feat: support mla model in vllm

[ztang2370]

Tested MLA model deepseek-ai/DeepSeek-V2-Lite on vllm 0.14.0 - 0.16.0.

[cui36]

Thanks @ztang2370!

[ivanium]

Nice contribution! At a high level, I feel alloc_mla_kv_caches should be merged into alloc_kv_caches since their logic are largely similar and we have already had a attention_type field for this. wdyt?

[ztang2370]

Nice contribution! At a high level, I feel alloc_mla_kv_caches should be merged into alloc_kv_caches since their logic are largely similar and we have already had a attention_type field for this. wdyt?

Agreed. Updated with alloc_mla_kv_cache merged into alloc_kv_cache and tested. I think the same should be applied to the sglang side cc @cui36 . I'll do the sglang part later in another PR.

[ztang2370]

Hi @cui36 , I encountered problem setting up sglang with kvcached from source on the DGX Spark device, and therefore does not have device to test the refactor of sglang part for now (the smallest MLA model DeepSeek-V2-Lite takes more than 32GB). I'll leave it to you if you don't mind.
On DGX Spark now we can set up sglang only with Docker.

[jiarong0907]

/gemini review

[gemini-code-assist]

Code Review

This pull request adds support for MLA models in vllm. The changes are mostly in kvcached/integration/vllm/interfaces.py and kvcached/integration/vllm/patches.py to handle the specifics of MLA, such as the combined KV buffer. The implementation is sound, but I've identified a couple of areas with code duplication that could be refactored to improve maintainability. My review includes suggestions for these refactorings.