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

[bugfix] fix bug when enable prefill delay and DP (#20134)

Browse Source
This commit is contained in:
LinyuanLi
2026-03-13 08:51:09 +08:00
committed by GitHub
parent c21ddbc785
commit 9865f11421
1 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/managers/prefill_delayer.py
View File

@@ -55,12 +55,14 @@ class PrefillDelayer:
)
# The global_info contains four pieces of information:
# prefillable, token_watermark_force_allow, running_batch, and max_prefill_bs.
self.dp_size = dp_size
self.enable_dp_attention = server_args.enable_dp_attention
dp_size_dim = dp_size if self.enable_dp_attention else 1
self._global_info_buffer = torch.empty(
(dp_size, attn_tp_size, 4),
(dp_size_dim, attn_tp_size, 4),
dtype=torch.int64,
device=device,
)
self.enable_dp_attention = server_args.enable_dp_attention
self._cpu_group = cpu_group
self._metrics_collector = metrics_collector
@@ -144,6 +146,10 @@ class PrefillDelayer:
)
max_running_requests = kwargs.get("max_running_requests", 0)
if not self.enable_dp_attention:
max_running_requests = (
max_running_requests + self.dp_size - 1
) // self.dp_size
if (
max_running_requests - global_running_batch.max().item()
< global_max_prefill_bs.max().item()