Enable CPU device on SGLang (#2806)

python/sglang/srt/managers/schedule_batch.py

@@ -65,6 +65,7 @@ global_server_args_dict = {
     "enable_nan_detection": ServerArgs.enable_nan_detection,
     "enable_dp_attention": ServerArgs.enable_dp_attention,
     "enable_ep_moe": ServerArgs.enable_ep_moe,
+    "device": ServerArgs.device,
 }
 
 

python/sglang/srt/managers/scheduler.py

@@ -317,6 +317,8 @@ class Scheduler:
         self.last_decode_stats_tic = time.time()
         self.stream_interval = server_args.stream_interval
         self.current_stream = torch.get_device_module(self.device).current_stream()
+        if self.device == "cpu":
+            self.current_stream.synchronize = lambda: None  # No-op for CPU
 
         # Session info
         self.sessions: Dict[str, Session] = {}

python/sglang/srt/managers/tp_worker_overlap_thread.py

@@ -82,6 +82,8 @@ class TpModelWorkerClient:
         self.forward_thread.start()
         self.parent_process = psutil.Process().parent()
         self.scheduler_stream = torch.get_device_module(self.device).current_stream()
+        if self.device == "cpu":
+            self.scheduler_stream.synchronize = lambda: None  # No-op for CPU
 
     def get_worker_info(self):
         return self.worker.get_worker_info()