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 9, 17:06 UTC
[Perf] Cache xpu_get_mem_info() result to avoid duplicate calls (#32983)
Signed-off-by: 7. Sun <jhao.sun@gmail.com>
7. Sun committed
Jan 24, 2026 at 04:56 UTC
14d03b8ddb383a114e9085f7ece48629e11991b5
1 file changed
+2
-1
vllm/v1/worker/xpu_worker.py
+2
-1
@@ -107,7 +107,8 @@ class XPUWorker(Worker):
107
108
torch.xpu.empty_cache()
109
torch_allocated_bytes = torch.xpu.memory_stats()["allocated_bytes.all.current"]
110
- total_allocated_bytes = self.xpu_get_mem_info()[1] - self.xpu_get_mem_info()[0]
110
+ free_mem, total_mem = self.xpu_get_mem_info()
111
+ total_allocated_bytes = total_mem - free_mem
112
113
non_torch_allocations = total_allocated_bytes - torch_allocated_bytes
114
if non_torch_allocations > 0: