chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[DP-Attn] Clarify MLP sync / idle batch preparation logic (#12843)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-11-08 23:23:14 +08:00
committed by GitHub
parent 6fee2c535c
commit 243ea585fc
7 changed files with 229 additions and 177 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/two_batch_overlap.py
View File

@@ -407,8 +407,8 @@ class TboDPAttentionPreparer:
return local_can_run_tbo, local_forward_mode
def compute_output(self, partial_global_info):
local_can_run_tbo_aggregated = min(partial_global_info[:, 0, 0].tolist())
forward_modes = partial_global_info[:, 0, 1].tolist()
local_can_run_tbo_aggregated = min(partial_global_info[:, 0].tolist())
forward_modes = partial_global_info[:, 1].tolist()
global_forward_mode, forward_mode_agree = self._compute_global_forward_mode(
forward_modes