[Perf] Add decode full-graph support to FlashInfer-MLA backend

[benchislett]

Purpose

The annotation was missing from FlashInfer-MLA while the implementation has support.

Running DSR1-FP4 on 4xB200 gets me 97 TPS:

VLLM_FLASHINFER_MOE_BACKEND=latency VLLM_USE_FLASHINFER_MOE_FP4=1 VLLM_ATTENTION_BACKEND=FLASHINFER_MLA vllm serve nvidia/DeepSeek-R1-FP4 -tp 4 --max-model-len 32768 --max-num-seqs 128 --no-enable-prefix-caching --async-scheduling --port 8049

I also tested on a local development branch for MTP containing #25984, and #25987.

On that branch, with 3 MTP speculative tokens, I get 165 TPS and passing GSM8k evals.

Test Plan

GSM8k run as follows:

lm_eval \
  --model local-completions \
  --tasks gsm8k \
  --model_args base_url=http://0.0.0.0:8049/v1/completions,model=nvidia/DeepSeek-R1-FP4,tokenized_requests=False,tokenizer_backend=None,num_concurrent=128,timeout=120,max_retries=5

Test Result

Matches the baseline:

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9439|±  |0.0063|
|     |       |strict-match    |     5|exact_match|↑  |0.9439|±  |0.0063|

[gemini-code-assist]

Code Review

This pull request correctly enables full CUDA graph support for decode operations in the FlashInfer-MLA attention backend. The change is implemented by creating a new FlashInferMLAMetadataBuilder class that inherits from MLACommonMetadataBuilder and sets the cudagraph_support attribute to AttentionCGSupport.UNIFORM_BATCH. The FlashInferMLABackend is then updated to use this new builder. The approach is clean, follows the existing design patterns in the codebase, and seems to correctly enable the feature as described. The changes are minimal and well-targeted. I found no issues of high or critical severity.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

[LucasWilkinson]

LGTM; thanks!