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

bugfix[schedule]: Excessive preemption occurs when preempting running requests to schedule new prefill requests. (#12494)

Browse Source 
Signed-off-by: CLFutureX <chenyongqyl@163.com>
This commit is contained in:
PiteXChen
2025-11-30 22:29:26 +08:00
committed by GitHub
parent 0a9d64530d
commit dc7bdc7329
1 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/schedule_policy.py
+12 -7
View File
@@ -667,14 +667,17 @@ class PrefillAdder:
Returns True if preemption was committed, and the new request can be scheduled. Returns True if preemption was committed, and the new request can be scheduled.
""" """
# Iterate running requests to find preemptible requests # Iterate running requests to find preemptible requests
valid_running_reqs = (
r for r in self.running_batch.reqs if r not in self.preempt_list
)
if server_args.schedule_low_priority_values_first: if server_args.schedule_low_priority_values_first:
sorted_running_reqs = sorted( sorted_valid_running_reqs = sorted(
self.running_batch.reqs, valid_running_reqs,
key=lambda x: (-x.priority, -x.time_stats.wait_queue_entry_time), key=lambda x: (-x.priority, -x.time_stats.wait_queue_entry_time),
) )
else: else:
sorted_running_reqs = sorted( sorted_valid_running_reqs = sorted(
self.running_batch.reqs, valid_running_reqs,
key=lambda x: (x.priority, -x.time_stats.wait_queue_entry_time), key=lambda x: (x.priority, -x.time_stats.wait_queue_entry_time),
) )
preemptible_reqs = [] preemptible_reqs = []
@@ -683,9 +686,7 @@ class PrefillAdder:
+ min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS) + min(req.sampling_params.max_new_tokens, CLIP_MAX_NEW_TOKENS)
- self.rem_total_tokens - self.rem_total_tokens
) )
for running_req in sorted_running_reqs: for running_req in sorted_valid_running_reqs:
if running_req in self.preempt_list:
continue
# Priority difference needs to meet the threshold to be preemptible. # Priority difference needs to meet the threshold to be preemptible.
priority_diff = req.priority - running_req.priority priority_diff = req.priority - running_req.priority
if server_args.schedule_low_priority_values_first: if server_args.schedule_low_priority_values_first:
@@ -695,6 +696,10 @@ class PrefillAdder:
min_tokens_to_remove -= self._get_running_request_total_token_offset( min_tokens_to_remove -= self._get_running_request_total_token_offset(
running_req running_req
) )
if min_tokens_to_remove <= 0:
break
else:
break
# Check max token count limit can be met # Check max token count limit can be met
if len(preemptible_reqs) == 0 or min_tokens_to_remove > 0: if len(preemptible_reqs) == 0 or min_tokens_to_remove > 0: