Deepseek v2 support

[hnyls2002]

To use deepseek v2, please sepcify the --context-length or --max-num-reqs to avoid oom. The context length for deepseek is quite large, for the current static req_to_token layout, we cannot support large requests num and large context length at the same time.

[m0g1cian]

Looking forward to see Deepseek v2 gets supported! I was trying to do the same thing two weeks ago but found the exact same issue of

RuntimeError: shape mismatch: value tensor of shape [7, 16, 256] cannot be broadcast to indexing result of shape [7, 16, 40]

[Xu-Chen]

Thank you for your excellent work. Will you support MLA in the future? Reduce KV cache to support larger context length.