EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix expert distribution recorder and profiler command stuck forever (#6284)

Browse Source
This commit is contained in:
fzyzcjy
2025-05-18 08:10:44 +08:00
committed by GitHub
parent fd08c04821
commit 4086566516
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/managers/tokenizer_manager.py
View File

@@ -792,6 +792,7 @@ class TokenizerManager:
with_stack: Optional[bool] = None, with_stack: Optional[bool] = None,
record_shapes: Optional[bool] = None, record_shapes: Optional[bool] = None,
): ):
self.auto_create_handle_loop()
req = ProfileReq( req = ProfileReq(
type=ProfileReqType.START_PROFILE, type=ProfileReqType.START_PROFILE,
output_dir=output_dir, output_dir=output_dir,
@@ -814,16 +815,20 @@ class TokenizerManager:
return result return result
def stop_profile(self): def stop_profile(self):
self.auto_create_handle_loop()
req = ProfileReq(type=ProfileReqType.STOP_PROFILE) req = ProfileReq(type=ProfileReqType.STOP_PROFILE)
self.send_to_scheduler.send_pyobj(req) self.send_to_scheduler.send_pyobj(req)
async def start_expert_distribution_record(self): async def start_expert_distribution_record(self):
self.auto_create_handle_loop()
await self.expert_distribution_communicator(ExpertDistributionReq.START_RECORD) await self.expert_distribution_communicator(ExpertDistributionReq.START_RECORD)
async def stop_expert_distribution_record(self): async def stop_expert_distribution_record(self):
self.auto_create_handle_loop()
await self.expert_distribution_communicator(ExpertDistributionReq.STOP_RECORD) await self.expert_distribution_communicator(ExpertDistributionReq.STOP_RECORD)
async def dump_expert_distribution_record(self): async def dump_expert_distribution_record(self):
self.auto_create_handle_loop()
await self.expert_distribution_communicator(ExpertDistributionReq.DUMP_RECORD) await self.expert_distribution_communicator(ExpertDistributionReq.DUMP_RECORD)
async def update_weights_from_disk( async def update_weights_from_disk(