From 4c1d9d0c10a21cc3596f38ff36f481c2db3f7386 Mon Sep 17 00:00:00 2001 From: xxi <95731198+xxi-nv@users.noreply.github.com> Date: Thu, 5 Feb 2026 15:07:59 +0800 Subject: [PATCH] [None][chore] Pass without_comm to cutlass and deepgemm (#11229) Signed-off-by: xxi --- tensorrt_llm/_torch/modules/fused_moe/create_moe.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tensorrt_llm/_torch/modules/fused_moe/create_moe.py b/tensorrt_llm/_torch/modules/fused_moe/create_moe.py index 1818daab7e..99a6847461 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/create_moe.py +++ b/tensorrt_llm/_torch/modules/fused_moe/create_moe.py @@ -196,6 +196,7 @@ def create_moe_backend( swiglu_beta=swiglu_beta, swiglu_limit=swiglu_limit, init_load_balancer=init_load_balancer, + without_comm=without_comm, activation_type=activation_type, ) elif moe_cls == WideEPMoE: @@ -256,6 +257,7 @@ def create_moe_backend( weight_loading_mode=weight_loading_mode, apply_router_weight_on_input=apply_router_weight_on_input, layer_idx=layer_idx, + without_comm=without_comm, ) elif moe_cls == TritonFusedMoE: assert not apply_router_weight_on_input, "apply_router_weight_on_input is not supported in TritonFusedMoE."