From 69f3150981e4bc8a09439eb7cae0095605b964b4 Mon Sep 17 00:00:00 2001 From: YiSheng5 Date: Tue, 7 Jul 2026 09:17:27 +0800 Subject: [PATCH] [XPU] Fix PP accuracy on XPU device (#47253) Signed-off-by: yisheng Co-authored-by: Kunshang Ji --- vllm/v1/worker/gpu/pp_utils.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm/v1/worker/gpu/pp_utils.py b/vllm/v1/worker/gpu/pp_utils.py index 9f5d4c2d80..9b1786cbe4 100644 --- a/vllm/v1/worker/gpu/pp_utils.py +++ b/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(