From 14d03b8ddb383a114e9085f7ece48629e11991b5 Mon Sep 17 00:00:00 2001 From: "7. Sun" Date: Sat, 24 Jan 2026 04:56:23 +0000 Subject: [PATCH] [Perf] Cache xpu_get_mem_info() result to avoid duplicate calls (#32983) Signed-off-by: 7. Sun --- vllm/v1/worker/xpu_worker.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/vllm/v1/worker/xpu_worker.py b/vllm/v1/worker/xpu_worker.py index fe0850771..611df090f 100644 --- a/vllm/v1/worker/xpu_worker.py +++ b/vllm/v1/worker/xpu_worker.py @@ -107,7 +107,8 @@ class XPUWorker(Worker): torch.xpu.empty_cache() torch_allocated_bytes = torch.xpu.memory_stats()["allocated_bytes.all.current"] - total_allocated_bytes = self.xpu_get_mem_info()[1] - self.xpu_get_mem_info()[0] + free_mem, total_mem = self.xpu_get_mem_info() + total_allocated_bytes = total_mem - free_mem non_torch_allocations = total_allocated_bytes - torch_allocated_bytes if non_torch_allocations > 0: