refactor: extract device-to-backend mapping into get_default_distributed_backend (#19202)

Signed-off-by: Wang, Yi <yi.a.wang@intel.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Liangsheng Yin <lsyincs@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>

python/sglang/srt/disaggregation/encode_server.py

@@ -25,6 +25,7 @@ from sglang.srt.configs.load_config import LoadConfig
 from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.disaggregation.encode_receiver import EmbeddingData
 from sglang.srt.distributed.parallel_state import (
+    get_default_distributed_backend,
     get_mooncake_transfer_engine,
     get_tp_group,
     init_distributed_environment,
@@ -176,6 +177,7 @@ class MMEncoder:
         self.use_image_processor_gpu = use_image_processor_gpu
 
         init_distributed_environment(
+            backend=get_default_distributed_backend(self.device),
             world_size=server_args.tp_size,
             rank=rank,
             distributed_init_method=dist_init_method,

python/sglang/srt/distributed/parallel_state.py

@@ -1578,6 +1578,20 @@ def set_torch_symm_mem_all_reduce(enable: bool):
     _ENABLE_TORCH_SYMM_MEM_ALL_REDUCE = enable
 
 
+_DEVICE_TO_DISTRIBUTED_BACKEND = {
+    "cuda": "nccl",
+    "xpu": "xccl",
+    "hpu": "hccl",
+    "cpu": "gloo",
+    "npu": "hccl",
+    "musa": "mccl",
+}
+
+
+def get_default_distributed_backend(device: str) -> str:
+    return _DEVICE_TO_DISTRIBUTED_BACKEND.get(device, "gloo")
+
+
 def init_distributed_environment(
     world_size: int = -1,
     rank: int = -1,

python/sglang/srt/model_executor/model_runner.py

@@ -55,6 +55,7 @@ from sglang.srt.debug_utils.tensor_dump_forward_hook import (
     register_forward_hook_for_model,
 )
 from sglang.srt.distributed import (
+    get_default_distributed_backend,
     get_pp_group,
     get_tp_group,
     get_world_group,
@@ -734,29 +735,17 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             )
             raise
 
-        if self.device == "cuda":
-            if self.server_args.elastic_ep_backend == "mooncake":
-                backend = "mooncake"
-                if self.server_args.mooncake_ib_device:
-                    mooncake_ib_device = self.server_args.mooncake_ib_device.split(",")
-                    try:
-                        from mooncake import ep as mooncake_ep
+        backend = get_default_distributed_backend(self.device)
+        if self.device == "cuda" and self.server_args.elastic_ep_backend == "mooncake":
+            backend = "mooncake"
+            if self.server_args.mooncake_ib_device:
+                mooncake_ib_device = self.server_args.mooncake_ib_device.split(",")
+                try:
+                    from mooncake import ep as mooncake_ep
 
-                        mooncake_ep.set_device_filter(mooncake_ib_device)
-                    except:
-                        pass  # A warning will be raised in `init_distributed_environment`
-            else:
-                backend = "nccl"
-        elif self.device == "xpu":
-            backend = "xccl"
-        elif self.device == "hpu":
-            backend = "hccl"
-        elif self.device == "cpu":
-            backend = "gloo"
-        elif self.device == "npu":
-            backend = "hccl"
-        elif self.device == "musa":
-            backend = "mccl"
+                    mooncake_ep.set_device_filter(mooncake_ib_device)
+                except:
+                    pass  # A warning will be raised in `init_distributed_environment`
 
         before_avail_memory = get_available_gpu_memory(self.device, self.gpu_id)
         if not self.server_args.enable_p2p_check: