ||
| else: | ||
| draft_model_path = download_and_convert_model(draft_model_id).models_path |
There was a problem hiding this comment.
The pattern of checking for "eagle3" in the model ID and conditionally adding trust_remote_code=True is duplicated in three locations (lines 590-593, 700-703, and in test_eagle3_sd_string_inputs). Consider extracting this logic into a helper function that wraps download_and_convert_model to handle EAGLE3 models. This would reduce code duplication and make the logic easier to maintain if the EAGLE3 handling needs to change in the future.
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
| if "eagle3" in str(draft_model_id).lower(): | ||
| draft_model_path = download_and_convert_model(draft_model_id, trust_remote_code=True).models_path | ||
| else: | ||
| draft_model_path = download_and_convert_model(draft_model_id).models_path |
There was a problem hiding this comment.
This is a duplicate of the conditional logic added at lines 590-593. Since hugging_face.py already automatically detects EAGLE3 models and sets trust_remote_code=True, this explicit parameter passing is redundant. Consider removing this conditional and relying on the automatic detection for better maintainability.
1 participant
Description
Update optimum intel for test cases only, doesn't change GenAI source code.
Use official optimum intel repo to test eagle3 pipeline since huggingface/optimum-intel#1588 which supports eagle3 draft model conversion has already been merged.
Tickets: CVS-179182
EAGLE3 enabling PR #3055
Checklist: