From aacfa8a231f36014c329ff799c02033460edbd96 Mon Sep 17 00:00:00 2001
From: "Li, Zonghang" <870644199@qq.com>
Date: Thu, 26 Jun 2025 20:45:45 +0400
Subject: [PATCH] fix compute buffer estimate: reserve 300 MiB VRAM to avoid
 potential OOM

---
 src/llama.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index b70f338b..abda2327 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -22043,8 +22043,8 @@ void llama_model_compute_buf_size(
         // context GPU memory usage, i.e. the initial memory cost of creating a CUDA context, 
         // even before you launch any kernels or allocate your own buffers.
         // this value may vary by GPU and CUDA version, but it's lower than 400 MiB in most cases,
-        // another 200 MiB is used to prevent accidental OOM.
-        *gpu_buf += 600 * 1024 * 1024;
+        // another 300 MiB is used to prevent accidental OOM.
+        *gpu_buf += 700 * 1024 * 1024;
     }
 }