minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix: avoid double KV release on disaggregated prefill grammar errors (#30937)

Browse Source 
Co-authored-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Kevin Flansburg
2026-07-14 20:16:58 +08:00
committed by GitHub
co-authored by Shangming Cai
parent 43241b7f3f
commit bbb5702a3c
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/disaggregation/prefill.py
+6 -3
View File
@@ -693,7 +693,6 @@ class SchedulerDisaggregationPrefillMixin:
req.grammar.accept_token(next_token_id) req.grammar.accept_token(next_token_id)
except ValueError as e: except ValueError as e:
error_message = f"Grammar accept_token failed for req {req.rid} with token {next_token_id}: {e}" error_message = f"Grammar accept_token failed for req {req.rid} with token {next_token_id}: {e}"
release_kv_cache(req, self.tree_cache)
prepare_abort( prepare_abort(
req, req,
error_message, error_message,
@@ -812,7 +811,8 @@ class SchedulerDisaggregationPrefillMixin:
undone_reqs.append(req) undone_reqs.append(req)
elif poll == KVPoll.Success: # transfer done elif poll == KVPoll.Success: # transfer done
release_kv_cache(req, self.tree_cache) # unlock the tree release_kv_cache(req, self.tree_cache) # unlock the tree
req.finished_reason = FINISH_LENGTH(length=0) if not isinstance(req.finished_reason, FINISH_ABORT):
req.finished_reason = FINISH_LENGTH(length=0)
# FIXME: clean up req's data in transfer engine # FIXME: clean up req's data in transfer engine
req.disagg_kv_sender.clear() req.disagg_kv_sender.clear()
done_reqs.append(req) done_reqs.append(req)
@@ -884,7 +884,10 @@ class SchedulerDisaggregationPrefillMixin:
logger.warning(error_message) logger.warning(error_message)
req.time_stats.trace_ctx.abort(abort_info={"reason": error_message}) req.time_stats.trace_ctx.abort(abort_info={"reason": error_message})
release_kv_cache(req, self.tree_cache) # unlock the tree release_kv_cache(req, self.tree_cache) # unlock the tree
prepare_abort(req, error_message, status_code=HTTPStatus.INTERNAL_SERVER_ERROR) if not isinstance(req.finished_reason, FINISH_ABORT):
prepare_abort(
req, error_message, status_code=HTTPStatus.INTERNAL_SERVER_ERROR
)
if self.metrics_reporter.enable_metrics: if self.metrics_reporter.enable_metrics:
self.metrics_collector.increment_transfer_failed_reqs() self.metrics_collector.increment_transfer_failed_reqs()
return exc return exc