Launch sglang engines gated and release them once they are running - #2096
Launch sglang engines gated and release them once they are running#2096fzyzcjy wants to merge 1 commit into
Conversation
|
Caution The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased. |
9a663a9 to
f3701c7
Compare
8423dcd to
0fdb385
Compare
f3701c7 to
0cfae7d
Compare
0fdb385 to
e0a4dce
Compare
0cfae7d to
2a9824b
Compare
e0a4dce to
ee99103
Compare
2a9824b to
13811cb
Compare
ee99103 to
109bed1
Compare
13811cb to
2589578
Compare
109bed1 to
b243dc4
Compare
2589578 to
206df79
Compare
b243dc4 to
ce501c4
Compare
bfc30f2 to
e024b0e
Compare
165008b to
0e6df86
Compare
14427e3 to
9d64207
Compare
4aef8a7 to
fd52c7c
Compare
9d64207 to
37be5f8
Compare
fd52c7c to
536beb7
Compare
37be5f8 to
5877b7a
Compare
536beb7 to
c472209
Compare
22b54d0 to
2cd2419
Compare
c472209 to
2441d03
Compare
97b950a to
cdc6b8b
Compare
71b8062 to
1b1e4c2
Compare
de3bffa to
cca083c
Compare
503f4b5 to
9442cbe
Compare
8ff168d to
b2fa901
Compare
Pass the allocated gate port to every engine so sglang idles at its out-of-band control port instead of initializing straight away, and have RayWorkerManager release each cell after its workers were started. The release retries until the port answers, because it only starts listening once the engine reaches its gate.
b2fa901 to
fd2f4ec
Compare
360867b to
fd391b2
Compare
guapisolo
left a comment
There was a problem hiding this comment.
Approved in stack context after reviewing PR #2096 changes. The upstream gate dependency is tracked in #1837, while #2098 removes the test hang before #2176 and the repair is retained through checked tip #2649; I left one non-blocking P2 about keeping the gate port allocator-owned.
| @@ -193,8 +197,6 @@ def _compute_server_args( | |||
| if sglang_overrides: | |||
| kwargs.update(sglang_overrides) | |||
There was a problem hiding this comment.
[P2] Please treat gated_launch_port as allocator-owned, like host/port. Under an explicit override, SGLang listens on that port while ServerCell activates the allocator port, so startup waits on the wrong endpoint; this is non-blocking for the intended configuration, but we should reject the override in ServerGroupConfig.resolve() and omit the corresponding CLI field.
Part of #1837