open-compass / open-compass/VLMEvalKit

The evaluation device_map setting is not optimized for quantized models using the ExLlamaV2 backend.

Open
#629 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
4.4k
Forks
768
Avg merge
1d 10h
Merged PRs (30d)
17

Description

The evaluation setting is not optimized for quantized models using the ExLlamaV2 backend, which serves as the default kernel for Auto-GPTQ and Transformers.

image

Changing device_map="cpu" to device_map='auto' will resolve the model loading issue.
image

The known problematic model files are:
vlmeval/vlm/llama_vision.py
vlmeval/vlm/llava/llava.py
vlmeval/vlm/qwen2_vl/model.py

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by inspecting vlmeval/vlm/llama_vision.py, vlmeval/vlm/llava/llava.py, and vlmeval/vlm/qwen2_vl/model.py, focusing on the evaluation device_map setting. Replace the problematic CPU mapping with the reported automatic mapping, then verify that the affected quantized models load successfully through the evaluation workflow.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Bug
Difficulty
2/5
Estimated time
1-3 hours
Activity status
Stale
Clarity
Clearly specified
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.