A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[XPU] Fix xpu model runner call torch.cuda APIs (#25011)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Sep 17, 2025 at 14:45 UTC
dd39baf7175c5e79faef071c67bb318eadb7752f
1 file changed
+5
-1
vllm/v1/worker/xpu_model_runner.py
+5
-1
@@ -45,8 +45,12 @@ def _torch_cuda_wrapper():
45
self.synchronize = lambda: None
46
47
try:
48
- # replace cuda Event with xpu Event, this should work by default
48
+ # replace cuda APIs with xpu APIs, this should work by default
49
torch.cuda.Event = torch.xpu.Event
50
+ torch.cuda.Stream = torch.xpu.Stream
51
+ torch.cuda.default_stream = torch.xpu.current_stream
52
+ torch.cuda.current_stream = torch.xpu.current_stream
53
+ torch.cuda.stream = torch.xpu.stream
54
yield
55
finally:
56
# if anything goes wrong, just patch it with a placeholder