baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

update rough cuda overhead estimate to 15% + 384MiB

Browse source
This commit is contained in:
Jeffrey Morgan 2024-01-09 11:47:30 -05:00
parent 62023177f6
commit 6df83e6daa
1 changed files with 6 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
gpu/gpu.go
View file

@ -131,8 +131,12 @@ func getCPUMem() (memInfo, error) {
func CheckVRAM() (int64, error) {
gpuInfo := GetGPUInfo()
if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") {
// leave 20% of VRAM free for unaccounted for overhead
return int64(gpuInfo.FreeMemory * 4 / 5), nil
// leave 10% or 384Mi of VRAM free for unaccounted for overhead
overhead := gpuInfo.FreeMemory / 10
if overhead < 384*1024*1024 {
overhead = 384 * 1024 * 1024
}
return int64(gpuInfo.FreeMemory - overhead), nil
}
return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation