修复ep的auto模式的崩溃bug

vllm/model_executor/layers/fused_moe/layer.py

@@ -911,7 +911,7 @@ class FusedMoE(torch.nn.Module):
         self.batched_hidden_states: Optional[torch.Tensor] = None
         self.batched_router_logits: Optional[torch.Tensor] = None
         if (self.moe_parallel_config.use_pplx_kernels
-                or self.moe_parallel_config.use_deepep_ll_kernels):
+                or self.moe_parallel_config.use_deepep_ll_kernels or self.moe_parallel_config.use_deepep_auto_kernels):
             self.batched_hidden_states = torch.zeros(
                 (moe.max_num_tokens, self.hidden_size),
                 dtype=moe.in_dtype,

vllm/model_executor/models/deepseek_v2.py

@@ -957,7 +957,8 @@ class DeepseekV2Model(nn.Module):
         parallel_config = vllm_config.parallel_config
         self.use_deepep = self.dp_size > 1 and parallel_config.enable_expert_parallel and \
             (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
-             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency")
+             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
+             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
         self.tp_size = get_tensor_model_parallel_world_size()
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: