refactor: move backends into the backends directory (#1279)

* refactor: move backends into the backends directory

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* refactor: move main close to implementation for every backend

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

backend/go/llm/bert/bert.go

@@ -0,0 +1,34 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	bert "github.com/go-skynet/go-bert.cpp"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+)
+
+type Embeddings struct {
+	base.SingleThread
+	bert *bert.Bert
+}
+
+func (llm *Embeddings) Load(opts *pb.ModelOptions) error {
+	model, err := bert.New(opts.ModelFile)
+	llm.bert = model
+	return err
+}
+
+func (llm *Embeddings) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
+
+	if len(opts.EmbeddingTokens) > 0 {
+		tokens := []int{}
+		for _, t := range opts.EmbeddingTokens {
+			tokens = append(tokens, int(t))
+		}
+		return llm.bert.TokenEmbeddings(tokens, bert.SetThreads(int(opts.Threads)))
+	}
+
+	return llm.bert.Embeddings(opts.Embeddings, bert.SetThreads(int(opts.Threads)))
+}

backend/go/llm/bert/main.go

@@ -0,0 +1,21 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &Embeddings{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/dolly/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Dolly{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/falcon-ggml/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Falcon{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/gpt2/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.GPT2{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/gpt4all/gpt4all.go

@@ -0,0 +1,62 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	gpt4all "github.com/nomic-ai/gpt4all/gpt4all-bindings/golang"
+)
+
+type LLM struct {
+	base.SingleThread
+
+	gpt4all *gpt4all.Model
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	model, err := gpt4all.New(opts.ModelFile,
+		gpt4all.SetThreads(int(opts.Threads)),
+		gpt4all.SetLibrarySearchPath(opts.LibrarySearchPath))
+	llm.gpt4all = model
+	return err
+}
+
+func buildPredictOptions(opts *pb.PredictOptions) []gpt4all.PredictOption {
+	predictOptions := []gpt4all.PredictOption{
+		gpt4all.SetTemperature(float64(opts.Temperature)),
+		gpt4all.SetTopP(float64(opts.TopP)),
+		gpt4all.SetTopK(int(opts.TopK)),
+		gpt4all.SetTokens(int(opts.Tokens)),
+	}
+
+	if opts.Batch != 0 {
+		predictOptions = append(predictOptions, gpt4all.SetBatch(int(opts.Batch)))
+	}
+	return predictOptions
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.gpt4all.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	predictOptions := buildPredictOptions(opts)
+
+	go func() {
+		llm.gpt4all.SetTokenCallback(func(token string) bool {
+			results <- token
+			return true
+		})
+		_, err := llm.gpt4all.Predict(opts.Prompt, predictOptions...)
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		llm.gpt4all.SetTokenCallback(nil)
+		close(results)
+	}()
+
+	return nil
+}

backend/go/llm/gpt4all/main.go

@@ -0,0 +1,21 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/gptj/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.GPTJ{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/gptneox/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.GPTNeoX{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/langchain/langchain.go

@@ -0,0 +1,58 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	"github.com/go-skynet/LocalAI/pkg/langchain"
+)
+
+type LLM struct {
+	base.Base
+
+	langchain *langchain.HuggingFace
+	model     string
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	llm.langchain, _ = langchain.NewHuggingFace(opts.Model)
+	llm.model = opts.Model
+	return nil
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	o := []langchain.PredictOption{
+		langchain.SetModel(llm.model),
+		langchain.SetMaxTokens(int(opts.Tokens)),
+		langchain.SetTemperature(float64(opts.Temperature)),
+		langchain.SetStopWords(opts.StopPrompts),
+	}
+	pred, err := llm.langchain.PredictHuggingFace(opts.Prompt, o...)
+	if err != nil {
+		return "", err
+	}
+	return pred.Completion, nil
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	o := []langchain.PredictOption{
+		langchain.SetModel(llm.model),
+		langchain.SetMaxTokens(int(opts.Tokens)),
+		langchain.SetTemperature(float64(opts.Temperature)),
+		langchain.SetStopWords(opts.StopPrompts),
+	}
+	go func() {
+		res, err := llm.langchain.PredictHuggingFace(opts.Prompt, o...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res.Completion
+		close(results)
+	}()
+
+	return nil
+}

backend/go/llm/langchain/main.go

@@ -0,0 +1,21 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/llama-stable/llama.go

@@ -0,0 +1,204 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	"github.com/go-skynet/go-llama.cpp"
+)
+
+type LLM struct {
+	base.SingleThread
+
+	llama *llama.LLama
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	ropeFreqBase := float32(10000)
+	ropeFreqScale := float32(1)
+
+	if opts.RopeFreqBase != 0 {
+		ropeFreqBase = opts.RopeFreqBase
+	}
+	if opts.RopeFreqScale != 0 {
+		ropeFreqScale = opts.RopeFreqScale
+	}
+
+	llamaOpts := []llama.ModelOption{
+		llama.WithRopeFreqBase(ropeFreqBase),
+		llama.WithRopeFreqScale(ropeFreqScale),
+	}
+
+	if opts.NGQA != 0 {
+		llamaOpts = append(llamaOpts, llama.WithGQA(int(opts.NGQA)))
+	}
+
+	if opts.RMSNormEps != 0 {
+		llamaOpts = append(llamaOpts, llama.WithRMSNormEPS(opts.RMSNormEps))
+	}
+
+	if opts.ContextSize != 0 {
+		llamaOpts = append(llamaOpts, llama.SetContext(int(opts.ContextSize)))
+	}
+	if opts.F16Memory {
+		llamaOpts = append(llamaOpts, llama.EnableF16Memory)
+	}
+	if opts.Embeddings {
+		llamaOpts = append(llamaOpts, llama.EnableEmbeddings)
+	}
+	if opts.NGPULayers != 0 {
+		llamaOpts = append(llamaOpts, llama.SetGPULayers(int(opts.NGPULayers)))
+	}
+
+	llamaOpts = append(llamaOpts, llama.SetMMap(opts.MMap))
+	llamaOpts = append(llamaOpts, llama.SetMainGPU(opts.MainGPU))
+	llamaOpts = append(llamaOpts, llama.SetTensorSplit(opts.TensorSplit))
+	if opts.NBatch != 0 {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(int(opts.NBatch)))
+	} else {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(512))
+	}
+
+	if opts.NUMA {
+		llamaOpts = append(llamaOpts, llama.EnableNUMA)
+	}
+
+	if opts.LowVRAM {
+		llamaOpts = append(llamaOpts, llama.EnabelLowVRAM)
+	}
+
+	model, err := llama.New(opts.ModelFile, llamaOpts...)
+	llm.llama = model
+
+	return err
+}
+
+func buildPredictOptions(opts *pb.PredictOptions) []llama.PredictOption {
+	ropeFreqBase := float32(10000)
+	ropeFreqScale := float32(1)
+
+	if opts.RopeFreqBase != 0 {
+		ropeFreqBase = opts.RopeFreqBase
+	}
+	if opts.RopeFreqScale != 0 {
+		ropeFreqScale = opts.RopeFreqScale
+	}
+	predictOptions := []llama.PredictOption{
+		llama.SetTemperature(opts.Temperature),
+		llama.SetTopP(opts.TopP),
+		llama.SetTopK(int(opts.TopK)),
+		llama.SetTokens(int(opts.Tokens)),
+		llama.SetThreads(int(opts.Threads)),
+		llama.WithGrammar(opts.Grammar),
+		llama.SetRopeFreqBase(ropeFreqBase),
+		llama.SetRopeFreqScale(ropeFreqScale),
+		llama.SetNegativePromptScale(opts.NegativePromptScale),
+		llama.SetNegativePrompt(opts.NegativePrompt),
+	}
+
+	if opts.PromptCacheAll {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheAll)
+	}
+
+	if opts.PromptCacheRO {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheRO)
+	}
+
+	// Expected absolute path
+	if opts.PromptCachePath != "" {
+		predictOptions = append(predictOptions, llama.SetPathPromptCache(opts.PromptCachePath))
+	}
+
+	if opts.Mirostat != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostat(int(opts.Mirostat)))
+	}
+
+	if opts.MirostatETA != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatETA(opts.MirostatETA))
+	}
+
+	if opts.MirostatTAU != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatTAU(opts.MirostatTAU))
+	}
+
+	if opts.Debug {
+		predictOptions = append(predictOptions, llama.Debug)
+	}
+
+	predictOptions = append(predictOptions, llama.SetStopWords(opts.StopPrompts...))
+
+	if opts.PresencePenalty != 0 {
+		predictOptions = append(predictOptions, llama.SetPenalty(opts.PresencePenalty))
+	}
+
+	if opts.NKeep != 0 {
+		predictOptions = append(predictOptions, llama.SetNKeep(int(opts.NKeep)))
+	}
+
+	if opts.Batch != 0 {
+		predictOptions = append(predictOptions, llama.SetBatch(int(opts.Batch)))
+	}
+
+	if opts.F16KV {
+		predictOptions = append(predictOptions, llama.EnableF16KV)
+	}
+
+	if opts.IgnoreEOS {
+		predictOptions = append(predictOptions, llama.IgnoreEOS)
+	}
+
+	if opts.Seed != 0 {
+		predictOptions = append(predictOptions, llama.SetSeed(int(opts.Seed)))
+	}
+
+	//predictOptions = append(predictOptions, llama.SetLogitBias(c.Seed))
+
+	predictOptions = append(predictOptions, llama.SetFrequencyPenalty(opts.FrequencyPenalty))
+	predictOptions = append(predictOptions, llama.SetMlock(opts.MLock))
+	predictOptions = append(predictOptions, llama.SetMemoryMap(opts.MMap))
+	predictOptions = append(predictOptions, llama.SetPredictionMainGPU(opts.MainGPU))
+	predictOptions = append(predictOptions, llama.SetPredictionTensorSplit(opts.TensorSplit))
+	predictOptions = append(predictOptions, llama.SetTailFreeSamplingZ(opts.TailFreeSamplingZ))
+	predictOptions = append(predictOptions, llama.SetTypicalP(opts.TypicalP))
+	return predictOptions
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.llama.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	predictOptions := buildPredictOptions(opts)
+
+	predictOptions = append(predictOptions, llama.SetTokenCallback(func(token string) bool {
+		results <- token
+		return true
+	}))
+
+	go func() {
+		_, err := llm.llama.Predict(opts.Prompt, predictOptions...)
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		close(results)
+	}()
+
+	return nil
+}
+
+func (llm *LLM) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
+	predictOptions := buildPredictOptions(opts)
+
+	if len(opts.EmbeddingTokens) > 0 {
+		tokens := []int{}
+		for _, t := range opts.EmbeddingTokens {
+			tokens = append(tokens, int(t))
+		}
+		return llm.llama.TokenEmbeddings(tokens, predictOptions...)
+	}
+
+	return llm.llama.Embeddings(opts.Embeddings, predictOptions...)
+}

backend/go/llm/llama-stable/main.go

@@ -0,0 +1,19 @@
+package main
+
+import (
+	"flag"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/llama/llama.go

@@ -0,0 +1,257 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+	"path/filepath"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	"github.com/go-skynet/go-llama.cpp"
+)
+
+type LLM struct {
+	base.SingleThread
+
+	llama      *llama.LLama
+	draftModel *llama.LLama
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	ropeFreqBase := float32(10000)
+	ropeFreqScale := float32(1)
+
+	if opts.RopeFreqBase != 0 {
+		ropeFreqBase = opts.RopeFreqBase
+	}
+	if opts.RopeFreqScale != 0 {
+		ropeFreqScale = opts.RopeFreqScale
+	}
+
+	llamaOpts := []llama.ModelOption{
+		llama.WithRopeFreqBase(ropeFreqBase),
+		llama.WithRopeFreqScale(ropeFreqScale),
+	}
+
+	if opts.NoMulMatQ {
+		llamaOpts = append(llamaOpts, llama.SetMulMatQ(false))
+	}
+
+	// Get base path of opts.ModelFile and use the same for lora (assume the same path)
+	basePath := filepath.Dir(opts.ModelFile)
+
+	if opts.LoraAdapter != "" {
+		llamaOpts = append(llamaOpts, llama.SetLoraAdapter(filepath.Join(basePath, opts.LoraAdapter)))
+	}
+
+	if opts.LoraBase != "" {
+		llamaOpts = append(llamaOpts, llama.SetLoraBase(filepath.Join(basePath, opts.LoraBase)))
+	}
+
+	if opts.ContextSize != 0 {
+		llamaOpts = append(llamaOpts, llama.SetContext(int(opts.ContextSize)))
+	}
+	if opts.F16Memory {
+		llamaOpts = append(llamaOpts, llama.EnableF16Memory)
+	}
+	if opts.Embeddings {
+		llamaOpts = append(llamaOpts, llama.EnableEmbeddings)
+	}
+	if opts.NGPULayers != 0 {
+		llamaOpts = append(llamaOpts, llama.SetGPULayers(int(opts.NGPULayers)))
+	}
+
+	llamaOpts = append(llamaOpts, llama.SetMMap(opts.MMap))
+	llamaOpts = append(llamaOpts, llama.SetMainGPU(opts.MainGPU))
+	llamaOpts = append(llamaOpts, llama.SetTensorSplit(opts.TensorSplit))
+	if opts.NBatch != 0 {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(int(opts.NBatch)))
+	} else {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(512))
+	}
+
+	if opts.NUMA {
+		llamaOpts = append(llamaOpts, llama.EnableNUMA)
+	}
+
+	if opts.LowVRAM {
+		llamaOpts = append(llamaOpts, llama.EnabelLowVRAM)
+	}
+
+	if opts.DraftModel != "" {
+		// https://github.com/ggerganov/llama.cpp/blob/71ca2fad7d6c0ef95ef9944fb3a1a843e481f314/examples/speculative/speculative.cpp#L40
+		llamaOpts = append(llamaOpts, llama.SetPerplexity(true))
+	}
+
+	model, err := llama.New(opts.ModelFile, llamaOpts...)
+
+	if opts.DraftModel != "" {
+		// opts.DraftModel is relative to opts.ModelFile, so we need to get the basepath of opts.ModelFile
+		if !filepath.IsAbs(opts.DraftModel) {
+			dir := filepath.Dir(opts.ModelFile)
+			opts.DraftModel = filepath.Join(dir, opts.DraftModel)
+		}
+
+		draftModel, err := llama.New(opts.DraftModel, llamaOpts...)
+		if err != nil {
+			return err
+		}
+		llm.draftModel = draftModel
+	}
+
+	llm.llama = model
+
+	return err
+}
+
+func buildPredictOptions(opts *pb.PredictOptions) []llama.PredictOption {
+	ropeFreqBase := float32(10000)
+	ropeFreqScale := float32(1)
+
+	if opts.RopeFreqBase != 0 {
+		ropeFreqBase = opts.RopeFreqBase
+	}
+	if opts.RopeFreqScale != 0 {
+		ropeFreqScale = opts.RopeFreqScale
+	}
+	predictOptions := []llama.PredictOption{
+		llama.SetTemperature(opts.Temperature),
+		llama.SetTopP(opts.TopP),
+		llama.SetTopK(int(opts.TopK)),
+		llama.SetTokens(int(opts.Tokens)),
+		llama.SetThreads(int(opts.Threads)),
+		llama.WithGrammar(opts.Grammar),
+		llama.SetRopeFreqBase(ropeFreqBase),
+		llama.SetRopeFreqScale(ropeFreqScale),
+		llama.SetNegativePromptScale(opts.NegativePromptScale),
+		llama.SetNegativePrompt(opts.NegativePrompt),
+	}
+
+	if opts.PromptCacheAll {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheAll)
+	}
+
+	if opts.PromptCacheRO {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheRO)
+	}
+
+	// Expected absolute path
+	if opts.PromptCachePath != "" {
+		predictOptions = append(predictOptions, llama.SetPathPromptCache(opts.PromptCachePath))
+	}
+
+	if opts.Mirostat != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostat(int(opts.Mirostat)))
+	}
+
+	if opts.MirostatETA != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatETA(opts.MirostatETA))
+	}
+
+	if opts.MirostatTAU != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatTAU(opts.MirostatTAU))
+	}
+
+	if opts.Debug {
+		predictOptions = append(predictOptions, llama.Debug)
+	}
+
+	predictOptions = append(predictOptions, llama.SetStopWords(opts.StopPrompts...))
+
+	if opts.PresencePenalty != 0 {
+		predictOptions = append(predictOptions, llama.SetPenalty(opts.PresencePenalty))
+	}
+
+	if opts.NKeep != 0 {
+		predictOptions = append(predictOptions, llama.SetNKeep(int(opts.NKeep)))
+	}
+
+	if opts.Batch != 0 {
+		predictOptions = append(predictOptions, llama.SetBatch(int(opts.Batch)))
+	}
+
+	if opts.F16KV {
+		predictOptions = append(predictOptions, llama.EnableF16KV)
+	}
+
+	if opts.IgnoreEOS {
+		predictOptions = append(predictOptions, llama.IgnoreEOS)
+	}
+
+	if opts.Seed != 0 {
+		predictOptions = append(predictOptions, llama.SetSeed(int(opts.Seed)))
+	}
+
+	if opts.NDraft != 0 {
+		predictOptions = append(predictOptions, llama.SetNDraft(int(opts.NDraft)))
+	}
+	//predictOptions = append(predictOptions, llama.SetLogitBias(c.Seed))
+
+	predictOptions = append(predictOptions, llama.SetFrequencyPenalty(opts.FrequencyPenalty))
+	predictOptions = append(predictOptions, llama.SetMlock(opts.MLock))
+	predictOptions = append(predictOptions, llama.SetMemoryMap(opts.MMap))
+	predictOptions = append(predictOptions, llama.SetPredictionMainGPU(opts.MainGPU))
+	predictOptions = append(predictOptions, llama.SetPredictionTensorSplit(opts.TensorSplit))
+	predictOptions = append(predictOptions, llama.SetTailFreeSamplingZ(opts.TailFreeSamplingZ))
+	predictOptions = append(predictOptions, llama.SetTypicalP(opts.TypicalP))
+	return predictOptions
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	if llm.draftModel != nil {
+		return llm.llama.SpeculativeSampling(llm.draftModel, opts.Prompt, buildPredictOptions(opts)...)
+	}
+	return llm.llama.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	predictOptions := buildPredictOptions(opts)
+
+	predictOptions = append(predictOptions, llama.SetTokenCallback(func(token string) bool {
+		results <- token
+		return true
+	}))
+
+	go func() {
+		var err error
+		if llm.draftModel != nil {
+			_, err = llm.llama.SpeculativeSampling(llm.draftModel, opts.Prompt, buildPredictOptions(opts)...)
+		} else {
+			_, err = llm.llama.Predict(opts.Prompt, predictOptions...)
+		}
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		close(results)
+	}()
+
+	return nil
+}
+
+func (llm *LLM) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
+	predictOptions := buildPredictOptions(opts)
+
+	if len(opts.EmbeddingTokens) > 0 {
+		tokens := []int{}
+		for _, t := range opts.EmbeddingTokens {
+			tokens = append(tokens, int(t))
+		}
+		return llm.llama.TokenEmbeddings(tokens, predictOptions...)
+	}
+
+	return llm.llama.Embeddings(opts.Embeddings, predictOptions...)
+}
+
+func (llm *LLM) TokenizeString(opts *pb.PredictOptions) (pb.TokenizationResponse, error) {
+	predictOptions := buildPredictOptions(opts)
+	l, tokens, err := llm.llama.TokenizeString(opts.Prompt, predictOptions...)
+	if err != nil {
+		return pb.TokenizationResponse{}, err
+	}
+	return pb.TokenizationResponse{
+		Length: l,
+		Tokens: tokens,
+	}, nil
+}

backend/go/llm/llama/main.go

@@ -0,0 +1,23 @@
+package main
+
+// GRPC Falcon server
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/mpt/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.MPT{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/replit/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Replit{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/rwkv/main.go

@@ -0,0 +1,21 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/rwkv/rwkv.go

@@ -0,0 +1,95 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+	"path/filepath"
+
+	"github.com/donomii/go-rwkv.cpp"
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+)
+
+const tokenizerSuffix = ".tokenizer.json"
+
+type LLM struct {
+	base.SingleThread
+
+	rwkv *rwkv.RwkvState
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	tokenizerFile := opts.Tokenizer
+	if tokenizerFile == "" {
+		modelFile := filepath.Base(opts.ModelFile)
+		tokenizerFile = modelFile + tokenizerSuffix
+	}
+	modelPath := filepath.Dir(opts.ModelFile)
+	tokenizerPath := filepath.Join(modelPath, tokenizerFile)
+
+	model := rwkv.LoadFiles(opts.ModelFile, tokenizerPath, uint32(opts.GetThreads()))
+
+	if model == nil {
+		return fmt.Errorf("could not load model")
+	}
+	llm.rwkv = model
+	return nil
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	stopWord := "\n"
+	if len(opts.StopPrompts) > 0 {
+		stopWord = opts.StopPrompts[0]
+	}
+
+	if err := llm.rwkv.ProcessInput(opts.Prompt); err != nil {
+		return "", err
+	}
+
+	response := llm.rwkv.GenerateResponse(int(opts.Tokens), stopWord, float32(opts.Temperature), float32(opts.TopP), nil)
+
+	return response, nil
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+
+		stopWord := "\n"
+		if len(opts.StopPrompts) > 0 {
+			stopWord = opts.StopPrompts[0]
+		}
+
+		if err := llm.rwkv.ProcessInput(opts.Prompt); err != nil {
+			fmt.Println("Error processing input: ", err)
+			return
+		}
+
+		llm.rwkv.GenerateResponse(int(opts.Tokens), stopWord, float32(opts.Temperature), float32(opts.TopP), func(s string) bool {
+			results <- s
+			return true
+		})
+		close(results)
+	}()
+
+	return nil
+}
+
+func (llm *LLM) TokenizeString(opts *pb.PredictOptions) (pb.TokenizationResponse, error) {
+	tokens, err := llm.rwkv.Tokenizer.Encode(opts.Prompt)
+	if err != nil {
+		return pb.TokenizationResponse{}, err
+	}
+
+	l := len(tokens)
+	i32Tokens := make([]int32, l)
+
+	for i, t := range tokens {
+		i32Tokens[i] = int32(t.ID)
+	}
+
+	return pb.TokenizationResponse{
+		Length: int32(l),
+		Tokens: i32Tokens,
+	}, nil
+}

backend/go/llm/starcoder/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/backend/go/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Starcoder{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/transformers/dolly.go

@@ -0,0 +1,44 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type Dolly struct {
+	base.SingleThread
+
+	dolly *transformers.Dolly
+}
+
+func (llm *Dolly) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewDolly(opts.ModelFile)
+	llm.dolly = model
+	return err
+}
+
+func (llm *Dolly) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.dolly.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *Dolly) PredictStream(opts *pb.PredictOptions, results chan string) error {
+
+	go func() {
+		res, err := llm.dolly.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+
+	return nil
+}

backend/go/llm/transformers/falcon.go

@@ -0,0 +1,43 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type Falcon struct {
+	base.SingleThread
+
+	falcon *transformers.Falcon
+}
+
+func (llm *Falcon) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewFalcon(opts.ModelFile)
+	llm.falcon = model
+	return err
+}
+
+func (llm *Falcon) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.falcon.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *Falcon) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.falcon.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+
+	return nil
+}

backend/go/llm/transformers/gpt2.go

@@ -0,0 +1,42 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type GPT2 struct {
+	base.SingleThread
+
+	gpt2 *transformers.GPT2
+}
+
+func (llm *GPT2) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.New(opts.ModelFile)
+	llm.gpt2 = model
+	return err
+}
+
+func (llm *GPT2) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.gpt2.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *GPT2) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.gpt2.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+	return nil
+}

backend/go/llm/transformers/gptj.go

@@ -0,0 +1,42 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type GPTJ struct {
+	base.SingleThread
+
+	gptj *transformers.GPTJ
+}
+
+func (llm *GPTJ) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewGPTJ(opts.ModelFile)
+	llm.gptj = model
+	return err
+}
+
+func (llm *GPTJ) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.gptj.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *GPTJ) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.gptj.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+	return nil
+}

backend/go/llm/transformers/gptneox.go

@@ -0,0 +1,42 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type GPTNeoX struct {
+	base.SingleThread
+
+	gptneox *transformers.GPTNeoX
+}
+
+func (llm *GPTNeoX) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewGPTNeoX(opts.ModelFile)
+	llm.gptneox = model
+	return err
+}
+
+func (llm *GPTNeoX) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.gptneox.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *GPTNeoX) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.gptneox.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+	return nil
+}

backend/go/llm/transformers/mpt.go

@@ -0,0 +1,42 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type MPT struct {
+	base.SingleThread
+
+	mpt *transformers.MPT
+}
+
+func (llm *MPT) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewMPT(opts.ModelFile)
+	llm.mpt = model
+	return err
+}
+
+func (llm *MPT) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.mpt.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *MPT) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.mpt.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+	return nil
+}

backend/go/llm/transformers/predict.go

@@ -0,0 +1,26 @@
+package transformers
+
+import (
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+func buildPredictOptions(opts *pb.PredictOptions) []transformers.PredictOption {
+	predictOptions := []transformers.PredictOption{
+		transformers.SetTemperature(float64(opts.Temperature)),
+		transformers.SetTopP(float64(opts.TopP)),
+		transformers.SetTopK(int(opts.TopK)),
+		transformers.SetTokens(int(opts.Tokens)),
+		transformers.SetThreads(int(opts.Threads)),
+	}
+
+	if opts.Batch != 0 {
+		predictOptions = append(predictOptions, transformers.SetBatch(int(opts.Batch)))
+	}
+
+	if opts.Seed != 0 {
+		predictOptions = append(predictOptions, transformers.SetSeed(int(opts.Seed)))
+	}
+
+	return predictOptions
+}

backend/go/llm/transformers/replit.go

@@ -0,0 +1,42 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type Replit struct {
+	base.SingleThread
+
+	replit *transformers.Replit
+}
+
+func (llm *Replit) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewReplit(opts.ModelFile)
+	llm.replit = model
+	return err
+}
+
+func (llm *Replit) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.replit.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *Replit) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.replit.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+	return nil
+}

backend/go/llm/transformers/starcoder.go

@@ -0,0 +1,43 @@
+package transformers
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	transformers "github.com/go-skynet/go-ggml-transformers.cpp"
+)
+
+type Starcoder struct {
+	base.SingleThread
+
+	starcoder *transformers.Starcoder
+}
+
+func (llm *Starcoder) Load(opts *pb.ModelOptions) error {
+	model, err := transformers.NewStarcoder(opts.ModelFile)
+	llm.starcoder = model
+	return err
+}
+
+func (llm *Starcoder) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.starcoder.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+// fallback to Predict
+func (llm *Starcoder) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	go func() {
+		res, err := llm.starcoder.Predict(opts.Prompt, buildPredictOptions(opts)...)
+
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		results <- res
+		close(results)
+	}()
+
+	return nil
+}