fix tests

Signed-off-by: yewentao256 <zhyanwentao@126.com>

tests/kernels/moe/modular_kernel_tools/common.py

@@ -594,7 +594,9 @@ def make_modular_kernel(
     )
 
     modular_kernel = mk.FusedMoEModularKernel(
-        prepare_finalize=prepare_finalize, fused_experts=fused_experts
+        prepare_finalize=prepare_finalize,
+        fused_experts=fused_experts,
+        moe_parallel_config=moe_parallel_config,
     )
 
     return modular_kernel

vllm/model_executor/layers/fused_moe/fused_moe_modular_method.py

@@ -43,7 +43,6 @@ class FusedMoEModularMethod(FusedMoEMethodBase, CustomOp):
         prepare_finalize: FusedMoEPrepareAndFinalize,
         shared_experts: torch.nn.Module | None,
     ) -> "FusedMoEModularMethod":
-        moe_parallel_config = getattr(moe_layer, "moe_parallel_config", None)
         return FusedMoEModularMethod(
             old_quant_method,
             FusedMoEModularKernel(
@@ -51,7 +50,7 @@ class FusedMoEModularMethod(FusedMoEMethodBase, CustomOp):
                 old_quant_method.select_gemm_impl(prepare_finalize, moe_layer),
                 shared_experts,
                 getattr(moe_layer, "shared_experts_stream", None),
-                moe_parallel_config=moe_parallel_config,
+                moe_parallel_config=moe_layer.moe_parallel_config,
             ),
         )
 

vllm/model_executor/layers/fused_moe/unquantized_fused_moe_method.py

@@ -334,7 +334,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 apply_router_weight_on_input=layer.apply_router_weight_on_input,
                 global_num_experts=layer.global_num_experts,
                 expert_map=layer.expert_map,
-                moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                moe_parallel_config=layer.moe_parallel_config,
             )
 
         if layer.zero_expert_num != 0 and layer.zero_expert_type is not None:

vllm/model_executor/layers/quantization/bitsandbytes.py

@@ -521,7 +521,7 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
             global_num_experts=layer.global_num_experts,
             expert_map=layer.expert_map,
             quant_config=self.moe_quant_config,
-            moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+            moe_parallel_config=layer.moe_parallel_config,
         )
 
     def _create_weights_4bit(

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

@@ -1244,7 +1244,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
                     if self.disable_expert_map
                     else layer.expert_map,  # ???
                     quant_config=self.moe_quant_config,
-                    moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                    moe_parallel_config=layer.moe_parallel_config,
                 )
             else:
                 from vllm.model_executor.layers.fused_moe.cutlass_moe import (
@@ -1267,7 +1267,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
                     ab_strides2=self.ab_strides2,
                     c_strides1=self.c_strides1,
                     c_strides2=self.ab_strides1_c_strides2,
-                    moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                    moe_parallel_config=layer.moe_parallel_config,
                 )
 
         else:
@@ -1287,7 +1287,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
                 global_num_experts=layer.global_num_experts,
                 expert_map=layer.expert_map,
                 quant_config=self.moe_quant_config,
-                moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                moe_parallel_config=layer.moe_parallel_config,
             )
 
     @property
@@ -1426,7 +1426,7 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
             global_num_experts=layer.global_num_experts,
             expert_map=layer.expert_map,
             quant_config=self.moe_quant_config,
-            moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+            moe_parallel_config=layer.moe_parallel_config,
         )
 
 
@@ -2015,7 +2015,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
             global_num_experts=layer.global_num_experts,
             expert_map=layer.expert_map,
             quant_config=self.moe_quant_config,
-            moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+            moe_parallel_config=layer.moe_parallel_config,
         )
 
     @property

vllm/model_executor/layers/quantization/experts_int8.py

@@ -159,7 +159,7 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
             global_num_experts=layer.global_num_experts,
             expert_map=layer.expert_map,
             quant_config=self.moe_quant_config,
-            moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+            moe_parallel_config=layer.moe_parallel_config,
         )
 
     @staticmethod

vllm/model_executor/layers/quantization/fp8.py

@@ -1376,7 +1376,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 allow_cutlass_block_scaled_grouped_gemm=(
                     self.allow_cutlass_block_scaled_grouped_gemm
                 ),
-                moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                moe_parallel_config=layer.moe_parallel_config,
             )
 
         if layer.zero_expert_num != 0 and layer.zero_expert_type is not None:

vllm/model_executor/layers/quantization/modelopt.py

@@ -768,7 +768,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
                 global_num_experts=layer.global_num_experts,
                 expert_map=layer.expert_map,
                 apply_router_weight_on_input=layer.apply_router_weight_on_input,
-                moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                moe_parallel_config=layer.moe_parallel_config,
             )
 
 

vllm/model_executor/layers/quantization/moe_wna16.py

@@ -381,7 +381,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
             global_num_experts=layer.global_num_experts,
             expert_map=layer.expert_map,
             quant_config=self.moe_quant_config,
-            moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+            moe_parallel_config=layer.moe_parallel_config,
         )
 
     @staticmethod

vllm/model_executor/layers/quantization/quark/quark_moe.py

@@ -393,7 +393,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
                 global_num_experts=layer.global_num_experts,
                 expert_map=layer.expert_map,
                 quant_config=self.moe_quant_config,
-                moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                moe_parallel_config=layer.moe_parallel_config,
             )
 
 
@@ -618,7 +618,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
                 apply_router_weight_on_input=layer.apply_router_weight_on_input,
                 expert_map=layer.expert_map,
                 quant_config=self.moe_quant_config,
-                moe_parallel_config=getattr(layer, "moe_parallel_config", None),
+                moe_parallel_config=layer.moe_parallel_config,
             )
 
         return out

vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

@@ -247,11 +247,6 @@ def flashinfer_cutlass_moe_fp8(
     assert quant_config is not None
 
     # Construct modular kernel with block-scale support when requested.
-    moe_parallel_config = getattr(
-        moe if moe is not None else layer,
-        "moe_parallel_config",
-        None,
-    )
     fused_experts = mk.FusedMoEModularKernel(
         build_flashinfer_fp8_cutlass_moe_prepare_finalize(
             moe=moe, use_deepseek_fp8_block_scale=use_deepseek_fp8_block_scale
@@ -262,7 +257,7 @@ def flashinfer_cutlass_moe_fp8(
             out_dtype=hidden_states.dtype,
             use_deepseek_fp8_block_scale=use_deepseek_fp8_block_scale,
         ),
-        moe_parallel_config=moe_parallel_config,
+        moe_parallel_config=layer.moe_parallel_config,
     )
 
     return fused_experts(