vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

adjusted layer estimation

Browse source
This commit is contained in:
Concedo 2024-07-24 21:54:49 +08:00
parent b7fc8e644a
commit e28c42d7f7
6 changed files with 9 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

2
koboldcpp.py
View file

@ -612,7 +612,7 @@ def autoset_gpu_layers(filepath,ctxsize,gpumem): #shitty algo to determine how m
headcount = ggufmeta[1]
headkvlen = (ggufmeta[2] if ggufmeta[2] > 0 else 128)
ratio = mem/(fsize*csmul*1.5)
computemem = layers*4*headkvlen*cs*4*1.25 # For now the first 4 is the hardcoded result for a blasbatchsize of 512. Ideally we automatically calculate blasbatchsize / 4 but I couldn't easily grab the value yet - Henk
computemem = layers*4*headkvlen*cs*4*1.35 # For now the first 4 is the hardcoded result for a blasbatchsize of 512. Ideally we automatically calculate blasbatchsize / 4 but I couldn't easily grab the value yet - Henk
contextmem = layers*headcount*headkvlen*cs*4
reservedmem = 1.5*1024*1024*1024 # Users often don't have their GPU's VRAM worth of memory, we assume 500MB to avoid driver swapping + 500MB for the OS + 500MB for background apps / browser - Henk
if headcount > 0: