Expose hash-proven recurrent replay boundaries - #142
Closed
FujitsuPolycom wants to merge 1 commit into
Closed
Conversation
The exact vLLM 0b67266 overlay exposes SchedulerOutput.recurrent_boundary_blocks as request-scoped group, physical-block, and token-boundary triples. Aligned Mamba entries require exact block-hash token and group identity; later running state and seven DFlash verification slots are never scanned. Connectors opt in through supports_recurrent_boundary_blocks. Existing partial_tail_offloads remains unchanged and its CoW targets join the generalized hand-off. Boundary pins use request cleanup lifetime, and overlapping producer batches defer block recycling until worker execution completes. The overlay pins SparkCache commit 9e779c32b285e00577a7829a75192069d12685dc, which requires complete recurrent-group proof and accepts the four vLLM postimages in its lease contract. Cache namespace impact: none; cache identities, chunk geometry, and persistent schemas are unchanged. Validation: exact composed-source verifier passed; exact-vLLM focused tests 5 passed; overlay tests 27 passed; runtime tests 235 passed, 3 skipped; Ruff passed. The broad available-tree suite reported 998 passed, 9 skipped, and one unrelated existing README assertion failure.
This was referenced Aug 30, 2026
Owner
Author
FujitsuPolycom
deleted the
codex/glm53-dflash7-recurrent-boundary-contract
branch
August 31, 2026 01:55
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Resulting behavior
The exact vLLM
0b67266a0f37d6146a8403fb8482403c62f412d5Python overlay exposes hash-proven recurrent replay boundaries to opt-in KV connectors:Each tuple is
(group_id, block_id, boundary_tokens). An aligned Mamba entry is emitted only when the physical block is non-null, its block hash belongs to the same KV group, andblock_hash_num_tokensequalsfloor((prompt_tokens - 1) / hash_block_size) * hash_block_size. Later running state and seven DFlash verification slots are never scanned.Connectors opt in through
supports_recurrent_boundary_blocks. Existingpartial_tail_offloadsremains unchanged; its copy-on-write targets are also present in the generalized hand-off. Non-opt-in connectors discard aligned candidates without taking pins.Lifetime and failure behavior
Boundary pages are pinned through the existing request-cleanup lifetime. Overlapping producer batches defer block recycling when the connector advertises the capability, so preemption cannot recycle a handed-off page before worker execution and synchronous snapshot completion.
Request cleanup releases pins. The exact tests advance beyond the 6,912-token boundary until arithmetic slot 2 becomes null, prove the exported block remains distinct from seven later speculative slots, and verify cleanup returns its reference.
Exact source identities
5a6561a5bbab990dcd03bfd6a485ea26c3b5a578c2fd61b76305767b16dbfba04a23f25b540760aa3f793427017e3ee39605e60de7af1b851cbbd6cc0faeddeckv_cache_manager.py:ee03dc9c…→c5b83d38…sched/output.py:65235eba…→9911b3f9…sched/scheduler.py:6d397c97…→260f36ce…single_type_kv_cache_manager.py:e4b1c5c3…→f67a1850…The builder pins SparkCache commit
9e779c32b285e00577a7829a75192069d12685dc, tree4df3ea1435241a688e6d44345687414605131450, deployable source SHA-25601cc59bf2c45af60f02813b771a484e65829e4f464eb27aa1756ca7067c82c9f, and lease-contract SHA-25645d7a92b38b836a4f829f02df85e339cfeea860e1080e4663a8340af6c125125.Compatibility and stacking
Cache namespace impact: none. Persistent cache identities and schemas are unchanged.
This draft is stacked on SparkRing #139 at commit
2f94b8405b54658a6b54d9b3377570d320879c6f. It consumes the capability implemented by SparkCache draft #31. Integration order is SparkCache #31, then this runtime overlay.No image was built or published, and no service or host state was changed.
Validation
python -m pytest runtime -q— 235 passed, 3 skippedpython -m ruff check .— passedC4 and C8 were capacity-limitedis absent)`n- GitHub maintained-tree suite — 1,944 passed, 9 skipped, with the same inherited assertion failure already present on base draft Remove unused DFlash runtime warning sources #139