From a2b6e6686b2e70a3b56dd2b4d99eeb8381ab2b9d Mon Sep 17 00:00:00 2001 From: shuyixiong <219646547+shuyixiong@users.noreply.github.com> Date: Wed, 1 Jul 2026 01:58:49 -0700 Subject: [PATCH] Fix moe backend setting Signed-off-by: shuyixiong <219646547+shuyixiong@users.noreply.github.com> --- tensorrt_llm/_torch/modules/fused_moe/create_moe.py | 5 ----- 1 file changed, 5 deletions(-) diff --git a/tensorrt_llm/_torch/modules/fused_moe/create_moe.py b/tensorrt_llm/_torch/modules/fused_moe/create_moe.py index f7b713e3de61..cfeab73b20a3 100644 --- a/tensorrt_llm/_torch/modules/fused_moe/create_moe.py +++ b/tensorrt_llm/_torch/modules/fused_moe/create_moe.py @@ -213,11 +213,6 @@ def resolve_moe_cls( moe_cls = get_moe_cls(model_config, override_quant_config, layer_idx) effective_quant_config = override_quant_config or model_config.quant_config - has_quant = (effective_quant_config is not None - and effective_quant_config.layer_quant_mode.has_any_quant( - exclude_kv_cache=True)) - if (moe_cls == TRTLLMGenFusedMoE and not has_quant): - moe_cls = CutlassFusedMoE # Routed-expert LoRA is supported only on CutlassFusedMoE with unquantized # base weights. Fail loudly here rather than at runtime if the user-selected