lucylyn/MSD-Qwen2.5VL-7B-Instruct

MSD draft checkpoint for Qwen2.5-VL-7B-Instruct

This repository contains the MSD draft-model weights and configuration. Use it with the corresponding target model and the MSD implementation.

• Target model: Qwen/Qwen2.5-VL-7B-Instruct
• Evaluation adapter: qwen2_vl_msd
• Draft visual compression: KMeans, retaining 32 visual tokens by default. The target model retains the complete visual input.

Usage

Install the packages from the MSD repository, then run:

accelerate launch --num_processes=1 -m lmms_eval \
    --model qwen2_vl_msd \
    --model_args pretrained=Qwen/Qwen2.5-VL-7B-Instruct \
    --msd_model_path lucylyn/MSD-Qwen2.5VL-7B-Instruct \
    --tasks chartqa --batch_size 1 --limit 100 \
    --gen_kwargs temperature=0,max_new_tokens=512 \
    --use_msd

Files: config.json and pytorch_model.bin. This is a draft checkpoint to pair with the target model.