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

use runner if cuda alloc won't fit

Browse source
This commit is contained in:
Jeffrey Morgan 2024-01-09 00:44:34 -05:00
parent 6566387ae3
commit f387e9631b
1 changed files with 3 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
llm/llm.go
View file

@ -100,10 +100,8 @@ func New(workDir, model string, adapters, projectors []string, opts api.Options)
break break
} }
// This handles two cases: // requiredAlloc is always loaded for the CUDA runner, so don't load it if it won't fit
// 1. overhead + tensors are always loaded into scratch memory even with num_gpu 0 if requiredAlloc > available {
// 2. it seems llama.cpp always tries to allocate the entire kv cache (even if later split into layers) into vram or crashes
if requiredAlloc > available || requiredKv > available {
log.Printf("not enough vram available, falling back to CPU only") log.Printf("not enough vram available, falling back to CPU only")
library = "cpu" library = "cpu"
opts.NumGPU = 0 opts.NumGPU = 0
@ -127,8 +125,7 @@ func New(workDir, model string, adapters, projectors []string, opts api.Options)
opts.NumGQA = 0 opts.NumGQA = 0
opts.RopeFrequencyBase = 0.0 opts.RopeFrequencyBase = 0.0
opts.RopeFrequencyScale = 0.0 opts.RopeFrequencyScale = 0.0
gpuInfo := gpu.GetGPUInfo() return newLlmServer(library, model, adapters, projectors, opts)
return newLlmServer(gpuInfo.Library, model, adapters, projectors, opts)
} }
// Give any native cgo implementations an opportunity to initialize // Give any native cgo implementations an opportunity to initialize