[XPU] Fix PP accuracy on XPU device (#47253)

Signed-off-by: yisheng <yi.sheng@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/v1/worker/gpu/pp_utils.py

@@ -9,6 +9,7 @@ import numpy as np
 import torch
 
 from vllm.distributed.parallel_state import get_pp_group
+from vllm.platforms import current_platform
 from vllm.v1.worker.gpu.buffer_utils import async_copy_to_gpu
 from vllm.v1.worker.gpu.input_batch import InputBatch
 
@@ -179,6 +180,10 @@ class PPHandler:
             return
 
         assert sampled_token_ids.dtype == torch.int64
+
+        if current_platform.is_xpu():
+            self.main_stream.synchronize()
+
         with torch.cuda.stream(self.broadcast_stream):
             self.broadcast_stream.wait_stream(self.main_stream)
             torch.distributed.broadcast(