From 6164f378f2780f8563f08e8cba315450f6b3b94e Mon Sep 17 00:00:00 2001 From: Jeffrey Morgan Date: Tue, 9 Jan 2024 00:54:25 -0500 Subject: [PATCH] revert cuda overhead to 20% --- gpu/gpu.go | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/gpu/gpu.go b/gpu/gpu.go index 07a3bee3..5bc1810e 100644 --- a/gpu/gpu.go +++ b/gpu/gpu.go @@ -131,9 +131,8 @@ func getCPUMem() (memInfo, error) { func CheckVRAM() (int64, error) { gpuInfo := GetGPUInfo() if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") { - // leave 25% of VRAM free for overhead - // TODO: improve cuda memory allocation to avoid needing this large of an amount - return int64(gpuInfo.FreeMemory * 3 / 4), nil + // leave 20% of VRAM free for unaccounted for overhead + return int64(gpuInfo.FreeMemory * 4 / 5), nil } return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation