fix-precommit & mypy

Signed-off-by: daishixun <[email protected]>

Author: dsxsteven

vllm_ascend/eplb/adaptor/vllm_adaptor.py

@@ -64,7 +64,6 @@ def __init__(self, model, mtp_instance, num_mtp_layers, **args):
             ]
         else:
             self.mtp_expert_weight_names = ["w13_weight", "w2_weight"]
-        
 
         self.expert_map_per_layer = dict(
         )  # reference to expert map on device for expert map update
@@ -79,7 +78,7 @@ def __init__(self, model, mtp_instance, num_mtp_layers, **args):
             for mtp_layer_idx in range(self.num_mtp_layers):
                 self.expert_map_per_layer[self.num_dense_layers + self.num_moe_layers + mtp_layer_idx] = \
                     self.mtp_instance.model.get_expert_map(self.num_dense_layers + self.num_moe_layers + mtp_layer_idx)
-        
+
         # TODO: here we set number of buffer tensor equal to number of expert in each laryer, which can be improved
         num_buffer_tensor = torch.where(
             self.expert_map_per_layer[self.num_dense_layers] != -1)[0].numel()
@@ -95,7 +94,7 @@ def __init__(self, model, mtp_instance, num_mtp_layers, **args):
         for layer_idx in range(self.num_moe_layers):
             self.log2phy_map_per_layer[self.num_dense_layers + layer_idx] = \
                 self.model.get_log2phy_map(self.num_dense_layers + layer_idx)
-                
+
         if self.mtp_instance is not None:
             for mtp_layer_idx in range(self.num_mtp_layers):
                 self.log2phy_map_per_layer[self.num_dense_layers + self.num_moe_layers + mtp_layer_idx] = \
@@ -127,7 +126,7 @@ def init_expert_param_per_layer(self):
                                     name].data[local_expert_id]
                     for name in self.expert_weight_names
                 ])
-                
+
         if self.mtp_instance is not None:
             mtp_param_dict = dict(self.mtp_instance.named_parameters())
             self.expert_param_per_layer[self.num_dense_layers +
@@ -153,8 +152,7 @@ def get_rank_expert_workload(self) -> torch.Tensor:
                 self.moe_load,
                 self.mtp_instance.model.get_all_moe_loads().to(
                     device=self.moe_load.device)
-            ],
-                                    dim=0)
+            ], dim=0)
         return self.moe_load
 
     def get_init_expert_map(self, num_moe_layers):
@@ -164,8 +162,7 @@ def get_init_expert_map(self, num_moe_layers):
                 expert_map,
                 self.mtp_instance.model.get_all_expert_map().to(
                     device=expert_map.device)
-            ],
-                                    dim=0)
+            ], dim=0)
         if dist.is_initialized():
             world_size = dist.get_world_size()
 

vllm_ascend/eplb/utils.py

@@ -85,8 +85,7 @@ def clear_all_moe_loads(self):
         for layer_id in range(self.mtp_start_layer_idx,
                               self.mtp_start_layer_idx + self.num_mtp_layers):
             self.layers[str(layer_id)].mtp_block.mlp.experts.clear_moe_load()
-    
-    
+
 
 def model_register(model, model_config):
     model.get_expert_map = types.MethodType(get_expert_map, model)

vllm_ascend/worker/model_runner_v1.py

@@ -3169,7 +3169,7 @@ def load_model(self) -> None:
                                       MtpProposer) and isinstance(
                                           self.drafter.model, DeepSeekMTP)
                     mtp_instance = self.drafter.model
-                model_register(mtp_instance.model, self.vllm_config)
+                    model_register(mtp_instance.model, self.vllm_config)
                 if self.drafter.name == SpecDcodeType.EAGLE3:
                     self.model.set_aux_hidden_state_layers(
                         self.model.get_eagle3_aux_hidden_state_layers())