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

default rope params to 0 for new models (#968)

Browse source
This commit is contained in:
Jeffrey Morgan 2023-11-02 08:41:30 -07:00 committed by GitHub
parent 527f9a7975
commit 2e53704685
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 12 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
llm/llama.go
View file

@ -306,13 +306,19 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
params := []string{ params := []string{
"--model", model, "--model", model,
"--ctx-size", fmt.Sprintf("%d", opts.NumCtx), "--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
"--batch-size", fmt.Sprintf("%d", opts.NumBatch), "--batch-size", fmt.Sprintf("%d", opts.NumBatch),
"--n-gpu-layers", fmt.Sprintf("%d", numGPU), "--n-gpu-layers", fmt.Sprintf("%d", numGPU),
"--embedding", "--embedding",
} }
if opts.RopeFrequencyBase > 0 {
params = append(params, "--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase))
}
if opts.RopeFrequencyScale > 0 {
params = append(params, "--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale))
}
if opts.NumGQA > 0 { if opts.NumGQA > 0 {
params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA)) params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA))
} }

5
llm/llm.go
View file

@ -85,7 +85,10 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
switch ggml.Name() { switch ggml.Name() {
case "gguf": case "gguf":
opts.NumGQA = 0 // TODO: remove this when llama.cpp runners differ enough to need separate newLlama functions // TODO: gguf will load these options automatically from the model binary
opts.NumGQA = 0
opts.RopeFrequencyBase = 0.0
opts.RopeFrequencyScale = 0.0
return newLlama(model, adapters, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts) return newLlama(model, adapters, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
case "ggml", "ggmf", "ggjt", "ggla": case "ggml", "ggmf", "ggjt", "ggla":
return newLlama(model, adapters, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts) return newLlama(model, adapters, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)