[5/N] MoE Refactor: Update MoE parallelism arguments (#8658)

python/sglang/srt/managers/scheduler.py

@@ -64,6 +64,7 @@ from sglang.srt.hf_transformers_utils import (
 )
 from sglang.srt.layers.dp_attention import compute_dp_attention_world_info
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.moe.utils import DeepEPMode, MoeA2ABackend
 from sglang.srt.managers.io_struct import (
     AbortReq,
     CloseSessionReqInput,
@@ -137,7 +138,6 @@ from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
 from sglang.srt.two_batch_overlap import TboDPAttentionPreparer
 from sglang.srt.utils import (
-    DeepEPMode,
     DynamicGradMode,
     broadcast_pyobj,
     configure_gc_logger,
@@ -1762,8 +1762,10 @@ class Scheduler(
             spec_algorithm=self.spec_algorithm,
             speculative_num_draft_tokens=self.server_args.speculative_num_draft_tokens,
             enable_two_batch_overlap=self.server_args.enable_two_batch_overlap,
-            enable_deepep_moe=self.server_args.enable_deepep_moe,
-            deepep_mode=DeepEPMode[self.server_args.deepep_mode],
+            enable_deepep_moe=MoeA2ABackend(
+                self.server_args.moe_a2a_backend
+            ).is_deepep(),
+            deepep_mode=DeepEPMode(self.server_args.deepep_mode),
             require_mlp_tp_gather=require_mlp_tp_gather(self.server_args),
             disable_overlap_schedule=self.server_args.disable_overlap_schedule,
         )