Repository navigation
Conversation
Rockdu
force-pushed
the
fix/sgld-scheduler-port-race
branch
from
October 4, 2026 00:58
f2144d4 to
d0dd3b2
Compare
Rockdu
force-pushed
the
fix/sgld-scheduler-port-race
branch
from
October 4, 2026 02:07
d0dd3b2 to
1b88975
Compare
Rockdu
force-pushed
the
fix/sgld-scheduler-port-race
branch
2 times, most recently
from
October 4, 2026 02:59
acd3c4d to
b25885f
Compare
Rockdu
marked this pull request as ready for review
October 4, 2026 03:00
Rockdu
requested review from
AgainstEntropy,
BBuf,
HaiShaw,
kevin-mii,
mickqian,
niehen6174,
ping1jing2 and
yichiche
as code owners
October 4, 2026 03:00
Rockdu
force-pushed
the
fix/sgld-scheduler-port-race
branch
4 times, most recently
from
October 4, 2026 07:41
bd88b5b to
09b745c
Compare
… clients Signed-off-by: rockdu <kangrdu@gmail.com>
Rockdu
force-pushed
the
fix/sgld-scheduler-port-race
branch
from
October 4, 2026 07:44
09b745c to
1853880
Compare
Rockdu
added a commit
that referenced
this pull request
Oct 4, 2026
…ler actually bound to its clients (#42470) Signed-off-by: rockdu <kangrdu@gmail.com>
nv-hwoo
pushed a commit
to nv-hwoo/sglang
that referenced
this pull request
Oct 20, 2026
…] fix: report the port each scheduler actually bound to its clients (sgl-project#42470) Signed-off-by: rockdu <kangrdu@gmail.com>
Collaborator
|
seems #43732 fixed the same issue the opposite way: the scheduler binds with However I believe in RL case (and most other cases), the losing server should still start, bind a free port and report it faithfully as this PR does. |
Rockdu
added a commit
that referenced
this pull request
Oct 30, 2026
…rt the port each scheduler actually bound to its clients (#42470) Signed-off-by: rockdu <kangrdu@gmail.com>
Rockdu
added a commit
that referenced
this pull request
Oct 30, 2026
…rt the port each scheduler actually bound to its clients (#42470) Signed-off-by: rockdu <kangrdu@gmail.com>
Rockdu
added a commit
that referenced
this pull request
Oct 31, 2026
…rt the port each scheduler actually bound to its clients (#42470) Signed-off-by: rockdu <kangrdu@gmail.com>
Rockdu
added a commit
that referenced
this pull request
Oct 31, 2026
…rt the port each scheduler actually bound to its clients (#42470) Signed-off-by: rockdu <kangrdu@gmail.com>
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.
Motivation
Diffusion servers starting together on one host can pick the same scheduler port: it is guessed as
5555 + randint(0, 100)and checked by a probe that releases it. The loser's scheduler then bindsport + 42, but its HTTP side still connects to the guessed port, so its requests are served by the other server's scheduler and GPU.We hit this as an intermittent RL CI failure, caught by the UUID check from #32685:
Modifications
dp_replica,actual_scheduler_port); the launcher records it inserver_args.scheduler_portsbefore any client connects. A scheduler that moved off a taken port is now reached on its new port.server_args, so the metricsreplicalabel matches the real endpoint.--strict-portsnow fillsscheduler_portsas well, so the list is always set.Accuracy Tests
Not applicable.
Speed Tests and Profiling
Not applicable.
Checklist
CI States
Latest PR Test (Base): ❌ Run #37186702005
Latest PR Test (Extra): ❌ Run #37186701834
Latest PR Test (AMD ROCm 10): ❌ Run #37186701981