[fix] Fix divide by zero error for llama4. (#8683)

This commit is contained in:
Varun Vinayak Shenoy
2025-08-03 00:55:55 -07:00
committed by GitHub
parent 32f2815451
commit 7ed8e51bc3

View File

@@ -415,6 +415,8 @@ class Llama4DecoderLayer(nn.Module):
)
def _is_moe_layer(self, layer_id: int) -> bool:
if self.config.interleave_moe_layer_step == 0:
return self.config.num_local_experts > 0
return (layer_id + 1) % self.config.interleave_moe_layer_step == 0
def forward(