Refactoring Mooncake TE as a shared distributed component (#17810)

Signed-off-by: Shangming Cai <csmthu@gmail.com>

python/sglang/srt/model_executor/model_runner.py

@@ -373,6 +373,9 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         if self.device == "cpu":
             self.init_threads_binding()
 
+        # Initialize MooncakeTransferEngine
+        self.init_shared_mooncake_transfer_engine()
+
         # Get memory before model loading
         min_per_gpu_memory = self.init_torch_distributed()
 
@@ -828,6 +831,46 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         )
         return min_per_gpu_memory
 
+    def init_shared_mooncake_transfer_engine(self):
+        """
+        Need MooncakeTransferEngine when:
+        1) PD disaggregation uses mooncake for KV transfer (prefill/decode)
+        2) HiCache uses mooncake storage backend
+        3) Encoder disaggregation uses mooncake
+        """
+        use_mooncake_te = (
+            (
+                self.server_args.disaggregation_mode != "null"
+                and self.server_args.disaggregation_transfer_backend == "mooncake"
+            )
+            or (
+                self.server_args.enable_hierarchical_cache
+                and self.server_args.hicache_storage_backend == "mooncake"
+            )
+            or (
+                self.server_args.encoder_only
+                and self.server_args.encoder_transfer_backend == "mooncake"
+            )
+            or (
+                self.server_args.language_only
+                and self.server_args.encoder_transfer_backend == "mooncake"
+            )
+        )
+
+        if use_mooncake_te:
+            from sglang.srt.distributed.device_communicators.mooncake_transfer_engine import (
+                init_mooncake_transfer_engine,
+            )
+
+            init_mooncake_transfer_engine(
+                hostname=get_local_ip_auto(),
+                gpu_id=self.gpu_id,
+                ib_device=(
+                    self.server_args.disaggregation_ib_device
+                    or self.server_args.mooncake_ib_device
+                ),
+            )
+
     def load_model(self):
         tic_total = time.perf_counter()
         before_avail_memory = get_available_gpu_memory(self.device, self.gpu_id)