fix(vllm): reset block hash on prefix-cache eviction

[shipiyouniao]

Summary

Fixes stale vLLM KVCacheBlock.block_hash state when kvcached evicts or frees prefix-cache blocks.

Closes #362.

Root Cause

ElasticBlockPool removes evicted blocks from kvcached's _cached_blocks, _block_id_to_key, and _evictable_blocks indexes, but the underlying vLLM block object can keep its old block_hash value. Reusing that block for a different prefix can leave stale block state behind, and newer vLLM block implementations require reset_hash() before assigning a new hash.

What Changed

• Add _reset_block_hash() helper that calls reset_hash() when available and falls back to clearing _block_hash for older/mock block objects.
• Reset block hash when blocks are evicted from the lazy prefix-cache pool.
• Reset block hash for uncached blocks that are immediately returned to kvcached.
• Reset block hash during explicit evict_blocks() and reset_prefix_cache().
• Add regression coverage to tests/test_prefix_cache.py for eviction, reset, explicit eviction, and reallocation/recache flows.

Tests

python -m pytest tests/test_prefix_cache.py -q

Result: 28 passed.

[Copilot]

Pull request overview

Fixes a correctness issue in the vLLM integration where KVCacheBlock.block_hash could remain stale after prefix-cache eviction/free, causing reused blocks to carry incorrect hash state across prefixes.

Changes:

• Add a _reset_block_hash() helper to clear vLLM block hash state (preferring reset_hash() when available).
• Invoke block-hash reset on eviction/free/reset flows in the patched ElasticBlockPool.
• Extend tests/test_prefix_cache.py to regress block-hash clearing across eviction, explicit eviction, reset, and reuse scenarios.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
kvcached/integration/vllm/patches.py	Clears KVCacheBlock hash state when blocks are returned to kvcached via eviction/free/reset paths to prevent stale reuse.
tests/test_prefix_cache.py	Strengthens the mock block to enforce write-once hash semantics and adds assertions ensuring hashes are cleared on eviction/free/reset/reuse.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.