Set the default value of routed_scaling_factor to 1

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1392,7 +1392,7 @@ def inplace_fused_experts(hidden_states: torch.Tensor,
                           block_shape: Optional[List[int]] = None,
                           use_nn_moe: Optional[bool] = False,
                           shared_output: Optional[torch.Tensor] = None,
-                          routed_scaling_factor: Optional[float] = None) -> None:
+                          routed_scaling_factor: Optional[float] = 1.0) -> None:
     fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids, True,
                        activation, apply_router_weight_on_input, use_fp8_w8a8,
                        use_int8_w8a8, use_int8_w8a16, use_int4_w4a16,use_int4_w4a8,
@@ -1426,7 +1426,7 @@ def inplace_fused_experts_fake(
         block_shape: Optional[List[int]] = None,
         use_nn_moe: Optional[bool] = False,
         shared_output: Optional[torch.Tensor] = None,
-        routed_scaling_factor: Optional[float] = None) -> None:
+        routed_scaling_factor: Optional[float] = 1.0) -> None:
     pass
 
 
@@ -1464,7 +1464,7 @@ def outplace_fused_experts(
         block_shape: Optional[List[int]] = None,
         use_nn_moe: Optional[bool] = False,
         shared_output: Optional[torch.Tensor] = None,
-        routed_scaling_factor: Optional[float] = None) -> torch.Tensor:
+        routed_scaling_factor: Optional[float] = 1.0) -> torch.Tensor:
     return fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids,
                               False, activation, apply_router_weight_on_input,
                               use_fp8_w8a8, use_int8_w8a8, use_int8_w8a16,
@@ -1498,7 +1498,7 @@ def outplace_fused_experts_fake(
         block_shape: Optional[List[int]] = None,
         use_nn_moe: Optional[bool] = False,
         shared_output: Optional[torch.Tensor] = None,
-        routed_scaling_factor: Optional[float] = None) -> torch.Tensor:
+        routed_scaling_factor: Optional[float] = 1.0) -> torch.Tensor:
     return torch.empty_like(hidden_states)
 
 
@@ -1557,7 +1557,7 @@ def fused_experts(
         allow_cutlass_block_scaled_grouped_gemm: bool = False,
         use_nn_moe: Optional[bool] = False,
         shared_output: Optional[torch.Tensor] = None,
-        routed_scaling_factor: Optional[float] = None) -> torch.Tensor:
+        routed_scaling_factor: Optional[float] = 1.0) -> torch.Tensor:
     # For now, disable DeepGemm for small N (<= 512) until better
     # permute/unpermute ops are available.
     N = w1.size(1)
@@ -1646,7 +1646,7 @@ def fused_experts_impl(
     block_shape: Optional[List[int]] = None,
     use_nn_moe: Optional[bool] = False,
     shared_output: Optional[torch.Tensor] = None,
-    routed_scaling_factor: Optional[float] = None,
+    routed_scaling_factor: Optional[float] = 1.0,
 ) -> torch.Tensor:
     num_tokens = hidden_states.size(0)
     if use_nn_moe:

vllm/model_executor/layers/fused_moe/layer.py

@@ -376,7 +376,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         logical_replica_count: Optional[torch.Tensor] = None,
         shared_output: Optional[torch.Tensor] = None,
         use_nn_moe: Optional[bool] = False,
-        routed_scaling_factor: Optional[float] = None,
+        routed_scaling_factor: Optional[float] = 1.0,
         use_fused_gate: Optional[bool] = False,
     ) -> torch.Tensor:
         if enable_eplb:
@@ -423,7 +423,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         activation: str = "silu",
         shared_output: Optional[torch.Tensor] = None,
         use_nn_moe: Optional[bool] = False,
-        routed_scaling_factor: Optional[float] = None,
+        routed_scaling_factor: Optional[float] = 1.0,
         use_fused_gate: Optional[bool] = False,
     ) -> torch.Tensor:
 
@@ -487,7 +487,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         apply_router_weight_on_input: bool = False,
         activation: str = "silu",
         use_nn_moe: Optional[bool] = False,
-        routed_scaling_factor: Optional[float] = None,
+        routed_scaling_factor: Optional[float] = 1.0,
         use_fused_gate: Optional[bool] = False,
         **kwargs,
     ):
@@ -527,7 +527,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         apply_router_weight_on_input: bool = False,
         activation: str = "silu",
         use_nn_moe: Optional[bool] = False,
-        routed_scaling_factor: Optional[float] = None,
+        routed_scaling_factor: Optional[float] = 1.0,
         use_fused_gate: Optional[bool] = False,
     ) -> torch.Tensor:
         assert not use_grouped_topk
@@ -683,7 +683,7 @@ class FusedMoE(torch.nn.Module):
         activation: str = "silu",
         enable_eplb: bool = False,
         num_redundant_experts: int = 0,
-        routed_scaling_factor: Optional[float] = None,
+        routed_scaling_factor: Optional[float] = 1.0,
     ):
         super().__init__()
         if params_dtype is None:
@@ -1262,7 +1262,7 @@ class FusedMoE(torch.nn.Module):
         expert_load_view: Optional[torch.Tensor] = None,
         logical_to_physical_map: Optional[torch.Tensor] = None,
         logical_replica_count: Optional[torch.Tensor] = None,
-        routed_scaling_factor: Optional[float] = None,
+        routed_scaling_factor: Optional[float] = 1.0,
         use_fused_gate: Optional[bool] = False
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """