From 17cd8ba618ae51157b4da1082950cb1a0d085f51 Mon Sep 17 00:00:00 2001 From: Lizonghang <870644199@qq.com> Date: Fri, 31 Jan 2025 16:24:44 +0400 Subject: [PATCH] reverse 300MiB for Metal kernel --- src/llama.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/llama.cpp b/src/llama.cpp index 741d3063..2531f0fe 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3622,8 +3622,8 @@ void llama_profile_device( dev_info->gpu_props.name = gpu_props.name; dev_info->gpu_props.description = gpu_props.description; - // reserved/limit memory to avoid potential OOM, default to 200 MiB - dev_info->gpu_props.memory_free = round(gpu_props.memory_free / (double)(1 << 30) * 100) / 100 - 0.2; + // reserved/limit memory to avoid potential OOM, default to 300 MiB + dev_info->gpu_props.memory_free = round(gpu_props.memory_free / (double)(1 << 30) * 100) / 100 - 0.3; dev_info->gpu_props.memory_free = std::min((float)gpu_mem, dev_info->gpu_props.memory_free); dev_info->gpu_props.memory_total = round(gpu_props.memory_total / (double)(1 << 30) * 100) / 100;