cnxup/Qwen2.5-VL-7B-MLA-stage1-rope32

Research Paper: "MHA2MLA-VLM: Enabling DeepSeek's Economical Multi-Head Latent Attention across Vision-Language Models"

Model Description

This is a Stage 1 checkpoint for the MHA2MLA-VLM conversion process. This model serves as an intermediate checkpoint trained with partial RoPE (rope_dim=32) and is specifically designed to provide weight initialization for Stage 2 MHA2MLA-VLM models through SVD decomposition.

Usage

This Stage 1 model is primarily used to generate SVD initialization weights for Stage 2 MLA models:

• Step 1: Download this Stage 1 checkpoint

• Step 2: For MHA2MLA-VLM models using Partial-RoPE MKL method, Download the MKL file.

• Step 3: Download and use the corresponding Stage 2 MHA2MLA-VLM models for inference. Refer to each Stage 2 model's README for detailed inference instructions. Choose from different d_kv dimensions:

  • Qwen2.5-VL-7B-MLA-stage2-rope32-d_kv_32
  • Qwen2.5-VL-7B-MLA-stage2-rope32-d_kv_64
  • Qwen2.5-VL-7B-MLA-stage2-rope32-d_kv_128

Citation

@misc{fan2026mha2mlavlmenablingdeepseekseconomical,
      title={MHA2MLA-VLM: Enabling DeepSeek's Economical Multi-Head Latent Attention across Vision-Language Models}, 
      author={Xiaoran Fan and Zhichao Sun and Tao Ji and Lixing Shen and Tao Gui},
      year={2026},
      eprint={2601.11464},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.11464}, 
}