{{ message }}
Repository navigation
Conversation
jongukc
requested review from
Ying1123,
alphabetc1,
hanming-lu,
hnyls2002,
huangtingwei9988,
hzh0425,
ispobock,
merrymercy,
xiezhq-hermann and
yizhang2077
as code owners
October 21, 2026 13:26
This comment has been minimized.
This comment has been minimized.
Contributor
This branch has not been deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.

Assisted-by: OpenAI Codex
Motivation
When a request finishes before its first Mamba tracking interval, the int8-checkpoint path reads an unwritten ping-pong state slot, then discards the zero-length checkpoint. The read is unnecessary; no output corruption was observed. Related PR #44718 and PR #39209 address different checkpoint cases.
Modifications
Skip the state read when tracked length is zero; nonempty checkpoints are unchanged. No user-facing documentation changes are needed.
Accuracy Tests
2e7c9d851c, the focused checkpoint suite passed2/2with CPU-device anddatasetsstubs. On the fix branch, Qwen3.5-0.8B-Base served 21/21 HTTP requests on an RTX A6000 with int8 Mamba checkpoints enabled. This was a functional smoke test, not a KVGuard-instrumented read check; the local kernel-version guard was bypassed becausesglang-kernel0.4.5 is older than main requires.Speed Tests and Profiling
RTX A6000, Qwen3.5-0.8B-Base, int8 checkpoints: two warmed baseline/fix pairs of 80-request rounds at concurrency 8. Output throughput ranges (baseline vs. fix) overlapped for short requests (571–650 vs. 616–647 tok/s) and longer requests (492–526 vs. 454–524 tok/s); no consistent change was observed. An unrelated GPU job overlapped later rounds, so small differences need an exclusive-GPU rerun.
Checklist
Review and Merge Process
/tag-and-rerun-ci,/tag-run-ci-label,/rerun-failed-ciCI States
Latest PR Test (Base): ❌ Run #39608468704
Latest PR Test (Extra): ❌ Run #39608468354
Latest PR Test (AMD ROCm 10): ❌ Run #39608468413