|
@@ -36,11 +36,11 @@ func New(model string, adapters []string, opts api.Options) (LLM, error) {
|
|
}
|
|
}
|
|
|
|
|
|
switch ggml.FileType().String() {
|
|
switch ggml.FileType().String() {
|
|
- case "F32", "F16", "Q5_0", "Q5_1", "Q8_0":
|
|
|
|
|
|
+ case "F32", "Q5_0", "Q5_1", "Q8_0":
|
|
if opts.NumGPU != 0 {
|
|
if opts.NumGPU != 0 {
|
|
// F32, F16, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
|
|
// F32, F16, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
|
|
// cause the runner to segmentation fault so disable GPU
|
|
// cause the runner to segmentation fault so disable GPU
|
|
- log.Printf("WARNING: GPU disabled for F32, F16, Q5_0, Q5_1, and Q8_0")
|
|
|
|
|
|
+ log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
|
|
opts.NumGPU = 0
|
|
opts.NumGPU = 0
|
|
}
|
|
}
|
|
}
|
|
}
|
|
@@ -48,19 +48,27 @@ func New(model string, adapters []string, opts api.Options) (LLM, error) {
|
|
totalResidentMemory := memory.TotalMemory()
|
|
totalResidentMemory := memory.TotalMemory()
|
|
switch ggml.ModelType() {
|
|
switch ggml.ModelType() {
|
|
case ModelType3B, ModelType7B:
|
|
case ModelType3B, ModelType7B:
|
|
- if totalResidentMemory < 8*1024*1024 {
|
|
|
|
|
|
+ if ggml.FileType().String() == "F16" && totalResidentMemory < 16*1024*1024 {
|
|
|
|
+ return nil, fmt.Errorf("F16 model requires at least 16GB of memory")
|
|
|
|
+ } else if totalResidentMemory < 8*1024*1024 {
|
|
return nil, fmt.Errorf("model requires at least 8GB of memory")
|
|
return nil, fmt.Errorf("model requires at least 8GB of memory")
|
|
}
|
|
}
|
|
case ModelType13B:
|
|
case ModelType13B:
|
|
- if totalResidentMemory < 16*1024*1024 {
|
|
|
|
|
|
+ if ggml.FileType().String() == "F16" && totalResidentMemory < 32*1024*1024 {
|
|
|
|
+ return nil, fmt.Errorf("F16 model requires at least 32GB of memory")
|
|
|
|
+ } else if totalResidentMemory < 16*1024*1024 {
|
|
return nil, fmt.Errorf("model requires at least 16GB of memory")
|
|
return nil, fmt.Errorf("model requires at least 16GB of memory")
|
|
}
|
|
}
|
|
case ModelType30B, ModelType34B:
|
|
case ModelType30B, ModelType34B:
|
|
- if totalResidentMemory < 32*1024*1024 {
|
|
|
|
|
|
+ if ggml.FileType().String() == "F16" && totalResidentMemory < 64*1024*1024 {
|
|
|
|
+ return nil, fmt.Errorf("F16 model requires at least 64GB of memory")
|
|
|
|
+ } else if totalResidentMemory < 32*1024*1024 {
|
|
return nil, fmt.Errorf("model requires at least 32GB of memory")
|
|
return nil, fmt.Errorf("model requires at least 32GB of memory")
|
|
}
|
|
}
|
|
case ModelType65B:
|
|
case ModelType65B:
|
|
- if totalResidentMemory < 64*1024*1024 {
|
|
|
|
|
|
+ if ggml.FileType().String() == "F16" && totalResidentMemory < 128*1024*1024 {
|
|
|
|
+ return nil, fmt.Errorf("F16 model requires at least 128GB of memory")
|
|
|
|
+ } else if totalResidentMemory < 64*1024*1024 {
|
|
return nil, fmt.Errorf("model requires at least 64GB of memory")
|
|
return nil, fmt.Errorf("model requires at least 64GB of memory")
|
|
}
|
|
}
|
|
}
|
|
}
|