[Hardware] Replace torch.cuda.synchronize() api with torch.accelerator.synchronize (#36085)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>

examples/offline_inference/rlhf_colocate.py

@@ -88,7 +88,7 @@ class RayTrainingActor:
         # Zero out all the parameters.
         for name, p in self.model.named_parameters():
             p.data.zero_()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         # The argument for `get_device_uuid` is the index of the GPU in the
         # list of visible devices.
         from vllm.platforms import current_platform
@@ -151,7 +151,7 @@ class RayTrainingActor:
                     p.data.view(-1).view(dtype=torch.uint8), non_blocking=True
                 )
                 offset += get_size(p)
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             s.send_pyobj(named_tensors)
             s.recv()
         s.send_pyobj(None)

examples/offline_inference/rlhf_utils.py

@@ -120,7 +120,7 @@ class ColocateWorkerExtension:
                 process_weights_after_loading(
                     self.model_runner.model, self.model_config, self.device
                 )
-                torch.cuda.synchronize()
+                torch.accelerator.synchronize()
                 socket.send(b"")
                 break
             if isinstance(payload, tuple):
@@ -144,7 +144,7 @@ class ColocateWorkerExtension:
                 weights.append((item["name"], tensor))
             self.model_runner.model.load_weights(weights=weights)
             del weights
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             socket.send(b"")
 
         socket.close()