Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

MindVL: Towards Efficient and Effective Training of Multimodal Large Language Models on Ascend NPUs

Created by
  • Haebom

Author

Feilong Chen, Yijiang Liu, Yi Huang, Hao Wang, Miren Tian, Ya-Qi Yu, Minghui Liao, Jihao Wu

MindVL: A Multimodal Large-Scale Language Model Based on Ascend NPUs

Outline

This paper proposes MindVL, a multimodal large-scale language model (MLLM) trained on Ascend NPUs. MindVL addresses the limited hardware platform dependency and closed data recipe issues of existing MLLM training. It supports stable and high-performance training of large-scale Dense and Mixture-of-Experts (MoE) models on Ascend hardware through an efficient training framework called MindSpeed-MLLM. Furthermore, it provides a systematic and open description of the training data generation method and mixing strategy. MindVL is a data-efficient MLLM trained end-to-end on Ascend NPUs. It improves performance by averaging the weights of checkpoints trained with various sequence lengths and by employing a test-time resolution search technique. MindVL-8B achieves the same performance as Qwen2.5VL-7B with 10% of the data, and MindVL-671B-A37B, an MoE model, shows similar performance with 3% of the data of Qwen2.5VL-72B.

Takeaways, Limitations

•
Takeaways:
◦
Presenting Ascend hardware as a valid alternative for MLLM training.
◦
Promoting reproducibility and openness in research by providing open data recipes.
◦
Presents effective performance enhancement techniques such as checkpoint weighted averaging and test time resolution exploration.
◦
Achieve competitive performance with less data through data-efficient model training.
•
Limitations:
◦
The paper may lack detailed information about the specific dataset size or model architecture.
◦
Comprehensive comparisons with other cutting-edge models and extensive benchmark results may not be sufficient.
◦
As this is a training framework specifically designed for Ascend NPUs, generalizability to other hardware environments may be limited.
◦
There may be a lack of analysis of the model's practical applicability and application to various real-world problems.
👍
Made with Slashpage