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 5, 09:45 UTC
[XPU] check is_xccl_available before oneccl warmup (#39302)
Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>
Xinyu Chen committed
Apr 10, 2026 at 03:42 UTC
a8134aef4e0cac7ffb72b24d89378df09d88f9cc
1 file changed
+2
-1
vllm/v1/worker/xpu_worker.py
+2
-1
@@ -86,7 +86,8 @@ class XPUWorker(Worker):
86
)
87
88
# global all_reduce needed for overall oneccl warm up
89
- torch.distributed.all_reduce(torch.zeros(1).xpu())
89
+ if torch.distributed.is_xccl_available():
90
+ torch.distributed.all_reduce(torch.zeros(1).xpu())
91
92
# Set random seed.
93
set_random_seed(self.model_config.seed)