[Fix] Add _MOE_TP to graph_capture for MoE models with ep>1 (#21907)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
This commit is contained in:
Kangyan-Zhou
2026-04-03 02:33:16 -07:00
committed by GitHub
co-authored by Claude Opus 4.6 Baizhou Zhang
parent ac593fed90
commit 56ac9c9932
2 changed files with 9 additions and 5 deletions
+3
View File
@@ -1133,6 +1133,9 @@ class ServerArgs:
# 16. Expert distribution recorder
if self.enable_eplb or self.expert_distribution_recorder_mode is not None:
self.disable_piecewise_cuda_graph = True
# 17. Context parallel
if self.attn_cp_size > 1:
self.disable_piecewise_cuda_graph = True
def _handle_gpu_memory_settings(self, gpu_mem):
"""