OUTDATED: Service Wiring / Concurrency pt 1

backend/go/llm/rwkv/rwkv.go

@@ -31,7 +31,7 @@ func (llm *LLM) Load(opts *pb.ModelOptions) error {
 	model := rwkv.LoadFiles(opts.ModelFile, tokenizerPath, uint32(opts.GetThreads()))
 
 	if model == nil {
-		return fmt.Errorf("could not load model")
+		return fmt.Errorf("rwkv could not load model")
 	}
 	llm.rwkv = model
 	return nil

core/application.go

@@ -1,6 +1,7 @@
 package core
 
 import (
+	"github.com/go-skynet/LocalAI/core/backend"
 	"github.com/go-skynet/LocalAI/core/config"
 	"github.com/go-skynet/LocalAI/core/services"
 	"github.com/go-skynet/LocalAI/pkg/model"
@@ -17,20 +18,23 @@ type Application struct {
 	// Core Low-Level Services
 	BackendConfigLoader *config.BackendConfigLoader
 	ModelLoader         *model.ModelLoader
+	StoresLoader        *model.ModelLoader
 
 	// Backend Services
-	// EmbeddingsBackendService      *backend.EmbeddingsBackendService
+	EmbeddingsBackendService *backend.EmbeddingsBackendService
 	// ImageGenerationBackendService *backend.ImageGenerationBackendService
 	// LLMBackendService             *backend.LLMBackendService
 	// TranscriptionBackendService *backend.TranscriptionBackendService
-	// TextToSpeechBackendService  *backend.TextToSpeechBackendService
+	TextToSpeechBackendService *backend.TextToSpeechBackendService
+	// RerankBackendService       *backend.RerankBackendService
 
 	// LocalAI System Services
 	BackendMonitorService *services.BackendMonitorService
 	GalleryService        *services.GalleryService
 	ListModelsService     *services.ListModelsService
 	LocalAIMetricsService *services.LocalAIMetricsService
 	// OpenAIService         *services.OpenAIService
+
 }
 
 // TODO [NEXT PR?]: Break up ApplicationConfig.

core/backend/embeddings.go

@@ -2,34 +2,128 @@ package backend
 
 import (
 	"fmt"
+	"time"
 
 	"github.com/go-skynet/LocalAI/core/config"
+	"github.com/go-skynet/LocalAI/core/schema"
+	"github.com/google/uuid"
+	"github.com/rs/zerolog/log"
 
+	"github.com/go-skynet/LocalAI/pkg/concurrency"
 	"github.com/go-skynet/LocalAI/pkg/grpc"
-	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/model"
 )
 
-func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendConfig config.BackendConfig, appConfig *config.ApplicationConfig) (func() ([]float32, error), error) {
+type EmbeddingsBackendService struct {
+	ml        *model.ModelLoader
+	bcl       *config.BackendConfigLoader
+	appConfig *config.ApplicationConfig
+}
+
+func NewEmbeddingsBackendService(ml *model.ModelLoader, bcl *config.BackendConfigLoader, appConfig *config.ApplicationConfig) *EmbeddingsBackendService {
+	return &EmbeddingsBackendService{
+		ml:        ml,
+		bcl:       bcl,
+		appConfig: appConfig,
+	}
+}
+
+func (ebs *EmbeddingsBackendService) Embeddings(request *schema.OpenAIRequest) *concurrency.JobResult[*schema.OpenAIRequest, *schema.OpenAIResponse] {
+
+	jr, wjr := concurrency.NewJobResult[*schema.OpenAIRequest, *schema.OpenAIResponse](request)
+
+	go func(wjr *concurrency.WritableJobResult[*schema.OpenAIRequest, *schema.OpenAIResponse]) {
+		id := uuid.New().String()
+		created := int(time.Now().Unix())
+		request = *wjr.Request // TODO is needed?
+
+		bc, err := ebs.bcl.LoadBackendConfigFileByName(request.Model, ebs.appConfig.ModelPath,
+			config.LoadOptionDebug(ebs.appConfig.Debug),
+			config.LoadOptionThreads(ebs.appConfig.Threads),
+			config.LoadOptionContextSize(ebs.appConfig.ContextSize),
+			config.LoadOptionF16(ebs.appConfig.F16),
+		)
+		if err != nil {
+			log.Error().Err(err).Str("modelPath", ebs.appConfig.ModelPath).Msg("unable to load backend config")
+			wjr.SetResult(nil, err)
+			return
+		}
+
+		// Set the parameters for the language model prediction
+		bc.UpdateFromOpenAIRequest(request)
+
+		items := []schema.Item{}
+
+		for i, s := range bc.InputToken {
+			// get the model function to call for the result
+			embedFn, err := ebs.modelEmbedding("", s, *bc)
+			if err != nil {
+				log.Error().Err(err).Ints("numeric tokens", s).Msg("error during modelEmbedding")
+				wjr.SetResult(nil, err)
+				return
+			}
+
+			embeddings, err := embedFn()
+			if err != nil {
+				log.Error().Err(err).Ints("numeric tokens", s).Msg("error during embedFn")
+				wjr.SetResult(nil, err)
+				return
+			}
+			items = append(items, schema.Item{Embedding: embeddings, Index: i, Object: "embedding"})
+		}
+
+		for i, s := range bc.InputStrings {
+			// get the model function to call for the result
+			embedFn, err := ebs.modelEmbedding(s, []int{}, *bc)
+			if err != nil {
+				log.Error().Err(err).Str("string tokens", s).Msg("error during modelEmbedding")
+				wjr.SetResult(nil, err)
+				return
+			}
+
+			embeddings, err := embedFn()
+			if err != nil {
+				log.Error().Err(err).Str("string tokens", s).Msg("error during embedFn")
+				wjr.SetResult(nil, err)
+				return
+			}
+			items = append(items, schema.Item{Embedding: embeddings, Index: i, Object: "embedding"})
+		}
+
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Model:   request.Model, // we have to return what the user sent here, due to OpenAI spec.
+			Data:    items,
+			Object:  "list",
+		}
+		wjr.SetResult(resp, nil)
+	}(wjr)
+
+	return jr
+}
+
+func (ebs *EmbeddingsBackendService) modelEmbedding(s string, tokens []int, backendConfig config.BackendConfig) (func() ([]float32, error), error) {
 	modelFile := backendConfig.Model
 
 	grpcOpts := gRPCModelOpts(backendConfig)
 
 	var inferenceModel interface{}
 	var err error
 
-	opts := modelOpts(backendConfig, appConfig, []model.Option{
+	opts := modelOpts(backendConfig, ebs.appConfig, []model.Option{
 		model.WithLoadGRPCLoadModelOpts(grpcOpts),
 		model.WithThreads(uint32(*backendConfig.Threads)),
-		model.WithAssetDir(appConfig.AssetsDestination),
+		model.WithAssetDir(ebs.appConfig.AssetsDestination),
 		model.WithModel(modelFile),
-		model.WithContext(appConfig.Context),
+		model.WithContext(ebs.appConfig.Context),
 	})
 
 	if backendConfig.Backend == "" {
-		inferenceModel, err = loader.GreedyLoader(opts...)
+		inferenceModel, err = ebs.ml.GreedyLoader(opts...)
 	} else {
 		opts = append(opts, model.WithBackendString(backendConfig.Backend))
-		inferenceModel, err = loader.BackendLoader(opts...)
+		inferenceModel, err = ebs.ml.BackendLoader(opts...)
 	}
 	if err != nil {
 		return nil, err
@@ -39,7 +133,7 @@ func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendCo
 	switch model := inferenceModel.(type) {
 	case grpc.Backend:
 		fn = func() ([]float32, error) {
-			predictOptions := gRPCPredictOpts(backendConfig, loader.ModelPath)
+			predictOptions := gRPCPredictOpts(backendConfig, ebs.appConfig.ModelPath)
 			if len(tokens) > 0 {
 				embeds := []int32{}
 
@@ -48,7 +142,7 @@ func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendCo
 				}
 				predictOptions.EmbeddingTokens = embeds
 
-				res, err := model.Embeddings(appConfig.Context, predictOptions)
+				res, err := model.Embeddings(ebs.appConfig.Context, predictOptions)
 				if err != nil {
 					return nil, err
 				}
@@ -57,7 +151,7 @@ func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendCo
 			}
 			predictOptions.Embeddings = s
 
-			res, err := model.Embeddings(appConfig.Context, predictOptions)
+			res, err := model.Embeddings(ebs.appConfig.Context, predictOptions)
 			if err != nil {
 				return nil, err
 			}

core/backend/tts.go

@@ -7,67 +7,105 @@ import (
 	"path/filepath"
 
 	"github.com/go-skynet/LocalAI/core/config"
+	"github.com/go-skynet/LocalAI/core/schema"
+	"github.com/rs/zerolog/log"
 
+	"github.com/go-skynet/LocalAI/pkg/concurrency"
 	"github.com/go-skynet/LocalAI/pkg/grpc/proto"
-	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/model"
 	"github.com/go-skynet/LocalAI/pkg/utils"
 )
 
-func generateUniqueFileName(dir, baseName, ext string) string {
-	counter := 1
-	fileName := baseName + ext
+type TextToSpeechBackendService struct {
+	ml        *model.ModelLoader
+	bcl       *config.BackendConfigLoader
+	appConfig *config.ApplicationConfig
+}
 
-	for {
-		filePath := filepath.Join(dir, fileName)
-		_, err := os.Stat(filePath)
-		if os.IsNotExist(err) {
-			return fileName
+func NewTextToSpeechBackendService(ml *model.ModelLoader, bcl *config.BackendConfigLoader, appConfig *config.ApplicationConfig) *TextToSpeechBackendService {
+	return &TextToSpeechBackendService{
+		ml:        ml,
+		bcl:       bcl,
+		appConfig: appConfig,
+	}
+}
+
+func (ttsbs *TextToSpeechBackendService) TextToAudioFile(request *schema.TTSRequest) *concurrency.JobResult[*schema.TTSRequest, string] {
+	jr, wjr := concurrency.NewJobResult[*schema.TTSRequest, string](request)
+
+	go func(wjr *concurrency.WritableJobResult[*schema.TTSRequest, string]) {
+		if request.Model == "" {
+			wjr.SetResult("", fmt.Errorf("model is required, no default available"))
+			return
+		}
+		bc, err := ttsbs.bcl.LoadBackendConfigFileByName(request.Model, ttsbs.appConfig.ModelPath,
+			config.LoadOptionDebug(ttsbs.appConfig.Debug),
+			config.LoadOptionThreads(ttsbs.appConfig.Threads),
+			config.LoadOptionContextSize(ttsbs.appConfig.ContextSize),
+			config.LoadOptionF16(ttsbs.appConfig.F16),
+		)
+		if err != nil || bc == nil {
+			log.Error().Err(err).Str("modelName", request.Model).Str("modelPath", ttsbs.appConfig.ModelPath).Msg("unable to load backend config")
+			wjr.SetResult("", err)
+			return
 		}
 
-		counter++
-		fileName = fmt.Sprintf("%s_%d%s", baseName, counter, ext)
-	}
+		if request.Backend != "" { // Allow users to specify a backend to use that overrides config.
+			bc.Backend = request.Backend
+		}
+		// TODO consider merging the below function in, but leave it seperated for diff reasons in the first PR
+		dst, err := ttsbs.modelTTS(request.Backend, request.Input, bc.Model, request.Voice, request.Language, *bc)
+		log.Debug().Str("dst", dst).Err(err).Msg("modelTTS result in goroutine")
+		wjr.SetResult(dst, err)
+	}(wjr)
+
+	return jr
 }
 
-func ModelTTS(
-	backend,
-	text,
-	modelFile,
-	voice ,
+func (ttsbs *TextToSpeechBackendService) modelTTS(
+	backend string,
+	text string,
+	modelFile string,
+	voice string,
 	language string,
-	loader *model.ModelLoader,
-	appConfig *config.ApplicationConfig,
 	backendConfig config.BackendConfig,
-) (string, *proto.Result, error) {
+) (string, error) {
 	bb := backend
 	if bb == "" {
 		bb = model.PiperBackend
 	}
 
 	grpcOpts := gRPCModelOpts(backendConfig)
 
-	opts := modelOpts(config.BackendConfig{}, appConfig, []model.Option{
+	opts := modelOpts(config.BackendConfig{}, ttsbs.appConfig, []model.Option{
 		model.WithBackendString(bb),
 		model.WithModel(modelFile),
-		model.WithContext(appConfig.Context),
-		model.WithAssetDir(appConfig.AssetsDestination),
+		model.WithContext(ttsbs.appConfig.Context),
+		model.WithAssetDir(ttsbs.appConfig.AssetsDestination),
 		model.WithLoadGRPCLoadModelOpts(grpcOpts),
 	})
-	ttsModel, err := loader.BackendLoader(opts...)
+	ttsModel, err := ttsbs.ml.BackendLoader(opts...)
 	if err != nil {
-		return "", nil, err
+		return "", err
 	}
 
 	if ttsModel == nil {
-		return "", nil, fmt.Errorf("could not load piper model")
+		return "", fmt.Errorf("could not load piper model")
 	}
 
-	if err := os.MkdirAll(appConfig.AudioDir, 0750); err != nil {
-		return "", nil, fmt.Errorf("failed creating audio directory: %s", err)
+	if ttsbs.appConfig.AudioDir == "" {
+		return "", fmt.Errorf("ApplicationConfig.AudioDir not set, cannot continue")
 	}
 
-	fileName := generateUniqueFileName(appConfig.AudioDir, "tts", ".wav")
-	filePath := filepath.Join(appConfig.AudioDir, fileName)
+	// Shouldn't be needed anymore. Consider removing later
+	if err := os.MkdirAll(ttsbs.appConfig.AudioDir, 0750); err != nil {
+		return "", fmt.Errorf("failed` creating audio directory: %s", err)
+	}
+
+	fileName := generateUniqueFileName(ttsbs.appConfig.AudioDir, "tts", ".wav")
+	filePath := filepath.Join(ttsbs.appConfig.AudioDir, fileName)
+
+	log.Debug().Str("filePath", filePath).Msg("computed output filePath")
 
 	// If the model file is not empty, we pass it joined with the model path
 	modelPath := ""
@@ -76,10 +114,10 @@ func ModelTTS(
 		// Checking first that it exists and is not outside ModelPath
 		// TODO: we should actually first check if the modelFile is looking like
 		// a FS path
-		mp := filepath.Join(loader.ModelPath, modelFile)
+		mp := filepath.Join(ttsbs.appConfig.ModelPath, modelFile)
 		if _, err := os.Stat(mp); err == nil {
-			if err := utils.VerifyPath(mp, appConfig.ModelPath); err != nil {
-				return "", nil, err
+			if err := utils.VerifyPath(mp, ttsbs.appConfig.ModelPath); err != nil {
+				return "", err
 			}
 			modelPath = mp
 		} else {
@@ -88,17 +126,33 @@ func ModelTTS(
 	}
 
 	res, err := ttsModel.TTS(context.Background(), &proto.TTSRequest{
-		Text:  text,
-		Model: modelPath,
-		Voice: voice,
-		Dst:   filePath,
+		Text:     text,
+		Model:    modelPath,
+		Voice:    voice,
+		Dst:      filePath,
 		Language: &language,
 	})
 
 	// return RPC error if any
 	if !res.Success {
-		return "", nil, fmt.Errorf(res.Message)
+		return "", fmt.Errorf(res.Message)
 	}
 
-	return filePath, res, err
+	return filePath, err
+}
+
+func generateUniqueFileName(dir, baseName, ext string) string {
+	counter := 1
+	fileName := baseName + ext
+
+	for {
+		filePath := filepath.Join(dir, fileName)
+		_, err := os.Stat(filePath)
+		if os.IsNotExist(err) {
+			return fileName
+		}
+
+		counter++
+		fileName = fmt.Sprintf("%s_%d%s", baseName, counter, ext)
+	}
 }