From beca88e59ea75a7aa1af72a5ae50188fa91d4e3d Mon Sep 17 00:00:00 2001 From: Ziming Huang Date: Wed, 5 Aug 2026 17:22:55 +0800 Subject: [PATCH] [BugFix][K3] Skip moe_intermediate padding when EP is enabled (#51131) --- vllm/models/kimi_k3/nvidia/model.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/models/kimi_k3/nvidia/model.py b/vllm/models/kimi_k3/nvidia/model.py index 7e06e23107..a008d17cb6 100644 --- a/vllm/models/kimi_k3/nvidia/model.py +++ b/vllm/models/kimi_k3/nvidia/model.py @@ -494,7 +494,7 @@ class KimiMoE(nn.Module): min_moe_intermediate_per_partition = getattr( config, "min_moe_intermediate_per_partition", 256 ) - if self.tp_size > 1: + if self.tp_size > 1 and not vllm_config.parallel_config.enable_expert_parallel: moe_intermediate_per_partition = moe_intermediate_size // self.tp_size if moe_intermediate_per_partition < min_moe_intermediate_per_partition: self.padded_moe_intermediate_size = (