Instrument MTP3 reuse and preserve continuation checkpoints in source builds - #239
Merged
Conversation
Emit optional admission, restore-finalization, accepted prompt-work, preemption and terminal events for the SparkCache request ledger. Preserve deferred GPU lease attribution and dispatch ranges across asynchronous scheduling. Apply an exact-source transform after verifying checkpoint ownership dependencies; retain immutable payloads and published receipts. Validation: 470 mesh tests passed, four optional skips, including five companion-ledger integration tests; Ruff and documentation links passed. Hardware validation requires a rebuilt image.
Install four byte-attested continuation sources after verifying checkpoint ownership, then apply the matching scheduler attribution transform. Preserve all unrelated dependencies and generate image receipts from the composed runtime. Enable both checkpoint flags and pin the SparkCache attribution source; published image receipts retain their identities. Avoid redundant empty prompt events during ordinary decode. Validation: 501 mesh tests passed, four optional skips, including both scheduler variants and companion-ledger cases; Ruff and 562 documentation links passed. Hardware validation remains pending.
… source Select the tested SparkCache package with same-pass survivor reconciliation, request-attribution accounting, and optional deletion pacing. Preserve immutable published image receipts and package the continuation runtime with the startup gate. Validation: 3,089 offline tests passed with 85 documented skips; Ruff, documentation links, and local context preparation passed. Image build and serving validation remain pending.
This was referenced Sep 6, 2026
…ions Retain raw cold-prefill observations and bounded recurrence/cache evidence for the exact continuation image, including unpaired-control and clock-resolution limits. Execute packaged scheduler/allocator methods for source retention, reserve overlap, publication pins and cleanup, and provide a portable harness with offline readiness/timing coverage. No cache identity or runtime source bytes change. Validation: 512 scoped CPU tests passed, 14 documented optional skips; Ruff passed. The attribution-containing image rebuild remains unqualified.
Reject absent, malformed, boolean or negative cached-token counts instead of accepting a missing field as zero. Apply the same accounting boundary to semantic reuse checks and cold timing. Preserve all raw observations: all 19 retained continuation rows contain explicit integer zero counts. Update portable-harness provenance. Validation: 14 offline readiness/timing tests pass, including six reproductions of missing or invalid accounting; Ruff passed. Runtime source packages are unchanged.
FujitsuPolycom
marked this pull request as ready for review
September 9, 2026 01:06
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.
Optional scheduler events distinguish admitted local prefixes, finalized persistent restores, and accepted prompt computation across preemption attempts. Events are enabled only when the connector opts in; cache offers and worker-local verification are not counted as completed reuse.
The research performance builder also preserves four exact continuation-checkpoint source files from image
sha256:489d1975619e9083d14f14bcd1c6cbb4a96c41e3ab3978af048dc3ed2bb452a8. It validates the complete checkpoint ownership contract before installing continuation and attribution transforms, then records the combined file inventory. Source archives, licenses, preimages, and replacement hashes remain pinned.This optional builder uses its own parent and SparkCache
b5aca7cd3d3f7e7a14636bf6e5fa1f50a9650168, from merged SparkCache #65. It preserves current startup admission and scheduler-liveness packaging. The common GLM source image, topology profiles, numerical source pins, and published-image receipts remain unchanged.Status: implemented and CPU-tested; a rebuilt attribution-containing image remains research-only pending supervised build/GPU/startup/serving validation. The continuation record applies only to its recorded image. Its three cold samples per shape yielded about 2,857–2,929 prompt tokens/s; no paired-control speedup is claimed, and historical Windows timing has about 15.625 ms granularity.
Validation: 557 mesh/profile/portable-harness CPU tests passed with 14 optional skips. The exact pinned SparkCache companion attribution suite passed all 50 tests. Ruff and whitespace checks passed. Tests cover scheduler variants, failed/cancelled restores, stale output, deferred leases, preemption, ownership drift, source tampering, and continuation allocation regressions. The full composed Docker build and public GPU harness were not rerun.