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

python/sglang/srt/model_executor/forward_batch_info.py

@@ -38,6 +38,7 @@ import torch
 import triton
 import triton.language as tl
 
+from sglang.srt.distributed.parallel_state import get_moe_expert_parallel_world_size
 from sglang.srt.layers.dp_attention import (
     DPPaddingMode,
     get_attention_dp_rank,
@@ -839,7 +840,7 @@ class ForwardBatch:
 
 
 def enable_num_token_non_padded(server_args):
-    return server_args.enable_ep_moe or server_args.enable_deepep_moe
+    return get_moe_expert_parallel_world_size() > 1
 
 
 class PPProxyTensors:

python/sglang/srt/model_executor/model_runner.py

@@ -60,6 +60,7 @@ from sglang.srt.layers.dp_attention import (
     initialize_dp_attention,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.moe.utils import DeepEPMode, MoeA2ABackend
 from sglang.srt.layers.quantization import (
     deep_gemm_wrapper,
     monkey_patch_isinstance_for_vllm_base_layer,
@@ -217,6 +218,10 @@ class ModelRunner:
                 "use_mla_backend": self.use_mla_backend,
                 "speculative_algorithm": self.spec_algorithm,
             }
+            | {
+                "moe_a2a_backend": MoeA2ABackend(server_args.moe_a2a_backend),
+                "deepep_mode": DeepEPMode(server_args.deepep_mode),
+            }
         )
 
         # CPU offload