[BugFix][K3] Skip moe_intermediate padding when EP is enabled (#51131)
This commit is contained in:
@@ -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 = (
|
||||
|
||||
Reference in New Issue
Block a user