Skip to content

[None][fix] Preserve reusable KV cache prefix snapshots - #16713

Closed
jiaganc wants to merge 1 commit into
NVIDIA:mainfrom
jiaganc:codex/investigate-block-reuse-cache-hit
Closed

[None][fix] Preserve reusable KV cache prefix snapshots#16713
jiaganc wants to merge 1 commit into
NVIDIA:mainfrom
jiaganc:codex/investigate-block-reuse-cache-hit

Conversation

@jiaganc

@jiaganc jiaganc commented Jul 22, 2026

Copy link
Copy Markdown
Collaborator

@coderabbitai summary

Description

Per-request block reuse can remove a shorter partial endpoint when a longer sibling covers the same token prefix, even when the longer sibling does not yet preserve all lifecycle state needed to reuse that endpoint. With commit_min_snapshot, a later rewind can then follow an exact radix-tree path whose pages are not reusable and miss an older reusable sibling prefix.

This change:

  • keeps covered partial endpoints until the covering block has committed pages for every lifecycle, including an SSM snapshot at the same token boundary;
  • removes redundant covered siblings after the replacement state has been committed;
  • evaluates partial sibling matches by reusable tokens after lifecycle pruning; and
  • falls back from an unusable exact path to the best reusable sibling prefix.

There is no public API change.

Test Coverage

  • Added test_commit_min_snapshot_preserves_partial_endpoint to cover preservation and reuse of a rewind endpoint with mixed full-attention and SWA lifecycles.
  • pre-commit run --from-ref upstream/main --to-ref HEAD passes.
  • python3 -m py_compile tensorrt_llm/runtime/kv_cache_manager_v2/_block_radix_tree.py passes.

Known validation gap: the latest AGA concurrency-144 end-to-end reproduction still reported 21 low-hit rewind requests with zero cached tokens. This PR remains a draft while that discrepancy is investigated.

PR Checklist

Please review the following before submitting your PR:

  • PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

  • PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

  • Test cases are provided for new code paths (see test instructions)

  • If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

  • Any new dependencies have been scanned for license and vulnerabilities

  • CODEOWNERS updated if ownership changes

  • Documentation updated as needed

  • Update tava architecture diagram if there is a significant design change in PR.

  • The reviewers assigned automatically/manually are appropriate for the PR.

  • Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

Preserve reusable partial block snapshots, reject redundant covered blocks, and fall back to reusable sibling prefixes when the exact path is unavailable.

Signed-off-by: Jiagan Cheng <jiaganc@nvidia.com>
@jiaganc

jiaganc commented Jul 31, 2026

Copy link
Copy Markdown
Collaborator Author

Close. Will be replaced by @lowsfer another PR.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant