fix

Author: hiworldwzj

lightllm/common/basemodel/layer_weights/meta_weights/fused_moe_weight_tp.py

@@ -67,7 +67,6 @@ def experts(self, input_tensor, router_logits, top_k, renormalize, use_grouped_t
             topk_group=topk_group,
             num_expert_group=num_expert_group,
             scoring_func=self.scoring_func,
-            num_fused_shared_experts=self.num_fused_shared_experts,
         )
         if self.num_fused_shared_experts > 0:
             topk_ids[:, -1] = self.n_routed_experts - 1

lightllm/common/fused_moe/grouped_topk.py

@@ -208,7 +208,6 @@ def triton_grouped_topk(
     topk_group: int = 0,
     scoring_func: str = "softmax",
     group_score_used_topk_num=2,
-    num_fused_shared_experts: int = 0,
 ):
 
     if correction_bias is not None:
@@ -223,8 +222,8 @@ def triton_grouped_topk(
         dtype = torch.float32
 
     scores_buffer = torch.empty((token_num, total_expert_num), dtype=dtype, device="cuda")
-    out_topk_weights = torch.empty((token_num, topk + num_fused_shared_experts), dtype=torch.float32, device="cuda")
-    out_topk_ids = torch.empty((token_num, topk + num_fused_shared_experts), dtype=torch.long, device="cuda")
+    out_topk_weights = torch.empty((token_num, topk), dtype=torch.float32, device="cuda")
+    out_topk_ids = torch.empty((token_num, topk), dtype=torch.long, device="cuda")
 
     assert total_expert_num % num_expert_group == 0
 

lightllm/common/fused_moe/topk_select.py

@@ -175,7 +175,6 @@ def select_experts(
     num_expert_group: Optional[int] = None,
     scoring_func: str = "softmax",
     custom_routing_function: Optional[Callable] = None,
-    num_fused_shared_experts: int = 0,
 ):
     from lightllm.common.fused_moe.topk_select import fused_topk
     from lightllm.common.fused_moe.grouped_topk import triton_grouped_topk
@@ -211,7 +210,6 @@ def select_experts(
                 topk_group=topk_group,
                 scoring_func=scoring_func,
                 group_score_used_topk_num=group_score_topk_num,
-                num_fused_shared_experts=num_fused_shared_experts,
             )
 
     elif custom_routing_function is None: