[sglang-miles] Remove obsolete PD retract guard after rebootstrap support (#25372); squash into 17c3be8e5d (#23672/#23887) - #39119
Merged
Conversation
Allow PD pause_generation(mode="retract") to reach the existing rebootstrap implementation. The old guard was replayed onto sglang-miles after main added hold_rebootstrap and resume-time enqueueing in #25372. When refreshing the sglang-miles patch stack, squash this fix into 17c3be8: [8/35] [sglang-miles] Improve PD pause handling (#23672, #23887). Verified: the unchanged pause suite improves from 3 failed / 17 passed to 20 passed. Five existing PD HTTP/GPU cases pass with Llama-3.1-8B-Instruct and Mooncake, with both endpoints on one H200 and CUDA graphs disabled.
guapisolo
requested review from
Ying1123,
hnyls2002,
merrymercy and
xiezhq-hermann
as code owners
September 11, 2026 18:54
guapisolo
changed the title
guapisolo
changed the title
42 tasks
HJSang
pushed a commit
to HJSang/sglang
that referenced
this pull request
Sep 21, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Restore PD retract pause by removing an obsolete rejection on
sglang-miles.Symptom & Reproduction
pause_generation(mode="retract")raises an assertion in both PD modes before pausing, blocking the supported retract/update/resume flow.32839114c4ada2ae237581405a8ff39dc0db9e25, runpython -m pytest -q test/registered/unit/managers/test_scheduler_pause_generation.py: 3 PD cases fail withpause_generation(mode='retract') is not supported in PD disaggregation mode yet; 17 pass. The existing tests expect PD retract to succeed.Root Cause
Scheduler.pause_generationreceived the old PD prohibition from [PD] Gate disagg event loops on _engine_paused; reject pause(retract) in PD #23672 when rebootstrap was unavailable.hold_rebootstrapand resume-time enqueueing were implemented in [PDD] Add true request retraction for PDD #25372, already merged intomain.17c3be8e5dc6579edce6715ba8e6ba05081170cdreplayed the obsolete guard ontosglang-miles, preventing entry into that implementation.Fix
Remove the 13-line guard and its outdated explanation. Keep mode validation, in-place pause handling, event-loop gates, and the existing retract/rebootstrap implementation unchanged.
Patch-stack maintenance: Squash this fix into 17c3be8e5dc6579edce6715ba8e6ba05081170cd, titled
[8/35] [sglang-miles] Improve PD pause handling (#23672, #23887), when refreshing thesglang-milespatch stack. That is the branch commit that reintroduced the outdated restriction.Verification
test_scheduler_pause_generation.py: all 20 unchanged tests pass after the fix.git diff --check: passed. Test processes exited; GPU memory and test ports were released.Review Focus
Scheduler.pause_generation: PD retract must reach the existing implementation while preserving unified and in-place behavior.CI States
Latest PR Test (Base): ❌ Run #34635912091
Latest PR Test (Extra): ❌ Run #34635911727
Latest PR Test (AMD ROCm 10): ❌ Run #34635911773