Merge branch 'v0.15.1_mul_add' into 'v0.15.1-dev'

fix(moe): 补齐非Marlin量化路径 shared_output/routed_scaling_factor 透传

See merge request dcutoolkit/deeplearing/vllm!529

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1881,6 +1881,8 @@ def fused_experts_impl(
                                        a2_scale=a2_scale,
                                        block_shape=block_shape,
                                        use_nn_moe=False,
+                                       routed_scaling_factor=routed_scaling_factor,
+                                       shared_output=shared_output,
                                        i_q=i_q,
                                        i_s=i_s
                                        )
@@ -1903,7 +1905,9 @@ def fused_experts_impl(
                                        a1_scale=a1_scale,
                                        a2_scale=a2_scale,
                                        block_shape=block_shape,
-                                       use_nn_moe=False
+                                       use_nn_moe=False,
+                                       routed_scaling_factor=routed_scaling_factor,
+                                       shared_output=shared_output
                                        )
 
     if use_int4_w4a16:

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -1111,7 +1111,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         topk_ids: torch.Tensor,
         use_nn_moe: bool | None = False,
         use_fused_gate: bool | None = False,
-
+        shared_output: torch.Tensor | None = None,
+        routed_scaling_factor: float = 1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
         assert not self.is_monolithic
         assert self.kernel is not None
@@ -1131,6 +1132,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             quant_config=self.moe_quant_config,
             use_fused_gate=use_fused_gate,
             use_nn_moe=False,
+            shared_output=shared_output,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
     @property
@@ -1256,7 +1259,9 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
         topk_ids: torch.Tensor,
         use_nn_moe: bool | None = False,
         i_q: torch.Tensor | None = None,
-        i_s: torch.Tensor | None = None
+        i_s: torch.Tensor | None = None,
+        shared_output: torch.Tensor | None = None,
+        routed_scaling_factor: float = 1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
         from vllm.model_executor.layers.fused_moe import fused_experts
 
@@ -1274,7 +1279,9 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
             quant_config=self.moe_quant_config,
             use_nn_moe=use_nn_moe,
             i_q=i_q,
-            i_s=i_s
+            i_s=i_s,
+            shared_output=shared_output,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
 
@@ -2515,4 +2522,4 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
 
     @property
     def supports_eplb(self) -> bool:
-        return False
+        return False
\ No newline at end of file

vllm/model_executor/layers/quantization/fp8.py

@@ -1028,6 +1028,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
         topk_ids: torch.Tensor,
         use_nn_moe: bool | None = False,
         use_fused_gate: bool | None = False,
+        shared_output: torch.Tensor | None = None,
+        routed_scaling_factor: float = 1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
         assert self.kernel is not None
         assert not self.is_monolithic
@@ -1047,6 +1049,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
             quant_config=self.moe_quant_config,
             use_fused_gate=use_fused_gate,
             use_nn_moe=False,
+            shared_output=shared_output,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
 

vllm/model_executor/layers/quantization/slimquant_w4a8.py

@@ -308,7 +308,9 @@ class SlimQuantW4A8Int8MoEMethod:
             use_nn_moe: bool | None = False,
             use_fused_gate: bool | None = False,
             i_q: torch.Tensor | None = None,
-            i_s: torch.Tensor | None = None
+            i_s: torch.Tensor | None = None,
+            shared_output: torch.Tensor | None = None,
+            routed_scaling_factor: float = 1.0,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
         from vllm.model_executor.layers.fused_moe import fused_experts
         return fused_experts(
@@ -324,4 +326,6 @@ class SlimQuantW4A8Int8MoEMethod:
             global_num_experts=layer.global_num_experts,
             quant_config=self.moe_quant_config,
             use_nn_moe=use_nn_moe,
+            shared_output=shared_output,
+            routed_scaling_factor=routed_scaling_factor,
         )