Transformers update fixes

#1
by varun92 - opened
Red Hat AI org
edited 3 days ago

Update modeling-phi4-visionr.py to be compatible across transformers versions

  • Account for the changes in image processing (Transformers 5.4.0 moved to TorchVisionBackend for image processing.)
  • Account for the removed filter_out_non_signature_kwargs decorator
  • Account for introduction of DynamicCache
varun92 changed pull request status to open
Red Hat AI org
edited 3 days ago

Testing :

vllm serve : vllm serve RedHatAI/Phi-4-reasoning-vision-15B --revision refs/pr/1 -tp 2 --trust-remote-code --async-scheduling --enable-chunked-prefill --enable-prefix-caching --limit-mm-per-prompt '{"image": 10}'

eval command : OPENAI_API_KEY=empty python -m lmms_eval --model openai --model_args model_version=RedHatAI/Phi-4-reasoning-vision-15B,base_url=http://localhost:8000/v1 --tasks mmmu_val --batch_size 1 --log_samples --output_path ./results/

transformers==5.5.3

| Tasks  |Filter|n-shot| Metric |   |Value |   |Stderr|
|--------|------|-----:|--------|---|-----:|---|------|
|mmmu_val|none  |     0|mmmu_acc|↑  |0.5122|±  |N/A   |

transformers==4.57.6

| Tasks  |Filter|n-shot| Metric |   |Value |   |Stderr|
|--------|------|-----:|--------|---|-----:|---|------|
|mmmu_val|none  |     0|mmmu_acc|↑  |0.5222|±  |N/A   |
Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment