From e7cba8f6b109c3baee107a222c79365e7f3ea70f Mon Sep 17 00:00:00 2001
From: Naman Lalit <nl2688@nyu.edu>
Date: Fri, 26 Sep 2025 17:07:36 -0700
Subject: [PATCH] [Bugfix] Optimize CpuGpuBuffer initialization (#25447)

Signed-off-by: Naman Lalit <nl2688@nyu.edu>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
---
 vllm/v1/utils.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/vllm/v1/utils.py b/vllm/v1/utils.py
index ec4417290f611..ee0c1168f3cd0 100644
--- a/vllm/v1/utils.py
+++ b/vllm/v1/utils.py
@@ -117,7 +117,7 @@ class CpuGpuBuffer:
                                dtype=dtype,
                                device="cpu",
                                pin_memory=pin_memory)
-        self.gpu = self.cpu.to(device)
+        self.gpu = torch.zeros_like(self.cpu, device=device)
         self.np: np.ndarray
         # To keep type hints simple (avoiding generics and subclasses), we
         # only conditionally create the numpy array attribute. This can cause