[Feature] Batch Invariant for R1 TP 8 on Blackwell

[yewentao256]

Purpose

Batch Invariant for R1 TP 8 on Blackwell

Backend: FlashinferMLA

Test

VLLM_ATTENTION_BACKEND=FLASHINFER_MLA VLLM_TEST_TP_SIZE=8 VLLM_TEST_MODEL="deepseek-ai/DeepSeek-R1" pytest -s -v tests/v1/generation/test_batch_invariance.py -k test_logprobs_bitwise_batch_invariance_bs1_vs_bsN[FLASHINFER]

PASSED

== 1 passed, 6 deselected, 4 warnings in 120.24s (0:02:00) ===

[gemini-code-assist]

Code Review

This pull request enables batch invariance for the FLASHINFER_MLA attention backend by adding it to the list of supported backends. The change is correct and aligns with the goal of supporting more backends for deterministic, batch-invariant execution. I've added one high-severity comment suggesting a refactor of the hardcoded list of supported backends. While the current change is fine, moving to a more dynamic approach where backends declare their own capabilities would improve maintainability and prevent this list from becoming a source of silent errors in the future.