[Feature] Integrate DeepEP into SGLang (#4232)

Co-authored-by: Cheng Wan <cwan39@gatech.edu>
Co-authored-by: Xuting Zhou <xutingz@nvidia.com>

python/sglang/srt/server_args.py

@@ -157,6 +157,7 @@ class ServerArgs:
     enable_mixed_chunk: bool = False
     enable_dp_attention: bool = False
     enable_ep_moe: bool = False
+    enable_deepep_moe: bool = False
     enable_torch_compile: bool = False
     torch_compile_max_bs: int = 32
     cuda_graph_max_bs: Optional[int] = None
@@ -281,6 +282,12 @@ class ServerArgs:
             logger.warning(
                 f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
             )
+            # DeepEP MoE
+            if self.enable_deepep_moe:
+                self.ep_size = self.dp_size
+                logger.info(
+                    f"DeepEP MoE is enabled. The expert parallel size is adjusted to be the same as the data parallel size[{self.dp_size}]."
+                )
 
         # Speculative Decoding
         if self.speculative_algorithm == "NEXTN":
@@ -1018,6 +1025,11 @@ class ServerArgs:
             default=ServerArgs.hicache_ratio,
             help="The ratio of the size of host KV cache memory pool to the size of device pool.",
         )
+        parser.add_argument(
+            "--enable-deepep-moe",
+            action="store_true",
+            help="Enabling DeepEP MoE implementation for EP MoE.",
+        )
 
         # Server warmups
         parser.add_argument(