Modularize fused experts and integrate PPLX kernels (#15956)

vllm/model_executor/models/qwen3_moe.py

@@ -30,9 +30,7 @@ from transformers import PretrainedConfig
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
-from vllm.distributed import (get_pp_group,
-                              get_tensor_model_parallel_world_size,
-                              tensor_model_parallel_all_reduce)
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -137,7 +135,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
                                            router_logits=router_logits)
         final_hidden_states = final_hidden_states
         if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(
+            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
                 final_hidden_states)
 
         return final_hidden_states.view(orig_shape)