Eagle3 DP attention for Qwen3 MoE (#12002)

This commit is contained in:
Rain H
2025-10-29 20:25:17 +08:00
committed by GitHub
parent 42f8ea4030
commit 750940ae36
9 changed files with 219 additions and 27 deletions
+1 -1
View File
@@ -822,7 +822,7 @@ class ServerArgs:
capture_bs = (
list(range(1, 9, 1))
+ list(range(10, 33, 2))
+ list(range(40, 64, 4))
+ list(range(40, 65, 4))
+ list(range(72, 257, 8))
+ list(range(272, self.cuda_graph_max_bs + 1, 16))
)