[BugFix][K3] Skip moe_intermediate padding when EP is enabled (#51131)
Ziming Huang committed
Aug 5, 2026 at 17:22 UTC
beca88e59ea75a7aa1af72a5ae50188fa91d4e3d
1 file changed
+1
-1
vllm/models/kimi_k3/nvidia/model.py
+1
-1
@@ -494,7 +494,7 @@ class KimiMoE(nn.Module):
494
min_moe_intermediate_per_partition = getattr(
495
config, "min_moe_intermediate_per_partition", 256
496
)
497
- if self.tp_size > 1:
497
+ if self.tp_size > 1 and not vllm_config.parallel_config.enable_expert_parallel:
498
moe_intermediate_per_partition = moe_intermediate_size // self.tp_size
499
if moe_intermediate_per_partition < min_moe_intermediate_per_partition:
500
self.padded_moe_intermediate_size = (