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

[PD] Propagate internal server errors from aborted requests to clients instead of blindly returning 200's (#8936)

Browse Source
This commit is contained in:
datdo-msft
2025-08-18 14:23:46 -07:00
committed by GitHub
parent 6805f6da40
commit 98b44e9e56
4 changed files with 10 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/disaggregation/decode.py
View File

@@ -259,7 +259,7 @@ class DecodePreallocQueue:
if len(req.origin_input_ids) > self.max_total_num_tokens:
message = f"Request {req.rid} exceeds the maximum number of tokens: {len(req.origin_input_ids)} > {self.max_total_num_tokens}"
logger.error(message)
prepare_abort(req, message)
prepare_abort(req, message, status_code=HTTPStatus.BAD_REQUEST)
self.scheduler.stream_output([req], req.return_logprob)
return True
return False

2
python/sglang/srt/disaggregation/prefill.py
View File

@@ -178,7 +178,7 @@ class PrefillBootstrapQueue:
if len(req.origin_input_ids) > self.max_total_num_tokens:
message = f"Request {req.rid} exceeds the maximum number of tokens: {len(req.origin_input_ids)} > {self.max_total_num_tokens}"
logger.error(message)
prepare_abort(req, message)
prepare_abort(req, message, status_code=HTTPStatus.BAD_REQUEST)
self.scheduler.stream_output([req], req.return_logprob)
return True
return False

2
python/sglang/srt/managers/scheduler.py
View File

@@ -1141,7 +1141,7 @@ class Scheduler(
f"boostrap room id. {req.rid=}"
)
logger.error(error_msg)
prepare_abort(req, error_msg)
prepare_abort(req, error_msg, status_code=HTTPStatus.BAD_REQUEST)
self.stream_output([req], req.return_logprob)
return

12
python/sglang/srt/managers/tokenizer_manager.py
View File

@@ -782,15 +782,17 @@ class TokenizerManager:
):
raise ValueError(finish_reason["message"])
if (
finish_reason.get("type") == "abort"
and finish_reason.get("status_code")
== HTTPStatus.SERVICE_UNAVAILABLE
if finish_reason.get("type") == "abort" and finish_reason.get(
"status_code"
) in (
HTTPStatus.SERVICE_UNAVAILABLE,
HTTPStatus.INTERNAL_SERVER_ERROR,
):
# This is an abort request initiated by scheduler.
# Delete the key to prevent resending abort request to the scheduler and
# to ensure aborted request state is cleaned up.
del self.rid_to_state[state.obj.rid]
if state.obj.rid in self.rid_to_state:
del self.rid_to_state[state.obj.rid]
# Mark ongoing LoRA request as finished.
if self.server_args.enable_lora and state.obj.lora_path: