[MoE Refactor] Rename "naive" all2all backend (#36294)

Signed-off-by: Bill Nell <bnell@redhat.com>

vllm/model_executor/layers/fused_moe/all2all_utils.py

@@ -229,7 +229,7 @@ def maybe_make_prepare_finalize(
             num_dispatchers=all2all_manager.world_size,
         )
 
-    elif moe.use_naive_all2all_kernels and allow_new_interface:
+    elif moe.use_ag_rs_all2all_kernels and allow_new_interface:
         prepare_finalize = make_moe_prepare_and_finalize_naive_dp_ep(
             use_monolithic=use_monolithic,
             is_sequence_parallel=moe.moe_parallel_config.is_sequence_parallel,

vllm/model_executor/layers/fused_moe/config.py

@@ -975,9 +975,10 @@ class FusedMoEParallelConfig:
         return self.use_deepep_ll_kernels
 
     @property
-    def use_naive_all2all_kernels(self):
-        return self.use_all2all_kernels and (
-            self.all2all_backend in ["naive", "allgather_reducescatter"]
+    def use_ag_rs_all2all_kernels(self):
+        return (
+            self.use_all2all_kernels
+            and self.all2all_backend == "allgather_reducescatter"
         )
 
     @property
@@ -1143,7 +1144,7 @@ class FusedMoEParallelConfig:
             ep_rank=0,
             sp_size=1,
             use_ep=False,
-            all2all_backend="naive",
+            all2all_backend="allgather_reducescatter",
             enable_eplb=False,
         )
 
@@ -1256,8 +1257,8 @@ class FusedMoEConfig:
         return self.moe_parallel_config.use_fi_nvl_one_sided_kernels
 
     @property
-    def use_naive_all2all_kernels(self):
-        return self.moe_parallel_config.use_naive_all2all_kernels
+    def use_ag_rs_all2all_kernels(self):
+        return self.moe_parallel_config.use_ag_rs_all2all_kernels
 
     @property
     def use_nixl_ep_kernels(self):

vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py

@@ -79,7 +79,7 @@ class TrtLlmFp8ExpertsBase:
         """Monolithic kernel so only use with naive DP/EP and TP."""
         return (
             not moe_parallel_config.use_all2all_kernels
-            or moe_parallel_config.use_naive_all2all_kernels
+            or moe_parallel_config.use_ag_rs_all2all_kernels
         ) and not moe_parallel_config.enable_eplb
 
     def supports_chunking(self) -> bool: