Remove PanguProMoEV1 related code

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

Author: Pr0Wh1teGivee

vllm_ascend/ops/fused_moe/moe_comm_method.py

@@ -30,7 +30,7 @@
     PrepareAndFinalizeWithMC2, QuantType)
 from vllm_ascend.ops.fused_moe.token_dispatcher import (
     TokenDispatcherWithAll2AllV, TokenDispatcherWithAllGather,
-    TokenDispatcherWithMC2, TokenDispatcherWithMoge)
+    TokenDispatcherWithMC2)
 
 _MoECommMethods: Dict[Optional[MoECommType], MoECommMethod] = {}
 
@@ -198,16 +198,10 @@ class AllGatherCommImpl(MoECommMethod):
     """
 
     def _get_token_dispatcher(self):
-        if self.model_type == "PanguProMoE":
-            return TokenDispatcherWithMoge(
-                top_k=self.moe_config.experts_per_token,
-                num_experts=self.moe_config.num_experts,
-                num_local_experts=self.moe_config.num_local_experts)
-        else:
-            return TokenDispatcherWithAllGather(
-                top_k=self.moe_config.experts_per_token,
-                num_experts=self.moe_config.num_experts,
-                num_local_experts=self.moe_config.num_local_experts)
+        return TokenDispatcherWithAllGather(
+            top_k=self.moe_config.experts_per_token,
+            num_experts=self.moe_config.num_experts,
+            num_local_experts=self.moe_config.num_local_experts)
 
     def _get_prepare_finalize(self):
         return PrepareAndFinalizeWithAllGather(self.moe_config)

vllm_ascend/ops/fused_moe/token_dispatcher.py

@@ -422,69 +422,6 @@ def token_combine(self,
         return final_hidden_states
 
 
-# mypy: disable-error-code="override"
-class TokenDispatcherWithMoge(MoETokenDispatcher):
-
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self.apply_router_weight_on_input = False
-        self.local_num_experts = self.num_experts // self.ep_size
-        self.local_num_group = self.top_k // self.ep_size
-        self.bsz = None
-
-    def token_dispatch(self,
-                       hidden_states: torch.Tensor,
-                       topk_weights: torch.Tensor,
-                       topk_ids: torch.Tensor,
-                       expert_map: Optional[torch.Tensor] = None,
-                       log2phy: Optional[torch.Tensor] = None,
-                       global_redundant_expert_num: int = 0,
-                       shared_experts: Optional[Any] = None,
-                       quantized_x_for_share: Optional[Any] = None,
-                       dynamic_scale_for_share: Optional[Any] = None,
-                       mc2_mask: Optional[torch.Tensor] = None,
-                       apply_router_weight_on_input: bool = False,
-                       with_quant: bool = False,
-                       dynamic_eplb: bool = False,
-                       pertoken_scale: Optional[torch.Tensor] = None):
-        self.bsz, _ = hidden_states.shape
-        flatten_topk_ids = topk_ids.view(-1)
-        self.sorted_topk_ids = torch.argsort(flatten_topk_ids.float())
-        self.sorted_topk_ids = self.sorted_topk_ids.to(torch.int32)
-        sorted_hidden_states = hidden_states.index_select(
-            0, self.sorted_topk_ids // self.local_num_group)
-
-        experts_id = torch.arange(0,
-                                  self.local_num_experts,
-                                  dtype=topk_ids.dtype,
-                                  device=topk_ids.device)
-        num_tokens_per_expert = (
-            flatten_topk_ids.unsqueeze(-1) == experts_id).to(
-                torch.float32).sum(0)
-        topk_scales = topk_weights.view(-1).index_select(
-            0, self.sorted_topk_ids).unsqueeze(-1)
-        group_list = num_tokens_per_expert.cumsum(dim=0).to(torch.int64)
-        group_list_type = 0
-        return {
-            "group_list_type": group_list_type,
-            "hidden_states": sorted_hidden_states,
-            "group_list": group_list,
-            "topk_scales": topk_scales
-        }
-
-    def token_combine(self,
-                      hidden_states: torch.Tensor,
-                      context_metadata: dict,
-                      bias: torch.Tensor = None):
-        unsorted_topk_ids = torch.argsort(self.sorted_topk_ids.float()).to(
-            torch.int32)
-        unsorted_hidden_states = hidden_states.index_select(
-            0, unsorted_topk_ids)
-        final_hidden_states = unsorted_hidden_states.reshape(
-            self.bsz, self.top_k // self.ep_size, -1).sum(1)
-        return final_hidden_states
-
-
 class TokenDispatcherWithAll2AllV(MoETokenDispatcher):
     """
     The implementation of the AlltoAll-based token dispatcher, which handles token

vllm_ascend/worker/model_runner_v1.py

@@ -1418,7 +1418,6 @@ def _select_moe_comm_method(self,
         quant_type = getattr(
             self.vllm_config.model_config.hf_config, 'moe_quantize',
             getattr(self.vllm_config.model_config.hf_config, 'quantize', None))
-        model_type = self.vllm_config.model_config.hf_config.model_type
 
         if not self.parallel_config.enable_expert_parallel:
             moe_comm_type = MoECommType.ALLGATHER
@@ -1445,10 +1444,6 @@ def _select_moe_comm_method(self,
         else:
             raise ValueError(f"Unsupported soc_version: {soc_version}")
 
-        # PanguProMoE only supports allgather
-        if model_type == "PanguProMoE":
-            moe_comm_type = MoECommType.ALLGATHER
-
         if is_global_first_rank():
             logger.debug(f"num_tokens: {num_tokens}, "
                          f"moe_comm_type: {moe_comm_type}")