dearwolf/LocalAI
1
0
Fork 0
mirror of https://github.com/mudler/LocalAI.git synced 2025-05-31 16:05:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 26

feat: add initial AutoGPTQ backend implementation

Browse source
This commit is contained in:
Ettore Di Giacinto 2023-08-07 22:39:10 +02:00
parent 91d49cfe9f
commit a843e64fc2
37 changed files with 660 additions and 148 deletions
Download patch file Download diff file Expand all files Collapse all files

2
pkg/grpc/image/stablediffusion.go
View file

@ -16,7 +16,7 @@ type StableDiffusion struct {
func (sd *StableDiffusion) Load(opts *pb.ModelOptions) error {
var err error
// Note: the Model here is a path to a directory containing the model files
sd.stablediffusion, err = stablediffusion.New(opts.Model)
sd.stablediffusion, err = stablediffusion.New(opts.ModelFile)
return err
}

2
pkg/grpc/llm/bert/bert.go
View file

@ -15,7 +15,7 @@ type Embeddings struct {
}
func (llm *Embeddings) Load(opts *pb.ModelOptions) error {
model, err := bert.New(opts.Model)
model, err := bert.New(opts.ModelFile)
llm.bert = model
return err
}

2
pkg/grpc/llm/bloomz/bloomz.go
View file

@ -18,7 +18,7 @@ type LLM struct {
}
func (llm *LLM) Load(opts *pb.ModelOptions) error {
model, err := bloomz.New(opts.Model)
model, err := bloomz.New(opts.ModelFile)
llm.bloomz = model
return err
}

2
pkg/grpc/llm/falcon/falcon.go
View file

@ -40,7 +40,7 @@ func (llm *LLM) Load(opts *pb.ModelOptions) error {
ggllmOpts = append(ggllmOpts, ggllm.SetNBatch(512))
}
model, err := ggllm.New(opts.Model, ggllmOpts...)
model, err := ggllm.New(opts.ModelFile, ggllmOpts...)
llm.falcon = model
return err
}

2
pkg/grpc/llm/gpt4all/gpt4all.go
View file

@ -17,7 +17,7 @@ type LLM struct {
}
func (llm *LLM) Load(opts *pb.ModelOptions) error {
model, err := gpt4all.New(opts.Model,
model, err := gpt4all.New(opts.ModelFile,
gpt4all.SetThreads(int(opts.Threads)),
gpt4all.SetLibrarySearchPath(opts.LibrarySearchPath))
llm.gpt4all = model

2
pkg/grpc/llm/llama/llama.go
View file

@ -71,7 +71,7 @@ func (llm *LLM) Load(opts *pb.ModelOptions) error {
llamaOpts = append(llamaOpts, llama.EnabelLowVRAM)
}
model, err := llama.New(opts.Model, llamaOpts...)
model, err := llama.New(opts.ModelFile, llamaOpts...)
llm.llama = model
return err
}

6
pkg/grpc/llm/rwkv/rwkv.go
View file

@ -20,9 +20,9 @@ type LLM struct {
}
func (llm *LLM) Load(opts *pb.ModelOptions) error {
modelPath := filepath.Dir(opts.Model)
modelFile := filepath.Base(opts.Model)
model := rwkv.LoadFiles(opts.Model, filepath.Join(modelPath, modelFile+tokenizerSuffix), uint32(opts.GetThreads()))
modelPath := filepath.Dir(opts.ModelFile)
modelFile := filepath.Base(opts.ModelFile)
model := rwkv.LoadFiles(opts.ModelFile, filepath.Join(modelPath, modelFile+tokenizerSuffix), uint32(opts.GetThreads()))
if model == nil {
return fmt.Errorf("could not load model")

2
pkg/grpc/llm/transformers/dolly.go
View file

@ -18,7 +18,7 @@ type Dolly struct {
}
func (llm *Dolly) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewDolly(opts.Model)
model, err := transformers.NewDolly(opts.ModelFile)
llm.dolly = model
return err
}

2
pkg/grpc/llm/transformers/falcon.go
View file

@ -18,7 +18,7 @@ type Falcon struct {
}
func (llm *Falcon) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewFalcon(opts.Model)
model, err := transformers.NewFalcon(opts.ModelFile)
llm.falcon = model
return err
}

2
pkg/grpc/llm/transformers/gpt2.go
View file

@ -18,7 +18,7 @@ type GPT2 struct {
}
func (llm *GPT2) Load(opts *pb.ModelOptions) error {
model, err := transformers.New(opts.Model)
model, err := transformers.New(opts.ModelFile)
llm.gpt2 = model
return err
}

2
pkg/grpc/llm/transformers/gptj.go
View file

@ -18,7 +18,7 @@ type GPTJ struct {
}
func (llm *GPTJ) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewGPTJ(opts.Model)
model, err := transformers.NewGPTJ(opts.ModelFile)
llm.gptj = model
return err
}

2
pkg/grpc/llm/transformers/gptneox.go
View file

@ -18,7 +18,7 @@ type GPTNeoX struct {
}
func (llm *GPTNeoX) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewGPTNeoX(opts.Model)
model, err := transformers.NewGPTNeoX(opts.ModelFile)
llm.gptneox = model
return err
}

2
pkg/grpc/llm/transformers/mpt.go
View file

@ -18,7 +18,7 @@ type MPT struct {
}
func (llm *MPT) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewMPT(opts.Model)
model, err := transformers.NewMPT(opts.ModelFile)
llm.mpt = model
return err
}

2
pkg/grpc/llm/transformers/replit.go
View file

@ -18,7 +18,7 @@ type Replit struct {
}
func (llm *Replit) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewReplit(opts.Model)
model, err := transformers.NewReplit(opts.ModelFile)
llm.replit = model
return err
}

2
pkg/grpc/llm/transformers/starcoder.go
View file

@ -18,7 +18,7 @@ type Starcoder struct {
}
func (llm *Starcoder) Load(opts *pb.ModelOptions) error {
model, err := transformers.NewStarcoder(opts.Model)
model, err := transformers.NewStarcoder(opts.ModelFile)
llm.starcoder = model
return err
}

209
pkg/grpc/proto/backend.pb.go
View file

@ -483,6 +483,11 @@ type ModelOptions struct {
RopeFreqScale float32 `protobuf:"fixed32,18,opt,name=RopeFreqScale,proto3" json:"RopeFreqScale,omitempty"`
RMSNormEps float32 `protobuf:"fixed32,19,opt,name=RMSNormEps,proto3" json:"RMSNormEps,omitempty"`
NGQA int32 `protobuf:"varint,20,opt,name=NGQA,proto3" json:"NGQA,omitempty"`
ModelFile string `protobuf:"bytes,21,opt,name=ModelFile,proto3" json:"ModelFile,omitempty"`
// AutoGPTQ
Device string `protobuf:"bytes,22,opt,name=Device,proto3" json:"Device,omitempty"`
UseTriton bool `protobuf:"varint,23,opt,name=UseTriton,proto3" json:"UseTriton,omitempty"`
ModelBaseName string `protobuf:"bytes,24,opt,name=ModelBaseName,proto3" json:"ModelBaseName,omitempty"`
}
func (x *ModelOptions) Reset() {
@ -657,6 +662,34 @@ func (x *ModelOptions) GetNGQA() int32 {
return 0
}
func (x *ModelOptions) GetModelFile() string {
if x != nil {
return x.ModelFile
}
return ""
}
func (x *ModelOptions) GetDevice() string {
if x != nil {
return x.Device
}
return ""
}
func (x *ModelOptions) GetUseTriton() bool {
if x != nil {
return x.UseTriton
}
return false
}
func (x *ModelOptions) GetModelBaseName() string {
if x != nil {
return x.ModelBaseName
}
return ""
}
type Result struct {
state protoimpl.MessageState
sizeCache protoimpl.SizeCache
@ -1207,7 +1240,7 @@ var file_pkg_grpc_proto_backend_proto_rawDesc = []byte{
0x0e, 0x4e, 0x65, 0x67, 0x61, 0x74, 0x69, 0x76, 0x65, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22,
0x21, 0x0a, 0x05, 0x52, 0x65, 0x70, 0x6c, 0x79, 0x12, 0x18, 0x0a, 0x07, 0x6d, 0x65, 0x73, 0x73,
0x61, 0x67, 0x65, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61,
0x67, 0x65, 0x22, 0xc8, 0x04, 0x0a, 0x0c, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x4f, 0x70, 0x74, 0x69,
0x67, 0x65, 0x22, 0xc2, 0x05, 0x0a, 0x0c, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x4f, 0x70, 0x74, 0x69,
0x6f, 0x6e, 0x73, 0x12, 0x14, 0x0a, 0x05, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x18, 0x01, 0x20, 0x01,
0x28, 0x09, 0x52, 0x05, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x20, 0x0a, 0x0b, 0x43, 0x6f, 0x6e,
0x74, 0x65, 0x78, 0x74, 0x53, 0x69, 0x7a, 0x65, 0x18, 0x02, 0x20, 0x01, 0x28, 0x05, 0x52, 0x0b,
@ -1243,90 +1276,98 @@ var file_pkg_grpc_proto_backend_proto_rawDesc = []byte{
0x46, 0x72, 0x65, 0x71, 0x53, 0x63, 0x61, 0x6c, 0x65, 0x12, 0x1e, 0x0a, 0x0a, 0x52, 0x4d, 0x53,
0x4e, 0x6f, 0x72, 0x6d, 0x45, 0x70, 0x73, 0x18, 0x13, 0x20, 0x01, 0x28, 0x02, 0x52, 0x0a, 0x52,
0x4d, 0x53, 0x4e, 0x6f, 0x72, 0x6d, 0x45, 0x70, 0x73, 0x12, 0x12, 0x0a, 0x04, 0x4e, 0x47, 0x51,
0x41, 0x18, 0x14, 0x20, 0x01, 0x28, 0x05, 0x52, 0x04, 0x4e, 0x47, 0x51, 0x41, 0x22, 0x3c, 0x0a,
0x06, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12, 0x18, 0x0a, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61,
0x67, 0x65, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67,
0x65, 0x12, 0x18, 0x0a, 0x07, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x18, 0x02, 0x20, 0x01,
0x28, 0x08, 0x52, 0x07, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x22, 0x31, 0x0a, 0x0f, 0x45,
0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12, 0x1e,
0x0a, 0x0a, 0x65, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x73, 0x18, 0x01, 0x20, 0x03,
0x28, 0x02, 0x52, 0x0a, 0x65, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x5b,
0x0a, 0x11, 0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x52, 0x65, 0x71, 0x75,
0x65, 0x73, 0x74, 0x12, 0x10, 0x0a, 0x03, 0x64, 0x73, 0x74, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09,
0x52, 0x03, 0x64, 0x73, 0x74, 0x12, 0x1a, 0x0a, 0x08, 0x6c, 0x61, 0x6e, 0x67, 0x75, 0x61, 0x67,
0x65, 0x18, 0x03, 0x20, 0x01, 0x28, 0x09, 0x52, 0x08, 0x6c, 0x61, 0x6e, 0x67, 0x75, 0x61, 0x67,
0x65, 0x12, 0x18, 0x0a, 0x07, 0x74, 0x68, 0x72, 0x65, 0x61, 0x64, 0x73, 0x18, 0x04, 0x20, 0x01,
0x28, 0x0d, 0x52, 0x07, 0x74, 0x68, 0x72, 0x65, 0x61, 0x64, 0x73, 0x22, 0x5e, 0x0a, 0x10, 0x54,
0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12,
0x36, 0x0a, 0x08, 0x73, 0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74, 0x73, 0x18, 0x01, 0x20, 0x03, 0x28,
0x0b, 0x32, 0x1a, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x54, 0x72, 0x61, 0x6e,
0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x53, 0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74, 0x52, 0x08, 0x73,
0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74, 0x73, 0x12, 0x12, 0x0a, 0x04, 0x74, 0x65, 0x78, 0x74, 0x18,
0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x04, 0x74, 0x65, 0x78, 0x74, 0x22, 0x77, 0x0a, 0x11, 0x54,
0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x53, 0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74,
0x12, 0x0e, 0x0a, 0x02, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x05, 0x52, 0x02, 0x69, 0x64,
0x12, 0x14, 0x0a, 0x05, 0x73, 0x74, 0x61, 0x72, 0x74, 0x18, 0x02, 0x20, 0x01, 0x28, 0x03, 0x52,
0x05, 0x73, 0x74, 0x61, 0x72, 0x74, 0x12, 0x10, 0x0a, 0x03, 0x65, 0x6e, 0x64, 0x18, 0x03, 0x20,
0x01, 0x28, 0x03, 0x52, 0x03, 0x65, 0x6e, 0x64, 0x12, 0x12, 0x0a, 0x04, 0x74, 0x65, 0x78, 0x74,
0x18, 0x04, 0x20, 0x01, 0x28, 0x09, 0x52, 0x04, 0x74, 0x65, 0x78, 0x74, 0x12, 0x16, 0x0a, 0x06,
0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x18, 0x05, 0x20, 0x03, 0x28, 0x05, 0x52, 0x06, 0x74, 0x6f,
0x6b, 0x65, 0x6e, 0x73, 0x22, 0xe4, 0x01, 0x0a, 0x14, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
0x65, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x12, 0x16, 0x0a,
0x06, 0x68, 0x65, 0x69, 0x67, 0x68, 0x74, 0x18, 0x01, 0x20, 0x01, 0x28, 0x05, 0x52, 0x06, 0x68,
0x65, 0x69, 0x67, 0x68, 0x74, 0x12, 0x14, 0x0a, 0x05, 0x77, 0x69, 0x64, 0x74, 0x68, 0x18, 0x02,
0x20, 0x01, 0x28, 0x05, 0x52, 0x05, 0x77, 0x69, 0x64, 0x74, 0x68, 0x12, 0x12, 0x0a, 0x04, 0x6d,
0x6f, 0x64, 0x65, 0x18, 0x03, 0x20, 0x01, 0x28, 0x05, 0x52, 0x04, 0x6d, 0x6f, 0x64, 0x65, 0x12,
0x12, 0x0a, 0x04, 0x73, 0x74, 0x65, 0x70, 0x18, 0x04, 0x20, 0x01, 0x28, 0x05, 0x52, 0x04, 0x73,
0x74, 0x65, 0x70, 0x12, 0x12, 0x0a, 0x04, 0x73, 0x65, 0x65, 0x64, 0x18, 0x05, 0x20, 0x01, 0x28,
0x05, 0x52, 0x04, 0x73, 0x65, 0x65, 0x64, 0x12, 0x27, 0x0a, 0x0f, 0x70, 0x6f, 0x73, 0x69, 0x74,
0x69, 0x76, 0x65, 0x5f, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x18, 0x06, 0x20, 0x01, 0x28, 0x09,
0x52, 0x0e, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x76, 0x65, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74,
0x12, 0x27, 0x0a, 0x0f, 0x6e, 0x65, 0x67, 0x61, 0x74, 0x69, 0x76, 0x65, 0x5f, 0x70, 0x72, 0x6f,
0x6d, 0x70, 0x74, 0x18, 0x07, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0e, 0x6e, 0x65, 0x67, 0x61, 0x74,
0x69, 0x76, 0x65, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x12, 0x10, 0x0a, 0x03, 0x64, 0x73, 0x74,
0x18, 0x08, 0x20, 0x01, 0x28, 0x09, 0x52, 0x03, 0x64, 0x73, 0x74, 0x22, 0x48, 0x0a, 0x0a, 0x54,
0x54, 0x53, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x12, 0x12, 0x0a, 0x04, 0x74, 0x65, 0x78,
0x74, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x04, 0x74, 0x65, 0x78, 0x74, 0x12, 0x14, 0x0a,
0x05, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x05, 0x6d, 0x6f,
0x64, 0x65, 0x6c, 0x12, 0x10, 0x0a, 0x03, 0x64, 0x73, 0x74, 0x18, 0x03, 0x20, 0x01, 0x28, 0x09,
0x52, 0x03, 0x64, 0x73, 0x74, 0x32, 0xeb, 0x03, 0x0a, 0x07, 0x42, 0x61, 0x63, 0x6b, 0x65, 0x6e,
0x64, 0x12, 0x32, 0x0a, 0x06, 0x48, 0x65, 0x61, 0x6c, 0x74, 0x68, 0x12, 0x16, 0x2e, 0x62, 0x61,
0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x48, 0x65, 0x61, 0x6c, 0x74, 0x68, 0x4d, 0x65, 0x73, 0x73,
0x61, 0x67, 0x65, 0x1a, 0x0e, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65,
0x70, 0x6c, 0x79, 0x22, 0x00, 0x12, 0x34, 0x0a, 0x07, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74,
0x12, 0x17, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69,
0x63, 0x74, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0e, 0x2e, 0x62, 0x61, 0x63, 0x6b,
0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x70, 0x6c, 0x79, 0x22, 0x00, 0x12, 0x35, 0x0a, 0x09, 0x4c,
0x6f, 0x61, 0x64, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x15, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65,
0x6e, 0x64, 0x2e, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a,
0x0f, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74,
0x22, 0x00, 0x12, 0x3c, 0x0a, 0x0d, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x53, 0x74, 0x72,
0x65, 0x61, 0x6d, 0x12, 0x17, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x50, 0x72,
0x65, 0x64, 0x69, 0x63, 0x74, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0e, 0x2e, 0x62,
0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x70, 0x6c, 0x79, 0x22, 0x00, 0x30, 0x01,
0x12, 0x40, 0x0a, 0x09, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x12, 0x17, 0x2e,
0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f,
0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x18, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64,
0x2e, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74,
0x22, 0x00, 0x12, 0x41, 0x0a, 0x0d, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x49, 0x6d,
0x61, 0x67, 0x65, 0x12, 0x1d, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x47, 0x65,
0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x52, 0x65, 0x71, 0x75, 0x65,
0x73, 0x74, 0x1a, 0x0f, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x73,
0x75, 0x6c, 0x74, 0x22, 0x00, 0x12, 0x4d, 0x0a, 0x12, 0x41, 0x75, 0x64, 0x69, 0x6f, 0x54, 0x72,
0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x12, 0x1a, 0x2e, 0x62, 0x61,
0x41, 0x18, 0x14, 0x20, 0x01, 0x28, 0x05, 0x52, 0x04, 0x4e, 0x47, 0x51, 0x41, 0x12, 0x1c, 0x0a,
0x09, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x46, 0x69, 0x6c, 0x65, 0x18, 0x15, 0x20, 0x01, 0x28, 0x09,
0x52, 0x09, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x46, 0x69, 0x6c, 0x65, 0x12, 0x16, 0x0a, 0x06, 0x44,
0x65, 0x76, 0x69, 0x63, 0x65, 0x18, 0x16, 0x20, 0x01, 0x28, 0x09, 0x52, 0x06, 0x44, 0x65, 0x76,
0x69, 0x63, 0x65, 0x12, 0x1c, 0x0a, 0x09, 0x55, 0x73, 0x65, 0x54, 0x72, 0x69, 0x74, 0x6f, 0x6e,
0x18, 0x17, 0x20, 0x01, 0x28, 0x08, 0x52, 0x09, 0x55, 0x73, 0x65, 0x54, 0x72, 0x69, 0x74, 0x6f,
0x6e, 0x12, 0x24, 0x0a, 0x0d, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x42, 0x61, 0x73, 0x65, 0x4e, 0x61,
0x6d, 0x65, 0x18, 0x18, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0d, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x42,
0x61, 0x73, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x22, 0x3c, 0x0a, 0x06, 0x52, 0x65, 0x73, 0x75, 0x6c,
0x74, 0x12, 0x18, 0x0a, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x18, 0x01, 0x20, 0x01,
0x28, 0x09, 0x52, 0x07, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x12, 0x18, 0x0a, 0x07, 0x73,
0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x18, 0x02, 0x20, 0x01, 0x28, 0x08, 0x52, 0x07, 0x73, 0x75,
0x63, 0x63, 0x65, 0x73, 0x73, 0x22, 0x31, 0x0a, 0x0f, 0x45, 0x6d, 0x62, 0x65, 0x64, 0x64, 0x69,
0x6e, 0x67, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12, 0x1e, 0x0a, 0x0a, 0x65, 0x6d, 0x62, 0x65,
0x64, 0x64, 0x69, 0x6e, 0x67, 0x73, 0x18, 0x01, 0x20, 0x03, 0x28, 0x02, 0x52, 0x0a, 0x65, 0x6d,
0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x5b, 0x0a, 0x11, 0x54, 0x72, 0x61, 0x6e,
0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x12, 0x10, 0x0a,
0x03, 0x64, 0x73, 0x74, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x03, 0x64, 0x73, 0x74, 0x12,
0x1a, 0x0a, 0x08, 0x6c, 0x61, 0x6e, 0x67, 0x75, 0x61, 0x67, 0x65, 0x18, 0x03, 0x20, 0x01, 0x28,
0x09, 0x52, 0x08, 0x6c, 0x61, 0x6e, 0x67, 0x75, 0x61, 0x67, 0x65, 0x12, 0x18, 0x0a, 0x07, 0x74,
0x68, 0x72, 0x65, 0x61, 0x64, 0x73, 0x18, 0x04, 0x20, 0x01, 0x28, 0x0d, 0x52, 0x07, 0x74, 0x68,
0x72, 0x65, 0x61, 0x64, 0x73, 0x22, 0x5e, 0x0a, 0x10, 0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72,
0x69, 0x70, 0x74, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12, 0x36, 0x0a, 0x08, 0x73, 0x65, 0x67,
0x6d, 0x65, 0x6e, 0x74, 0x73, 0x18, 0x01, 0x20, 0x03, 0x28, 0x0b, 0x32, 0x1a, 0x2e, 0x62, 0x61,
0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74,
0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x1a, 0x19, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e,
0x64, 0x2e, 0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x52, 0x65, 0x73, 0x75,
0x6c, 0x74, 0x22, 0x00, 0x12, 0x2d, 0x0a, 0x03, 0x54, 0x54, 0x53, 0x12, 0x13, 0x2e, 0x62, 0x61,
0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x54, 0x54, 0x53, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74,
0x1a, 0x0f, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c,
0x74, 0x22, 0x00, 0x42, 0x5a, 0x0a, 0x19, 0x69, 0x6f, 0x2e, 0x73, 0x6b, 0x79, 0x6e, 0x65, 0x74,
0x2e, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x61, 0x69, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64,
0x42, 0x0e, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x41, 0x49, 0x42, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64,
0x50, 0x01, 0x5a, 0x2b, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67,
0x6f, 0x2d, 0x73, 0x6b, 0x79, 0x6e, 0x65, 0x74, 0x2f, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x41, 0x49,
0x2f, 0x70, 0x6b, 0x67, 0x2f, 0x67, 0x72, 0x70, 0x63, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62,
0x06, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x33,
0x53, 0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74, 0x52, 0x08, 0x73, 0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74,
0x73, 0x12, 0x12, 0x0a, 0x04, 0x74, 0x65, 0x78, 0x74, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x52,
0x04, 0x74, 0x65, 0x78, 0x74, 0x22, 0x77, 0x0a, 0x11, 0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72,
0x69, 0x70, 0x74, 0x53, 0x65, 0x67, 0x6d, 0x65, 0x6e, 0x74, 0x12, 0x0e, 0x0a, 0x02, 0x69, 0x64,
0x18, 0x01, 0x20, 0x01, 0x28, 0x05, 0x52, 0x02, 0x69, 0x64, 0x12, 0x14, 0x0a, 0x05, 0x73, 0x74,
0x61, 0x72, 0x74, 0x18, 0x02, 0x20, 0x01, 0x28, 0x03, 0x52, 0x05, 0x73, 0x74, 0x61, 0x72, 0x74,
0x12, 0x10, 0x0a, 0x03, 0x65, 0x6e, 0x64, 0x18, 0x03, 0x20, 0x01, 0x28, 0x03, 0x52, 0x03, 0x65,
0x6e, 0x64, 0x12, 0x12, 0x0a, 0x04, 0x74, 0x65, 0x78, 0x74, 0x18, 0x04, 0x20, 0x01, 0x28, 0x09,
0x52, 0x04, 0x74, 0x65, 0x78, 0x74, 0x12, 0x16, 0x0a, 0x06, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73,
0x18, 0x05, 0x20, 0x03, 0x28, 0x05, 0x52, 0x06, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x22, 0xe4,
0x01, 0x0a, 0x14, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x49, 0x6d, 0x61, 0x67, 0x65,
0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x12, 0x16, 0x0a, 0x06, 0x68, 0x65, 0x69, 0x67, 0x68,
0x74, 0x18, 0x01, 0x20, 0x01, 0x28, 0x05, 0x52, 0x06, 0x68, 0x65, 0x69, 0x67, 0x68, 0x74, 0x12,
0x14, 0x0a, 0x05, 0x77, 0x69, 0x64, 0x74, 0x68, 0x18, 0x02, 0x20, 0x01, 0x28, 0x05, 0x52, 0x05,
0x77, 0x69, 0x64, 0x74, 0x68, 0x12, 0x12, 0x0a, 0x04, 0x6d, 0x6f, 0x64, 0x65, 0x18, 0x03, 0x20,
0x01, 0x28, 0x05, 0x52, 0x04, 0x6d, 0x6f, 0x64, 0x65, 0x12, 0x12, 0x0a, 0x04, 0x73, 0x74, 0x65,
0x70, 0x18, 0x04, 0x20, 0x01, 0x28, 0x05, 0x52, 0x04, 0x73, 0x74, 0x65, 0x70, 0x12, 0x12, 0x0a,
0x04, 0x73, 0x65, 0x65, 0x64, 0x18, 0x05, 0x20, 0x01, 0x28, 0x05, 0x52, 0x04, 0x73, 0x65, 0x65,
0x64, 0x12, 0x27, 0x0a, 0x0f, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x76, 0x65, 0x5f, 0x70, 0x72,
0x6f, 0x6d, 0x70, 0x74, 0x18, 0x06, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0e, 0x70, 0x6f, 0x73, 0x69,
0x74, 0x69, 0x76, 0x65, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x12, 0x27, 0x0a, 0x0f, 0x6e, 0x65,
0x67, 0x61, 0x74, 0x69, 0x76, 0x65, 0x5f, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x18, 0x07, 0x20,
0x01, 0x28, 0x09, 0x52, 0x0e, 0x6e, 0x65, 0x67, 0x61, 0x74, 0x69, 0x76, 0x65, 0x50, 0x72, 0x6f,
0x6d, 0x70, 0x74, 0x12, 0x10, 0x0a, 0x03, 0x64, 0x73, 0x74, 0x18, 0x08, 0x20, 0x01, 0x28, 0x09,
0x52, 0x03, 0x64, 0x73, 0x74, 0x22, 0x48, 0x0a, 0x0a, 0x54, 0x54, 0x53, 0x52, 0x65, 0x71, 0x75,
0x65, 0x73, 0x74, 0x12, 0x12, 0x0a, 0x04, 0x74, 0x65, 0x78, 0x74, 0x18, 0x01, 0x20, 0x01, 0x28,
0x09, 0x52, 0x04, 0x74, 0x65, 0x78, 0x74, 0x12, 0x14, 0x0a, 0x05, 0x6d, 0x6f, 0x64, 0x65, 0x6c,
0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x05, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x10, 0x0a,
0x03, 0x64, 0x73, 0x74, 0x18, 0x03, 0x20, 0x01, 0x28, 0x09, 0x52, 0x03, 0x64, 0x73, 0x74, 0x32,
0xeb, 0x03, 0x0a, 0x07, 0x42, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x12, 0x32, 0x0a, 0x06, 0x48,
0x65, 0x61, 0x6c, 0x74, 0x68, 0x12, 0x16, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e,
0x48, 0x65, 0x61, 0x6c, 0x74, 0x68, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x1a, 0x0e, 0x2e,
0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x70, 0x6c, 0x79, 0x22, 0x00, 0x12,
0x34, 0x0a, 0x07, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x12, 0x17, 0x2e, 0x62, 0x61, 0x63,
0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f, 0x70, 0x74, 0x69,
0x6f, 0x6e, 0x73, 0x1a, 0x0e, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65,
0x70, 0x6c, 0x79, 0x22, 0x00, 0x12, 0x35, 0x0a, 0x09, 0x4c, 0x6f, 0x61, 0x64, 0x4d, 0x6f, 0x64,
0x65, 0x6c, 0x12, 0x15, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x4d, 0x6f, 0x64,
0x65, 0x6c, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0f, 0x2e, 0x62, 0x61, 0x63, 0x6b,
0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x22, 0x00, 0x12, 0x3c, 0x0a, 0x0d,
0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x53, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x12, 0x17, 0x2e,
0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f,
0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x1a, 0x0e, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64,
0x2e, 0x52, 0x65, 0x70, 0x6c, 0x79, 0x22, 0x00, 0x30, 0x01, 0x12, 0x40, 0x0a, 0x09, 0x45, 0x6d,
0x62, 0x65, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x12, 0x17, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e,
0x64, 0x2e, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73,
0x1a, 0x18, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x45, 0x6d, 0x62, 0x65, 0x64,
0x64, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x22, 0x00, 0x12, 0x41, 0x0a, 0x0d,
0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x12, 0x1d, 0x2e,
0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65,
0x49, 0x6d, 0x61, 0x67, 0x65, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x1a, 0x0f, 0x2e, 0x62,
0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x22, 0x00, 0x12,
0x4d, 0x0a, 0x12, 0x41, 0x75, 0x64, 0x69, 0x6f, 0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69,
0x70, 0x74, 0x69, 0x6f, 0x6e, 0x12, 0x1a, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e,
0x54, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73,
0x74, 0x1a, 0x19, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x54, 0x72, 0x61, 0x6e,
0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x22, 0x00, 0x12, 0x2d,
0x0a, 0x03, 0x54, 0x54, 0x53, 0x12, 0x13, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x2e,
0x54, 0x54, 0x53, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x1a, 0x0f, 0x2e, 0x62, 0x61, 0x63,
0x6b, 0x65, 0x6e, 0x64, 0x2e, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x22, 0x00, 0x42, 0x5a, 0x0a,
0x19, 0x69, 0x6f, 0x2e, 0x73, 0x6b, 0x79, 0x6e, 0x65, 0x74, 0x2e, 0x6c, 0x6f, 0x63, 0x61, 0x6c,
0x61, 0x69, 0x2e, 0x62, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x42, 0x0e, 0x4c, 0x6f, 0x63, 0x61,
0x6c, 0x41, 0x49, 0x42, 0x61, 0x63, 0x6b, 0x65, 0x6e, 0x64, 0x50, 0x01, 0x5a, 0x2b, 0x67, 0x69,
0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x6f, 0x2d, 0x73, 0x6b, 0x79, 0x6e,
0x65, 0x74, 0x2f, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x41, 0x49, 0x2f, 0x70, 0x6b, 0x67, 0x2f, 0x67,
0x72, 0x70, 0x63, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62, 0x06, 0x70, 0x72, 0x6f, 0x74, 0x6f,
0x33,
}
var (

6
pkg/grpc/proto/backend.proto
View file

@ -89,6 +89,12 @@ message ModelOptions {
float RopeFreqScale = 18;
float RMSNormEps = 19;
int32 NGQA = 20;
string ModelFile = 21;
// AutoGPTQ
string Device = 22;
bool UseTriton = 23;
string ModelBaseName = 24;
}
message Result {

2
pkg/grpc/transcribe/whisper.go
View file

@ -17,7 +17,7 @@ type Whisper struct {
func (sd *Whisper) Load(opts *pb.ModelOptions) error {
// Note: the Model here is a path to a directory containing the model files
w, err := whisper.New(opts.Model)
w, err := whisper.New(opts.ModelFile)
sd.whisper = w
return err
}

4
pkg/grpc/tts/piper.go
View file

@ -18,8 +18,8 @@ type Piper struct {
}
func (sd *Piper) Load(opts *pb.ModelOptions) error {
if filepath.Ext(opts.Model) != ".onnx" {
return fmt.Errorf("unsupported model type %s (should end with .onnx)", opts.Model)
if filepath.Ext(opts.ModelFile) != ".onnx" {
return fmt.Errorf("unsupported model type %s (should end with .onnx)", opts.ModelFile)
}
var err error
// Note: the Model here is a path to a directory containing the model files

29
pkg/model/initializers.go
View file

@ -124,8 +124,8 @@ func (ml *ModelLoader) startProcess(grpcProcess, id string, serverAddress string
// starts the grpcModelProcess for the backend, and returns a grpc client
// It also loads the model
func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string) (*grpc.Client, error) {
return func(s string) (*grpc.Client, error) {
func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string, string) (*grpc.Client, error) {
return func(modelName, modelFile string) (*grpc.Client, error) {
log.Debug().Msgf("Loading GRPC Model %s: %+v", backend, *o)
var client *grpc.Client
@ -148,7 +148,7 @@ func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string) (*grpc
return nil, fmt.Errorf("failed allocating free ports: %s", err.Error())
}
// Make sure the process is executable
if err := ml.startProcess(uri, o.modelFile, serverAddress); err != nil {
if err := ml.startProcess(uri, o.model, serverAddress); err != nil {
return nil, err
}
@ -172,7 +172,7 @@ func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string) (*grpc
}
// Make sure the process is executable
if err := ml.startProcess(grpcProcess, o.modelFile, serverAddress); err != nil {
if err := ml.startProcess(grpcProcess, o.model, serverAddress); err != nil {
return nil, err
}
@ -198,7 +198,8 @@ func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string) (*grpc
}
options := *o.gRPCOptions
options.Model = s
options.Model = modelName
options.ModelFile = modelFile
log.Debug().Msgf("GRPC: Loading model with options: %+v", options)
@ -217,14 +218,14 @@ func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string) (*grpc
func (ml *ModelLoader) BackendLoader(opts ...Option) (model *grpc.Client, err error) {
o := NewOptions(opts...)
log.Debug().Msgf("Loading model %s from %s", o.backendString, o.modelFile)
log.Debug().Msgf("Loading model %s from %s", o.backendString, o.model)
backend := strings.ToLower(o.backendString)
// if an external backend is provided, use it
_, externalBackendExists := o.externalBackends[backend]
if externalBackendExists {
return ml.LoadModel(o.modelFile, ml.grpcModel(backend, o))
return ml.LoadModel(o.model, ml.grpcModel(backend, o))
}
switch backend {
@ -232,13 +233,13 @@ func (ml *ModelLoader) BackendLoader(opts ...Option) (model *grpc.Client, err er
MPTBackend, Gpt2Backend, FalconBackend,
GPTNeoXBackend, ReplitBackend, StarcoderBackend, BloomzBackend,
RwkvBackend, LCHuggingFaceBackend, BertEmbeddingsBackend, FalconGGMLBackend, StableDiffusionBackend, WhisperBackend:
return ml.LoadModel(o.modelFile, ml.grpcModel(backend, o))
return ml.LoadModel(o.model, ml.grpcModel(backend, o))
case Gpt4AllLlamaBackend, Gpt4AllMptBackend, Gpt4AllJBackend, Gpt4All:
o.gRPCOptions.LibrarySearchPath = filepath.Join(o.assetDir, "backend-assets", "gpt4all")
return ml.LoadModel(o.modelFile, ml.grpcModel(Gpt4All, o))
return ml.LoadModel(o.model, ml.grpcModel(Gpt4All, o))
case PiperBackend:
o.gRPCOptions.LibrarySearchPath = filepath.Join(o.assetDir, "backend-assets", "espeak-ng-data")
return ml.LoadModel(o.modelFile, ml.grpcModel(PiperBackend, o))
return ml.LoadModel(o.model, ml.grpcModel(PiperBackend, o))
default:
return nil, fmt.Errorf("backend unsupported: %s", o.backendString)
}
@ -249,8 +250,8 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (*grpc.Client, error) {
// Is this really needed? BackendLoader already does this
ml.mu.Lock()
if m := ml.checkIsLoaded(o.modelFile); m != nil {
log.Debug().Msgf("Model '%s' already loaded", o.modelFile)
if m := ml.checkIsLoaded(o.model); m != nil {
log.Debug().Msgf("Model '%s' already loaded", o.model)
ml.mu.Unlock()
return m, nil
}
@ -263,13 +264,13 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (*grpc.Client, error) {
for _, b := range o.externalBackends {
allBackendsToAutoLoad = append(allBackendsToAutoLoad, b)
}
log.Debug().Msgf("Loading model '%s' greedly from all the available backends: %s", o.modelFile, strings.Join(allBackendsToAutoLoad, ", "))
log.Debug().Msgf("Loading model '%s' greedly from all the available backends: %s", o.model, strings.Join(allBackendsToAutoLoad, ", "))
for _, b := range allBackendsToAutoLoad {
log.Debug().Msgf("[%s] Attempting to load", b)
options := []Option{
WithBackendString(b),
WithModelFile(o.modelFile),
WithModel(o.model),
WithLoadGRPCLLMModelOpts(o.gRPCOptions),
WithThreads(o.threads),
WithAssetDir(o.assetDir),

4
pkg/model/loader.go
View file

@ -98,7 +98,7 @@ func (ml *ModelLoader) ListModels() ([]string, error) {
return models, nil
}
func (ml *ModelLoader) LoadModel(modelName string, loader func(string) (*grpc.Client, error)) (*grpc.Client, error) {
func (ml *ModelLoader) LoadModel(modelName string, loader func(string, string) (*grpc.Client, error)) (*grpc.Client, error) {
ml.mu.Lock()
defer ml.mu.Unlock()
@ -111,7 +111,7 @@ func (ml *ModelLoader) LoadModel(modelName string, loader func(string) (*grpc.Cl
modelFile := filepath.Join(ml.ModelPath, modelName)
log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
model, err := loader(modelFile)
model, err := loader(modelName, modelFile)
if err != nil {
return nil, err
}

6
pkg/model/options.go
View file

@ -8,7 +8,7 @@ import (
type Options struct {
backendString string
modelFile string
model string
threads uint32
assetDir string
context context.Context
@ -35,9 +35,9 @@ func WithBackendString(backend string) Option {
}
}
func WithModelFile(modelFile string) Option {
func WithModel(modelFile string) Option {
return func(o *Options) {
o.modelFile = modelFile
o.model = modelFile
}
}