Repository navigation
0.3.2: hash image ids and cache_salt into recovered blocks - #1
Merged
Merged
Conversation
apply_stored takes block_extras: per-block lists of u64 ids that are mixed into each block's tokens_hash after its tokens, via the BlockExtraInfo that create_stored_block_from_parts already accepts (it was passed None, with a "TODO multimodal support" comment). This lets /kv_recover dumps and gap replays hash image blocks and a request's cache_salt the same way the standalone indexer's live ZMQ path does (deepinfra/dynamo #53/#54). Without it, recovered blocks went on the token-only path. Also export cache_salt_extra_hash(salt) -> u64, so the engine derives the salt id with the same XXH3 as the indexer instead of needing Python xxhash. block_extras defaults to None, which keeps 0.3.1 behaviour. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Shang-Pin
marked this pull request as ready for review
October 19, 2026 21:22
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
apply_stored(..., block_extras=None)adds per-block u64 ids to each block'stokens_hashafter its tokens. It uses theBlockExtraInfothatcreate_stored_block_from_partsalready takes; until now that argument was passedNonewith aTODO multimodal supportcomment.cache_salt_extra_hash(salt) -> u64is now exported, the same id the standalone indexer derives: a 64-hex salt uses its first 16 hex digits, any other salt is its XXH3-64./kv_recoverTreeDumps and gap replays now hash image andcache_saltblocks the way the live ZMQ path does (standalone-indexer: hash a vLLM cache_salt into its sequence's first block dynamo#53, #54). Before this, recovered blocks went on the token-only path until evicted.block_extrasdefaults toNone, which keeps 0.3.1 behavior. Version bumped to 0.3.2. Based ontier-aware-recovery, where 0.3.1 lives.Validation
./build_wheel.sh(manylinux_2_28, cp312) builds.vllm:v0.29.1rc0-ds4-patches-*image, with this wheel installed and the engine feed from vllm-private branchv0.29.1rc0-ds4-ready-kvextras: thetokens_hashvalues in a local-indexer dump match the hash vectors shared with the dynamo and deepapi tests (salted, salted plus image, plain). 34/34 KV-event tests pass.After merge
./release.sh v0.3.2: the v0.29.1rc0-ds4 engine Dockerfile installs from that release URL. This also closes the gap left by 0.3.1, which never got a release.🤖 Generated with Claude Code