[EPLB][Bugfix] Dispatch Allgather use log2phy if enable eplb (#5933)

### What this PR does / why we need it?
1. Move the logic of expert mapping forward to prevent shotgun changes
2. Disable the update of expert map.

### How was this patch tested?
a2
| dataset | version | metric | mode | vllm-api-general-chat |
|----- | ----- | ----- | ----- | -----|
| GPQA_diamond | 53064e | accuracy | gen | 73.23 |

a3
| dataset | version | metric | mode | vllm-api-general-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 83.33 |


- vLLM version: v0.13.0
- vLLM main:
11b6af5280

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>

vllm_ascend/eplb/adaptor/vllm_adaptor.py

@@ -33,12 +33,7 @@ class VllmEplbAdaptor(EplbAdaptor):
         self.rank_id = dist.get_rank()
         self.world_size = dist.get_world_size()
         self.param_dict = dict(self.model.named_parameters())
-        if self.model.config.model_type == "qwen3_moe":
-            self.num_dense_layers = 0
-            self.global_expert_num = self.model.config.num_experts
-        else:
-            self.num_dense_layers = self.model.config.first_k_dense_replace
-            self.global_expert_num = self.model.config.n_routed_experts
+        self.num_dense_layers = getattr(self.model.config, "first_k_dense_replace", 0)
         self.num_moe_layers = self.model.config.num_hidden_layers - self.num_dense_layers
 
         for i in range(self.num_dense_layers,
@@ -64,17 +59,10 @@ class VllmEplbAdaptor(EplbAdaptor):
         else:
             self.expert_weight_names = ["w13_weight", "w2_weight"]
 
-        self.expert_map_per_layer = dict(
-        )  # reference to expert map on device for expert map update
         self.expert_map_per_layer_cpu = dict(
         )  # copy of expert map on CPU to avoid device synchronize frequently
-        for layer_idx in range(self.num_moe_layers):
-            self.expert_map_per_layer[self.num_dense_layers + layer_idx] = \
-                self.model.get_expert_map(self.num_dense_layers + 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()
+        num_buffer_tensor = self.model.model.layers[-1].mlp.experts.local_num_experts
         self.buffer_tensor_list: list[list[Any]] = [
             [] for _ in range(num_buffer_tensor)
         ]
@@ -88,8 +76,6 @@ class VllmEplbAdaptor(EplbAdaptor):
             self.log2phy_map_per_layer[self.num_dense_layers + layer_idx] = \
                 self.model.get_log2phy_map(self.num_dense_layers + layer_idx)
 
-        self.all_topk_ids = []
-
     def init_buffer_tensor(self, num_buffer_tensor):
         for buffer_id in range(num_buffer_tensor):
             for name in self.expert_weight_names:
@@ -169,7 +155,6 @@ class VllmEplbAdaptor(EplbAdaptor):
                 json.dump(record, f, indent=4)
 
     def do_update_expert_map(self, layer_id, updated_expert_map):
-        self.expert_map_per_layer[layer_id].copy_(updated_expert_map)
         self.expert_map_per_layer_cpu[layer_id].copy_(updated_expert_map)
 
     def do_update_expert_weight(self, layer_id, local_expert_to_replace,

vllm_ascend/eplb/eplb_updator.py

@@ -38,7 +38,6 @@ class EplbUpdator:
     def set_adaptor(self, adaptor):
         self.adaptor = adaptor
         self.num_moe_layers = self.adaptor.num_moe_layers
-        self.global_expert_num = self.adaptor.global_expert_num
 
     def init_eplb(self, expert_map_path, process):
         self.rank_id = dist.get_rank()