[Bugfix] fixes the decoding metadata of dense mla's fp8 kvcache.

[sighingnow]

Require the flashmla patch vllm-project/FlashMLA#7 to be landed first.

[gemini-code-assist]

Code Review

This pull request fixes an issue with decoding metadata for dense MLA's FP8 K/V cache by introducing a specialized operator. The changes in the Python code correctly route the execution to this new operator when appropriate. However, there is a critical issue in the CMake configuration where the flashmla dependency is pointed to a personal fork. This practice introduces significant risks and should be rectified by merging the required changes into the official upstream repository and updating the commit hash accordingly.

[gemini-code-assist]

Pointing the GIT_REPOSITORY to a personal fork (sighingnow/FlashMLA) introduces a significant dependency risk. For project stability, security, and long-term maintainability, dependencies should point to official repositories. The required changes should be merged into the official vllm-project/FlashMLA repository first. Afterward, this pull request can be updated to use the new commit hash from the official repository.

        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
        GIT_TAG <new_commit_hash_from_official_repo>

[sighingnow]

@LucasWilkinson could you please take a look? Thanks!

[LucasWilkinson]

LGTM; will run evals and post here

edit: gsm8k looks good: DeepSeek-V2-Lite-Chat with fp8 kv-cache, 64.6%