Repository navigation
Repair prefix cache retention - #38
Merged
Merged
Conversation
Count retiring backing once, retry blocked stores, and refill compacted rows atomically. CPU: Helios 900 passed, 22 ignored, 3 memlock failures; protocol 48 passed; CLI 134 passed, 1 ignored; package 98 passed, 1 missing-passwd-home failure. Release and extension builds and normal clippy passed. Strict clippy and full format have existing failures. GPU and integration checks are to run in the test round. Design-Signed-Off-By: yatesdr; design-review=6aeb
Distinguish failed stores that replaced a file on some ranks from failures that leave every prior file intact. Invalidate the damaged ancestor while retaining descendant links, charge the full replacement size until reclaim, and allow missing-file cleanup to complete successfully. Intact replacements still roll back to their previous record. Cover partial writes on either rank, cleanup failure, ancestor invalidation, GPU-resident prefix bypass, reclamation, and all-rank write failure. Validation: full Helios tests with server,runtime-package passed (953 passed, 34 ignored); the Flare release build and formatting checks passed. GPU concurrency, isolation performance, and live multi-host qualification were not run. Design-Signed-Off-By: Codex GPT-6; design-review=6aeb
Include ARM package hashing dependencies and the Qwen recurrence documentation. The prefix-cache changes retain their shared runtime interfaces and preserve package compatibility. Validation: full Helios suite with server,runtime-package passed (953 passed, 34 ignored), and the Flare release build passed on x86_64. GPU concurrency, isolation performance, live multi-host qualification, and ARM execution were not run. Design-Signed-Off-By: Codex GPT-6; design-review=6aeb
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
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.
Problem
Under pressure from one long conversation, the prefix cache drops other conversations' recent checkpoints. The host tier then often recomputes them instead of restoring them. On DeepSeek V4.1 Flash TP4, an 855k-token session beside a ~170k-token classifier transcript turned 2 s classifier calls into 16–17 s full prefills.
Change
Tests
Helios 900, protocol 48, CLI 134 and package 98 pass on CPU. DeepSeek V4.1 TP4 concurrency-32 replay, memory/disk/composite isolation against the documented 2% limits, and multi-host qualification are still to do.