Merge pull request #9741 from ollama/mxyng/visionless

fix: error if image requested without vision model

model/model.go

@@ -22,6 +22,8 @@ import (
 	"github.com/ollama/ollama/model/input"
 )
 
+var ErrNoVisionModel = errors.New("this model is missing data required for image input")
+
 // Model implements a specific model architecture, defining the forward pass and any model-specific configuration
 type Model interface {
 	Forward(ml.Context, input.Options) (ml.Tensor, error)

model/models/gemma3/model.go

@@ -84,6 +84,10 @@ func New(c ml.Config) (model.Model, error) {
 }
 
 func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+	if len(m.VisionModel.Layers) == 0 {
+		return nil, model.ErrNoVisionModel
+	}
+
 	image, _, err := image.Decode(bytes.NewReader(multimodalData))
 	if err != nil {
 		return nil, err

model/models/mllama/model.go

@@ -63,6 +63,10 @@ func New(c ml.Config) (model.Model, error) {
 }
 
 func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+	if len(m.VisionModel.Transformer.Layers) == 0 || len(m.GlobalTransformer.Layers) == 0 {
+		return nil, model.ErrNoVisionModel
+	}
+
 	image, _, err := image.Decode(bytes.NewReader(multimodalData))
 	if err != nil {
 		return nil, err