load projectors

llm/llama.go

@@ -325,7 +325,7 @@ func (w *StatusWriter) Write(b []byte) (int, error) {
 	return os.Stderr.Write(b)
 }
 
-func newLlama(model string, adapters []string, runners []ModelRunner, numLayers int64, opts api.Options) (*llama, error) {
+func newLlama(model string, adapters, projectors []string, runners []ModelRunner, numLayers int64, opts api.Options) (*llama, error) {
 	fileInfo, err := os.Stat(model)
 	if err != nil {
 		return nil, err
@@ -365,6 +365,11 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
 		params = append(params, "--lora", adapters[0])
 	}
 
+	if len(projectors) > 0 {
+		// TODO: applying multiple projectors is not supported by the llama.cpp server yet
+		params = append(params, "--mmproj", projectors[0])
+	}
+
 	if opts.NumThread > 0 {
 		params = append(params, "--threads", fmt.Sprintf("%d", opts.NumThread))
 	}

llm/llm.go

@@ -23,7 +23,7 @@ type LLM interface {
 	Ping(context.Context) error
 }
 
-func New(workDir, model string, adapters []string, opts api.Options) (LLM, error) {
+func New(workDir, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -82,9 +82,9 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 		opts.NumGQA = 0
 		opts.RopeFrequencyBase = 0.0
 		opts.RopeFrequencyScale = 0.0
-		return newLlama(model, adapters, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
+		return newLlama(model, adapters, projectors, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
 	case "ggml", "ggmf", "ggjt", "ggla":
-		return newLlama(model, adapters, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)
+		return newLlama(model, adapters, projectors, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)
 	default:
 		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
 	}

server/images.go

@@ -35,16 +35,17 @@ type RegistryOptions struct {
 }
 
 type Model struct {
-	Name          string `json:"name"`
-	ShortName     string
-	ModelPath     string
-	OriginalModel string
-	AdapterPaths  []string
-	Template      string
-	System        string
-	License       []string
-	Digest        string
-	Options       map[string]interface{}
+	Name           string `json:"name"`
+	ShortName      string
+	ModelPath      string
+	OriginalModel  string
+	AdapterPaths   []string
+	ProjectorPaths []string
+	Template       string
+	System         string
+	License        []string
+	Digest         string
+	Options        map[string]interface{}
 }
 
 type PromptVars struct {
@@ -250,6 +251,8 @@ func GetModel(name string) (*Model, error) {
 			log.Print("WARNING: model contains embeddings, but embeddings in modelfiles have been deprecated and will be ignored.")
 		case "application/vnd.ollama.image.adapter":
 			model.AdapterPaths = append(model.AdapterPaths, filename)
+		case "application/vnd.ollama.image.projector":
+			model.ProjectorPaths = append(model.ProjectorPaths, filename)
 		case "application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
 			if err != nil {

server/routes.go

@@ -105,7 +105,7 @@ func load(c *gin.Context, modelName string, reqOpts map[string]interface{}, sess
 			loaded.Options = nil
 		}
 
-		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, opts)
+		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, model.ProjectorPaths, opts)
 		if err != nil {
 			// some older models are not compatible with newer versions of llama.cpp
 			// show a generalized compatibility error until there is a better way to