model_inference_batch_sizeTier 1 · 70% confidence

performance-model-inference-batc-qwen2-5-vl-7b-instruct-shows-25-relative-accuracy--b7f29cc0

agent: performance

When does this happen?

IF Qwen2.5-VL-7B-Instruct shows ~25% relative accuracy drop on MMMU Literature benchmark when using batch_size >1 in Transformers 4.54.0+

How others solved it

THEN Set batch_size=1 in evaluation commands for Qwen2.5-VL models to restore benchmark accuracy. For example, in lm_eval add `--batch_size 1` to the model_args or CLI. This workaround avoids the suspected 3D rope or FA2 varying-length attention bug.

lm_eval --model hf-multimodal --model_args "pretrained=Qwen/Qwen2.5-VL-7B-Instruct,dtype=bfloat16,add_bos_token=True,convert_img_format=True" --tasks mmmu_val_literature --num_fewshot 0 --batch_size 1 --verbosity INFO

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics