bugfix

Signed-off-by: Dingli <dingli40@huawei.com>

Author: Dingli

vllm_ascend/models/deepseek_v2.py

@@ -228,7 +228,7 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
 
         cv_parallel = VLLM_ENABLE_CV_PARALLEL and not is_prefill
 
-        if self.n_shared_experts is not None
+        if self.n_shared_experts is not None:
             if not cv_parallel:
                 shared_output = self.shared_experts(hidden_states)
             else:

vllm_ascend/quantization/w8a8_dynamic.py

@@ -127,7 +127,7 @@ def apply_mlp(hidden_states_wrapper: List[torch.Tensor],
                 output_dtype=torch.bfloat16,
             )
             if shared_experts.down_proj.reduce_results and shared_experts.down_proj.tp_size > 1:
-                shared_output = tensor_model_parallel_all_reduce(x)
+                shared_output = tensor_model_parallel_all_reduce(shared_output)
     if shared_experts:
         return hidden_states, shared_output
     return hidden_states