fc1 for glm

Author: unknown

vllm_ascend/models/glm4_moe.py

@@ -76,7 +76,7 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
             final_hidden_states = (
                 self.experts.maybe_all_reduce_tensor_model_parallel(
                     final_hidden_states))
-        return final_hidden_states.view(num_tokens, hidden_dim)
+        return final_hidden_states.view(-1, hidden_dim)
 
 
 class CustomGlm4MoeDecoderLayer(nn.Module):