Enhancements (#34)

Signed-off-by: mudler <mudler@c3os.io>

pkg/model/loader.go

@@ -10,6 +10,7 @@ import (
 	"sync"
 	"text/template"
 
+	gptj "github.com/go-skynet/go-gpt4all-j.cpp"
 	llama "github.com/go-skynet/go-llama.cpp"
 )
 
@@ -17,11 +18,12 @@ type ModelLoader struct {
 	modelPath        string
 	mu               sync.Mutex
 	models           map[string]*llama.LLama
+	gptmodels        map[string]*gptj.GPTJ
 	promptsTemplates map[string]*template.Template
 }
 
 func NewModelLoader(modelPath string) *ModelLoader {
-	return &ModelLoader{modelPath: modelPath, models: make(map[string]*llama.LLama), promptsTemplates: make(map[string]*template.Template)}
+	return &ModelLoader{modelPath: modelPath, gptmodels: make(map[string]*gptj.GPTJ), models: make(map[string]*llama.LLama), promptsTemplates: make(map[string]*template.Template)}
 }
 
 func (ml *ModelLoader) ListModels() ([]string, error) {
@@ -62,16 +64,81 @@ func (ml *ModelLoader) TemplatePrefix(modelName string, in interface{}) (string,
 	return buf.String(), nil
 }
 
-func (ml *ModelLoader) LoadModel(modelName string, opts ...llama.ModelOption) (*llama.LLama, error) {
+func (ml *ModelLoader) loadTemplate(modelName, modelFile string) error {
+	modelTemplateFile := fmt.Sprintf("%s.tmpl", modelFile)
+
+	// Check if the model path exists
+	if _, err := os.Stat(modelTemplateFile); err != nil {
+		return nil
+	}
+
+	dat, err := os.ReadFile(modelTemplateFile)
+	if err != nil {
+		return err
+	}
+
+	// Parse the template
+	tmpl, err := template.New("prompt").Parse(string(dat))
+	if err != nil {
+		return err
+	}
+	ml.promptsTemplates[modelName] = tmpl
+
+	return nil
+}
+
+func (ml *ModelLoader) LoadGPTJModel(modelName string) (*gptj.GPTJ, error) {
 	ml.mu.Lock()
 	defer ml.mu.Unlock()
 
 	// Check if we already have a loaded model
 	modelFile := filepath.Join(ml.modelPath, modelName)
 
+	if m, ok := ml.gptmodels[modelFile]; ok {
+		return m, nil
+	}
+
+	// Check if the model path exists
+	if _, err := os.Stat(modelFile); os.IsNotExist(err) {
+		// try to find a s.bin
+		modelBin := fmt.Sprintf("%s.bin", modelFile)
+		if _, err := os.Stat(modelBin); os.IsNotExist(err) {
+			return nil, err
+		} else {
+			modelName = fmt.Sprintf("%s.bin", modelName)
+			modelFile = modelBin
+		}
+	}
+
+	// Load the model and keep it in memory for later use
+	model, err := gptj.New(modelFile)
+	if err != nil {
+		return nil, err
+	}
+
+	// If there is a prompt template, load it
+	if err := ml.loadTemplate(modelName, modelFile); err != nil {
+		return nil, err
+	}
+
+	ml.gptmodels[modelFile] = model
+	return model, err
+}
+
+func (ml *ModelLoader) LoadLLaMAModel(modelName string, opts ...llama.ModelOption) (*llama.LLama, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	modelFile := filepath.Join(ml.modelPath, modelName)
 	if m, ok := ml.models[modelFile]; ok {
 		return m, nil
 	}
+	// TODO: This needs refactoring, it's really bad to have it in here
+	// Check if we have a GPTJ model loaded instead
+	if _, ok := ml.gptmodels[modelFile]; ok {
+		return nil, fmt.Errorf("this model is a GPTJ one")
+	}
 
 	// Check if the model path exists
 	if _, err := os.Stat(modelFile); os.IsNotExist(err) {
@@ -92,21 +159,8 @@ func (ml *ModelLoader) LoadModel(modelName string, opts ...llama.ModelOption) (*
 	}
 
 	// If there is a prompt template, load it
-
-	modelTemplateFile := fmt.Sprintf("%s.tmpl", modelFile)
-	// Check if the model path exists
-	if _, err := os.Stat(modelTemplateFile); err == nil {
-		dat, err := os.ReadFile(modelTemplateFile)
-		if err != nil {
-			return nil, err
-		}
-
-		// Parse the template
-		tmpl, err := template.New("prompt").Parse(string(dat))
-		if err != nil {
-			return nil, err
-		}
-		ml.promptsTemplates[modelName] = tmpl
+	if err := ml.loadTemplate(modelName, modelFile); err != nil {
+		return nil, err
 	}
 
 	ml.models[modelFile] = model