[Bugfix] Zero recycled KV cache blocks for FullAttention models

[AjAnubolu]

Summary

Closes #39146. The KV block zeroing pipeline from #35219 was gated to Mamba-only models; enabling it for FullAttention prevents stale K/V in partial-block tail slots from propagating NaN through masked softmax.

[gemini-code-assist]

Code Review

This pull request updates the needs_kv_cache_zeroing property to include models using FullAttentionSpec, preventing stale K/V data leakage in partial-block tail slots as identified in issue #39146. A regression test was added to verify this behavior. Feedback suggests using isinstance() for type checking to ensure compatibility with subclasses like MLAAttentionSpec and to follow PEP 8 guidelines.

[gemini-code-assist]

Using type(g.kv_cache_spec) is FullAttentionSpec is overly restrictive as it excludes subclasses like MLAAttentionSpec and SinkFullAttentionSpec. These variants of full attention likely suffer from the same stale K/V issues in partial blocks and should also benefit from zeroing. Following PEP 8 recommendations, object type comparisons should use isinstance() instead of comparing types directly, which also ensures consistency with the has_mamba_layers implementation.

Suggested change

	return self.has_mamba_layers or any(
	type(g.kv_cache_spec) is FullAttentionSpec for g in self.kv_cache_groups
	)
	return self.has_mamba_layers or any(
	isinstance(g.kv_cache_spec, FullAttentionSpec) for g in self.kv_cache_groups
	)

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @AjAnubolu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @AjAnubolu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @AjAnubolu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[ppham-nv]

GHSA-x368-4g9h-fvv4

anyone have contacts to be able to fix this advisory? This PR is still not merged so the claim that the patch from 1ad6786 (only exists in the fork) is fixed on "0.19.1" vLLM is wrong.

[kamb-code]

Cross-referencing for context: I ran a replication/discrimination study of #39146 on current main (b01728b08, RTX A6000, 10 runs per configuration). With this PR's zeroing change applied to main and instrumented (canary-verified zeroing of 30,341 recycled blocks, 0 verification failures), the temperature=0 divergence persists unchanged; VLLM_BATCH_INVARIANT=1 on the stock allocator eliminates it while scheduler batch compositions still vary. Full evidence and raw data: #39146 (comment)