[Feature]: MLA Support

[chengtbf]

🚀 The feature, motivation and pitch

DeepSeek-V2 design MLA (Multi-head Latent Attention), which utilizes low-rank key-value union compression to eliminate the bottleneck of inference-time key-value cache, thus supporting efficient inference.

Can VLLM support MLA for accelerated inference?

@misc{deepseek-v2,
author = {DeepSeek-AI},
title = {DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model},
year = {2024},
note = {GitHub repository},
url = {https://github.com/deepseek-ai/deepseek-v2}
}

https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat
https://github.com/deepseek-ai/DeepSeek-V2/blob/main/deepseek-v2-tech-report.pdf

Alternatives

No response

Additional context

No response

[datalee]

mark

[obitoquilt]

mark

[chenrui17]

mark

[lengyueyang]

mark

[RanchiZhao]

mark

[Jiayi-Pan]

mark

[quwu0820]

mark

[qianchen94]

mark

[lumosity4tpj]

mark

[zhyncs]

ref #4650 (comment)

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!