[Bugfix][Attention] Zero reused FlashInfer MLA KV blocks

[andylolu2]

Why

FlashInfer MLA kernels can consume values from the unwritten suffix of a KV-cache block. When a recycled block still contains non-finite data from its previous owner, masking is insufficient (0 * NaN remains NaN), which can corrupt attention outputs.

vLLM already has a scheduler-to-worker path for zeroing newly allocated KV blocks, but uniform-precision attention models do not currently enable it. This change adds a generic attention-backend capability and opts in only FlashInfer MLA dense/sparse today.

Summary

• Add a generic AttentionBackend.requires_kv_cache_zeroing capability and publish it on the base AttentionSpec.
• Preserve the capability across attention-spec merges, uniform-type scheduler projection, and composite backends.
• Enable it for FlashInfer MLA dense and all registered FlashInfer MLA sparse variants, including DeepSeek V4/V4.1 compressed and sliding-window caches.
• Leave FlashMLA dense/sparse and all other attention backends unchanged.

Reviewer guide

• vllm/v1/attention/backend.py and vllm/v1/kv_cache_interface.py contain the core logic: the generic backend opt-in, base attention-spec metadata, merge propagation, and zeroing decision.
• vllm/v1/core/kv_cache_utils.py preserves the capability when a UniformTypeKVCacheSpecs group is collapsed for the scheduler.
• vllm/v1/attention/backends/composite.py ORs the requirement from either child backend.
• The FlashInfer backend files only opt the relevant dense/sparse backend classes into the generic capability.
• The tests cover backend scoping and propagation through base/MLA, sliding-window, uniform-type, and composite paths.
• No kernel or block-zeroer implementation changes are needed; this uses the existing newly-allocated-block zeroing pipeline.

Duplicate check

This is not a duplicate of the closest related work:

• [Bugfix] Generalize KV block zeroing to AttentionSpec #51749 generalized the existing worker zeroer to all AttentionSpec subclasses, but did not enable zeroing for uniform-precision MLA configurations.
• Closed, unmerged [Bugfix] Fix MLA KV cache blocks not zeroed on reuse, causing CUDA crashes under concurrent load #37530 enabled zeroing for every MLA backend; this change is deliberately restricted to FlashInfer MLA dense/sparse.
• [Core][Perf] Zero new KV blocks by cache group #52187 optimizes zeroing by cache group.
• [Bugfix][Attention] Zero the null KV block after CUDA graph capture #57158 restores the reserved null block after CUDA graph capture.
• [Bugfix][KV] Zero the never-written tail of async-loaded blocks #55085 zeroes unwritten tails after partial Mooncake loads.

Searches performed:

gh pr list --repo vllm-project/vllm --state open --search "FlashInfer MLA KV cache zeroing"
gh pr list --repo vllm-project/vllm --state open --search "KV cache corruption FlashInfer"
gh pr list --repo vllm-project/vllm --state open --search "flashinfer mla zero OR stale OR reused"

Test plan

.venv/bin/python -m pytest -q tests/v1/core/test_kv_cache_utils.py -k attention_kv_cache_zeroing
# 16 passed

.venv/bin/python -m pytest -q tests/v1/attention/test_mm_prefix.py -k composite_backend_preserves_kv_cache_zeroing_requirement
# 1 passed

.venv/bin/python -m pytest -q tests/v1/attention/test_mla_backends.py -k only_flashinfer_mla_backends_require_kv_cache_zeroing
# 1 skipped locally because FlashInfer is not installed

# Directly imported and exercised customize_spec for all seven sparse FlashInfer MLA backend classes.
# verified sparse customize_spec paths

.venv/bin/pre-commit run --files <changed files>
# all applicable hooks passed, including Ruff and mypy

Model evaluation

Not run in this CPU-only development environment because FlashInfer and a CUDA accelerator are unavailable. GPU reproduction of the incident and an output-correctness check are required before this draft is marked ready for review.

AI assistance

AI assistance was used to investigate the relevant backend/cache-spec paths, implement the change, and draft this description. The submitter must review every changed line and validate the fix on the affected GPU configuration before merge.