[Feature, Hardware] Enable SGLang on XPU GPUs via PyTorch (#1480)

python/sglang/srt/model_executor/forward_batch_info.py

@@ -118,7 +118,7 @@ class ForwardBatch:
         batch: ModelWorkerBatch,
         model_runner: ModelRunner,
     ):
-        device = "cuda"
+        device = model_runner.device
 
         ret = cls(
             forward_mode=batch.forward_mode,

python/sglang/srt/model_executor/model_runner.py

@@ -138,6 +138,7 @@ class ModelRunner:
             self.init_attention_backend()
             self.init_cuda_graphs()
         else:
+            self.cuda_graph_runner = None
             self.init_attention_backend()
 
     def init_torch_distributed(self):
@@ -146,6 +147,11 @@ class ModelRunner:
         if self.device == "cuda":
             torch.cuda.set_device(self.gpu_id)
             backend = "nccl"
+        # ToDO(liangan1):Just use gloo to bypass the initilization fail
+        # Need to use xccl for xpu backend in the future
+        elif self.device == "xpu":
+            torch.xpu.set_device(self.gpu_id)
+            backend = "gloo"
 
         if not self.server_args.enable_p2p_check:
             monkey_patch_vllm_p2p_access_check(self.gpu_id)