ollamarunner: Multi-modal worst case graph

We currently preallocate compute graph memory for the worst case
batch of text tokens. This adds support for doing the same for
images.

Note that image models are more complicated than text models in
how they process their inputs so there may be cases where this
approach isn't completely generic for all models. It covers all
currently supported models though.

model/input/input.go

@@ -2,16 +2,30 @@ package input
 
 import "github.com/ollama/ollama/ml"
 
+// Multimodal is a multimodal embedding or a component of one.
+// For example, it could be a row of an image that can be processed
+// independently.
+type Multimodal struct {
+	// Tensor is the embedding data. Implementations may chose what to
+	// store here or it may be nil if not needed. However, any ml.Tensor
+	// objects must be stored here and not in Data.
+	Tensor ml.Tensor
+
+	// Data is implementation-specific opaque data, such as metadata on how
+	// to layout Tensor. It may be nil if not needed. It may also store larger
+	// objects such as complete images if they are to be processed later.
+	Data any
+}
+
 // Input represents one token in the input stream
 type Input struct {
 	// Token is a single element of text.
 	Token int32
 
-	// Multimodal is opaque data representing a non-text
-	// element such as an image (or part of one if the image
-	// can be processed in pieces). It may be either together
-	// with Token or on its own.
-	Multimodal any
+	// Multimodal is represents a non-text element such as an
+	// image (or part of one if the image can be processed in pieces).
+	// It may be used either together with Token or on its own.
+	Multimodal []Multimodal
 
 	// MultimodalHash is a unique representation of the data
 	// stored in Multimodal, used for caching and comparing
@@ -32,7 +46,7 @@ type Input struct {
 // Positions slice.
 type MultimodalIndex struct {
 	Index      int
-	Multimodal any
+	Multimodal []Multimodal
 }
 
 // Batch contains the inputs for a model forward pass

model/model.go

@@ -40,12 +40,13 @@ type MultimodalProcessor interface {
 	// EncodeMultimodal processes a single input (such as an image) and
 	// generates an output (typically an embedding) that can be used by the model.
 	//
-	// The return value is most typically an ml.Tensor, however, different
-	// type are possible, such as an object containing a tensor plus
-	// additional metadata, a slice of tensors or even just the original input.
+	// The return value is one or more tensors, each with optional model-specific
+	// opaque metadata. Typically, the tensors might be views into an embedding
+	// with each view representing a chunk of data that can be processed independently
+	// in different batches.
 	//
 	// The result may be cached by the runner.
-	EncodeMultimodal(ml.Context, []byte) (any, error)
+	EncodeMultimodal(ml.Context, []byte) ([]input.Multimodal, error)
 
 	// PostTokenize is called after tokenization to allow the model to edit the
 	// input stream to correctly arrange multimodal elements.

model/models/gemma3/model.go

@@ -82,7 +82,7 @@ func New(c fs.Config) (model.Model, error) {
 	return &m, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -108,22 +108,22 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 
 	visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
 	visionOutputs = m.MultiModalProjector.Forward(ctx, visionOutputs, m.imageSize, m.patchSize, m.VisionModel.eps)
-	return visionOutputs, nil
+	return []input.Multimodal{{Tensor: visionOutputs}}, nil
 }
 
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	var result []input.Input
 
 	for _, inp := range inputs {
-		if inp.Multimodal == nil {
+		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 		} else {
-			inputMultimodal := inp.Multimodal.(ml.Tensor)
+			inputMultimodal := inp.Multimodal[0].Tensor
 
 			result = append(result,
 				input.Input{Token: 108, SameBatch: inputMultimodal.Dim(1) + 3}, // "\n\n"
 				input.Input{Token: 255999},                                     // "<start_of_image>""
-				input.Input{Multimodal: inputMultimodal, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
+				input.Input{Multimodal: []input.Multimodal{{Tensor: inputMultimodal}}, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
 			)
 
 			// add image token placeholders

model/models/gemma3/model_text.go

@@ -178,7 +178,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	// set image embeddings
 	var except []int
 	for _, image := range batch.Multimodal {
-		visionOutputs := image.Multimodal.(ml.Tensor)
+		visionOutputs := image.Multimodal[0].Tensor
 		ctx.Forward(visionOutputs.Copy(ctx, hiddenState.View(ctx, image.Index*hiddenState.Stride(1), visionOutputs.Dim(0)*visionOutputs.Dim(1))))
 
 		for i := range visionOutputs.Dim(1) {

model/models/llama4/model.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"image"
 	"slices"
-	"sync"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -60,7 +59,7 @@ func New(c fs.Config) (model.Model, error) {
 	return &m, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) < 1 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -100,69 +99,78 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 	visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
 	visionOutputs = visionOutputs.Reshape(ctx, visionOutputs.Dim(0), visionOutputs.Dim(1)*visionOutputs.Dim(2)*visionOutputs.Dim(3))
 	projectedOutputs := m.Projector.Forward(ctx, visionOutputs)
-	return &chunks{Model: m, Tensor: projectedOutputs, aspectRatio: image.Point{ratioW, ratioH}}, nil
+
+	var multimodal []input.Multimodal
+	aspectRatio := image.Point{ratioW, ratioH}
+
+	var offset int
+	patchesPerChunk := projectedOutputs.Dim(1)
+	if aspectRatio.Y*aspectRatio.X > 1 {
+		patchesPerChunk = projectedOutputs.Dim(1) / (aspectRatio.X*aspectRatio.Y + 1)
+
+		for range aspectRatio.Y {
+			for x := range aspectRatio.X {
+				view := projectedOutputs.View(ctx, projectedOutputs.Stride(1)*offset,
+					projectedOutputs.Dim(0), projectedOutputs.Stride(1),
+					patchesPerChunk)
+				var separator separator
+				if x < aspectRatio.X-1 {
+					separator.x = true // <|tile_x_separator|>
+				} else {
+					separator.y = true // <|tile_y_separator|>
+				}
+				multimodal = append(multimodal, input.Multimodal{Tensor: view, Data: &separator})
+				offset += patchesPerChunk
+			}
+		}
 	}
 
-type chunks struct {
-	*Model
-	ml.Tensor
-	aspectRatio image.Point
+	view := projectedOutputs.View(ctx, projectedOutputs.Stride(1)*offset,
+		projectedOutputs.Dim(0), projectedOutputs.Stride(1),
+		patchesPerChunk)
+	multimodal = append(multimodal, input.Multimodal{Tensor: view, Data: &separator{}})
 
-	dataOnce sync.Once
-	data     []float32
+	return multimodal, nil
 }
 
-type chunk struct {
-	*chunks
-	s, n int
-}
-
-func (r *chunk) floats() []float32 {
-	r.dataOnce.Do(func() {
-		temp := r.Backend().NewContext()
-		defer temp.Close()
-		temp.Forward(r.Tensor).Compute(r.Tensor)
-		r.data = r.Floats()
-	})
-
-	return r.data[r.s*r.Dim(0) : (r.s+r.n)*r.Dim(0)]
+type separator struct {
+	x bool
+	y bool
 }
 
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	var result []input.Input
 	for _, inp := range inputs {
-		if inp.Multimodal == nil {
+		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 			continue
 		}
 
-		t := inp.Multimodal.(*chunks)
 		var imageInputs []input.Input
 		imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_start|>
 
-		var offset int
-		patchesPerChunk := t.Dim(1)
-		if t.aspectRatio.Y*t.aspectRatio.X > 1 {
-			patchesPerChunk = t.Dim(1) / (t.aspectRatio.X*t.aspectRatio.Y + 1)
+		for i, mm := range inp.Multimodal {
+			patchesPerChunk := mm.Tensor.Dim(1)
 
-			for range t.aspectRatio.Y {
-				for x := range t.aspectRatio.X {
-					imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: &chunk{t, offset, patchesPerChunk}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
+			if i < len(inp.Multimodal)-1 {
+				separator := mm.Data.(*separator)
+
+				imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
 				imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
-					if x < t.aspectRatio.X-1 {
+
+				if separator.x {
 					imageInputs = append(imageInputs, input.Input{Token: 200084}) // <|tile_x_separator|>
 				}
-					offset += patchesPerChunk
-				}
-
+				if separator.y {
 					imageInputs = append(imageInputs, input.Input{Token: 200085}) // <|tile_y_separator|>
 				}
-		}
-
+			} else {
 				imageInputs = append(imageInputs, input.Input{Token: 200090})                                                                                                                      // <|image|>
-		imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: &chunk{t, offset, patchesPerChunk}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
+				imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
 				imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
 				imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_end|>
+			}
+		}
 
 		result = append(result, imageInputs...)
 	}

model/models/llama4/model_text.go

@@ -210,12 +210,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	hiddenStates := m.TokenEmbedding.Forward(ctx, inputs).Duplicate(ctx)
 
 	for _, mi := range batch.Multimodal {
-		f32s := mi.Multimodal.(*chunk).floats()
-		img, err := ctx.Input().FromFloatSlice(f32s, len(f32s)/m.hiddenSize, m.hiddenSize)
-		if err != nil {
-			panic(err)
-		}
-
+		img := mi.Multimodal[0].Tensor
 		ctx.Forward(img.Copy(ctx, hiddenStates.View(ctx, mi.Index*hiddenStates.Stride(1), img.Dim(0)*img.Dim(1))))
 	}
 

model/models/mistral3/model.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"image"
 	"slices"
-	"sync"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -88,7 +87,7 @@ func newMultiModalProjector(c fs.Config) *MultiModalProjector {
 	}
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -112,37 +111,14 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 	features, size := m.MultiModalProjector.Forward(ctx, visionOutputs, size)
 
 	// split into patches to be sent to the text transformer
-	parent := imageFeatures{tensor: features}
-	rows := make([]*imageRow, size.Y)
+	rows := make([]input.Multimodal, size.Y)
 	for i := range rows {
-		rows[i] = &imageRow{parent: &parent, s: i, shape: []int{features.Dim(0), size.X}}
+		rows[i].Tensor = features.View(ctx, features.Stride(1)*size.X*i, features.Dim(0), features.Stride(1), size.X)
 	}
 
 	return rows, nil
 }
 
-type imageFeatures struct {
-	tensor ml.Tensor
-
-	dataOnce sync.Once
-	data     []float32
-}
-
-type imageRow struct {
-	parent *imageFeatures
-	s      int
-	shape  []int
-}
-
-func (r *imageRow) data() []float32 {
-	n := 1
-	for _, s := range r.shape {
-		n *= s
-	}
-
-	return r.parent.data[r.s*n : (r.s+1)*n]
-}
-
 // PostTokenize arranges Mistral 3's inputs for the forward pass
 // In Mistral 3 and Pixtral, the input patches are arranged as follows:
 // [IMG]...[IMG][IMG_BREAK][IMG]...[IMG][IMG_BREAK][IMG]...[IMG][IMG_END]
@@ -151,15 +127,14 @@ func (r *imageRow) data() []float32 {
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	var result []input.Input
 	for _, inp := range inputs {
-		if inp.Multimodal == nil {
+		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 		} else {
-			inputMultimodal := inp.Multimodal.([]*imageRow)
-			for i, row := range inputMultimodal {
+			for i, row := range inp.Multimodal {
 				// [IMG]
-				result = append(result, input.Input{Token: 10, Multimodal: row, MultimodalHash: inp.MultimodalHash, SameBatch: row.shape[1]})
-				result = append(result, slices.Repeat([]input.Input{{Token: 10}}, row.shape[1]-1)...)
-				if i == len(inputMultimodal)-1 {
+				result = append(result, input.Input{Token: 10, Multimodal: []input.Multimodal{{Tensor: row.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: row.Tensor.Dim(1)})
+				result = append(result, slices.Repeat([]input.Input{{Token: 10}}, row.Tensor.Dim(1)-1)...)
+				if i == len(inp.Multimodal)-1 {
 					// [IMG_END]
 					result = append(result, input.Input{Token: 13})
 				} else {

model/models/mistral3/model_text.go

@@ -110,20 +110,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 
 	// image embeddings
 	for _, image := range batch.Multimodal {
-		row := image.Multimodal.(*imageRow)
-		row.parent.dataOnce.Do(func() {
-			// use a new, throwaway context so the image tensor is not added to the graph
-			temp := m.Backend().NewContext()
-			temp.Forward(row.parent.tensor).Compute(row.parent.tensor)
-			row.parent.data = row.parent.tensor.Floats()
-			temp.Close()
-		})
-
-		imageFeature, err := ctx.Input().FromFloatSlice(row.data(), row.shape...)
-		if err != nil {
-			panic(err)
-		}
-
+		imageFeature := image.Multimodal[0].Tensor
 		ctx.Forward(imageFeature.Copy(ctx, hiddenState.View(ctx, image.Index*hiddenState.Stride(1), imageFeature.Dim(0)*imageFeature.Dim(1))))
 	}
 

model/models/mllama/model.go

@@ -63,7 +63,7 @@ func New(c fs.Config) (model.Model, error) {
 	return &m, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Transformer.Layers) == 0 || len(m.GlobalTransformer.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -95,7 +95,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 
 	positionIDs := ctx.Arange(0, 1601, 1, ml.DTypeI32)
 	crossAttentionStates := m.VisionModel.Forward(ctx, pixelValues, positionIDs, aspectRatio)
-	return m.Projector.Forward(ctx, crossAttentionStates), nil
+	return []input.Multimodal{{Tensor: m.Projector.Forward(ctx, crossAttentionStates)}}, nil
 }
 
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
@@ -103,12 +103,12 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	fnvHash := fnv.New64a()
 
 	for i := range inputs {
-		if inputs[i].Multimodal == nil {
+		if len(inputs[i].Multimodal) == 0 {
 			if len(images) > 0 {
-				inputs[i].Multimodal = []ml.Tensor{images[0].Multimodal.(ml.Tensor)}
+				inputs[i].Multimodal = images[0].Multimodal
 				inputs[i].MultimodalHash = images[0].MultimodalHash
 				for j := 1; j < len(images); j++ {
-					inputs[i].Multimodal = append(inputs[i].Multimodal.([]ml.Tensor), images[0].Multimodal.(ml.Tensor))
+					inputs[i].Multimodal = append(inputs[i].Multimodal, images[j].Multimodal...)
 					fnvHash.Reset()
 					binary.Write(fnvHash, binary.NativeEndian, inputs[i].MultimodalHash)
 					binary.Write(fnvHash, binary.NativeEndian, inputs[j].MultimodalHash)
@@ -130,9 +130,9 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 	var crossAttentionStates ml.Tensor
 	if len(batch.Multimodal) > 0 {
-		images := batch.Multimodal[len(batch.Multimodal)-1].Multimodal.([]ml.Tensor)
+		images := batch.Multimodal[len(batch.Multimodal)-1].Multimodal
 		if len(images) > 0 {
-			crossAttentionStates = images[len(images)-1]
+			crossAttentionStates = images[len(images)-1].Tensor
 		}
 	}
 

runner/llamarunner/cache.go

@@ -104,8 +104,8 @@ func (c *InputCache) LoadCacheSlot(prompt []input, cachePrompt bool) (*InputCach
 	slog.Debug("loading cache slot", "id", slot.Id, "cache", len(slot.Inputs), "prompt", len(prompt),
 		"used", numPast, "remaining", len(prompt)-numPast)
 
+	slot.Inputs = prompt[:numPast]
 	prompt = prompt[numPast:]
-	slot.Inputs = slot.Inputs[:numPast]
 
 	return slot, prompt, nil
 }

runner/ollamarunner/cache.go

@@ -136,8 +136,8 @@ func (c *InputCache) LoadCacheSlot(prompt []input.Input) (*InputCacheSlot, []inp
 	slog.Debug("loading cache slot", "id", slot.Id, "cache", len(slot.Inputs), "prompt", len(prompt),
 		"used", numPast, "remaining", int32(len(prompt))-numPast)
 
+	slot.Inputs = prompt[:numPast]
 	prompt = prompt[numPast:]
-	slot.Inputs = slot.Inputs[:numPast]
 
 	return slot, prompt, nil
 }

runner/ollamarunner/cache_test.go

@@ -3,7 +3,6 @@ package ollamarunner
 import (
 	"errors"
 	"fmt"
-	"image"
 	"testing"
 	"time"
 
@@ -12,10 +11,6 @@ import (
 )
 
 func TestCountCommon(t *testing.T) {
-	imgA := image.NewRGBA(image.Rect(0, 0, 100, 100))
-	imgB := image.NewRGBA(image.Rect(0, 0, 50, 50))
-	imgC := image.NewRGBA(image.Rect(50, 50, 100, 100))
-
 	tests := []struct {
 		name     string
 		t1       []input.Input
@@ -36,20 +31,20 @@ func TestCountCommon(t *testing.T) {
 		},
 		{
 			name:     "Image Prefix",
-			t1:       []input.Input{{Multimodal: imgA, MultimodalHash: 1}},
-			t2:       []input.Input{{Multimodal: imgA, MultimodalHash: 1}, {Multimodal: imgB, MultimodalHash: 2}, {Multimodal: imgC, MultimodalHash: 3}},
+			t1:       []input.Input{{MultimodalHash: 1}},
+			t2:       []input.Input{{MultimodalHash: 1}, {MultimodalHash: 2}, {MultimodalHash: 3}},
 			expected: 1,
 		},
 		{
 			name:     "Mixed",
-			t1:       []input.Input{{Token: 1}, {Multimodal: imgA, MultimodalHash: 1}},
-			t2:       []input.Input{{Token: 1}, {Multimodal: imgA, MultimodalHash: 1}, {Token: 5}},
+			t1:       []input.Input{{Token: 1}, {MultimodalHash: 1}},
+			t2:       []input.Input{{Token: 1}, {MultimodalHash: 1}, {Token: 5}},
 			expected: 2,
 		},
 		{
 			name:     "Mixed, Same Length",
-			t1:       []input.Input{{Token: 1}, {Multimodal: imgA, MultimodalHash: 1}},
-			t2:       []input.Input{{Token: 1}, {Multimodal: imgB, MultimodalHash: 2}},
+			t1:       []input.Input{{Token: 1}, {MultimodalHash: 1}},
+			t2:       []input.Input{{Token: 1}, {MultimodalHash: 2}},
 			expected: 1,
 		},
 		{

runner/ollamarunner/multimodal.go

@@ -0,0 +1,116 @@
+package ollamarunner
+
+import (
+	"errors"
+
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
+)
+
+// Tensors can't be used across multiple compute graphs. This is a problem
+// if a single embedding is split across batches using views since all of
+// the views will have the same source tensor. We also don't want to
+// recompute the entire embedding for each batch.
+//
+// To avoid this, we compute all of the tensors for the embedding on the
+// first use and then store the result in system memory. When we need
+// additional tensors, we recreate them from the stored data.
+
+// multimodalEntry represents the embeddings of a single object (such
+// as an image).
+type multimodalEntry struct {
+	// mm is the original set of tensors created by EncodeMultimodal
+	mm []input.Multimodal
+
+	// data is the computed result of mm. Nil if not yet computed
+	data [][]float32
+}
+
+// multimodalStore maps from an individual tensor (of which there
+// may be many in a single multimodal object) to its parent embedding
+type multimodalStore map[ml.Tensor]*multimodalEntry
+
+func newMultimodalStore() multimodalStore {
+	return make(multimodalStore)
+}
+
+// addMultimodal stores an embedding for later use in a compute graph
+func (m multimodalStore) addMultimodal(embedding []input.Multimodal) {
+	entry := &multimodalEntry{mm: embedding}
+
+	for _, e := range embedding {
+		if e.Tensor != nil {
+			m[e.Tensor] = entry
+		}
+	}
+}
+
+// getMultimodal takes a source set of tensors (which may contain a whole or
+// parts of one or more images) and returns the equivalent that can be used in
+// the current context
+func (m multimodalStore) getMultimodal(backend ml.Backend, ctx ml.Context, in []input.Multimodal, reserve bool) ([]input.Multimodal, error) {
+	out := make([]input.Multimodal, len(in))
+	for i := range out {
+		if in[i].Tensor != nil {
+			var err error
+			out[i].Tensor, err = m.getTensor(backend, ctx, in[i].Tensor, reserve)
+			if err != nil {
+				return nil, err
+			}
+		}
+
+		out[i].Data = in[i].Data
+	}
+
+	return out, nil
+}
+
+func (m multimodalStore) getTensor(backend ml.Backend, ctx ml.Context, in ml.Tensor, reserve bool) (ml.Tensor, error) {
+	entry := m[in]
+
+	if entry.data == nil {
+		computeCtx := backend.NewContext()
+		defer computeCtx.Close()
+
+		var tensors []ml.Tensor
+		for _, t := range entry.mm {
+			if t.Tensor != nil {
+				tensors = append(tensors, t.Tensor)
+			}
+		}
+
+		if len(tensors) == 0 {
+			return nil, nil
+		}
+
+		computeCtx.Forward(tensors...)
+		entry.data = make([][]float32, len(entry.mm))
+
+		if !reserve {
+			computeCtx.Compute(tensors...)
+
+			for i, t := range entry.mm {
+				if t.Tensor != nil {
+					entry.data[i] = t.Tensor.Floats()
+				}
+			}
+		} else {
+			err := computeCtx.Reserve()
+			if err != nil {
+				return nil, err
+			}
+		}
+	}
+
+	for i, t := range entry.mm {
+		if in == t.Tensor {
+			if !reserve {
+				return ctx.Input().FromFloatSlice(entry.data[i], t.Tensor.Shape()...)
+			} else {
+				return ctx.Input().Empty(t.Tensor.DType(), t.Tensor.Shape()...), nil
+			}
+		}
+	}
+
+	return nil, errors.New("multimodal tensor not found")
+}

runner/ollamarunner/runner.go

@@ -1,12 +1,14 @@
 package ollamarunner
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
 	"flag"
 	"fmt"
 	"hash/maphash"
+	"image"
 	"log"
 	"log/slog"
 	"net"
@@ -20,6 +22,7 @@ import (
 	"time"
 	"unicode/utf8"
 
+	"golang.org/x/image/bmp"
 	"golang.org/x/sync/semaphore"
 
 	"github.com/ollama/ollama/api"
@@ -40,6 +43,9 @@ type Sequence struct {
 	// multimodal embeddings
 	ctxs []ml.Context
 
+	// mmStore holds multimodal embeddings to mange memory and enable splitting across batches
+	mmStore multimodalStore
+
 	// batch index
 	iBatch int
 
@@ -101,7 +107,7 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 
 	startTime := time.Now()
 
-	inputs, ctxs, err := s.inputs(prompt, images)
+	inputs, ctxs, mmStore, err := s.inputs(prompt, images)
 	if err != nil {
 		return nil, fmt.Errorf("failed to process inputs: %w", err)
 	} else if len(inputs) == 0 {
@@ -156,6 +162,7 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 
 	return &Sequence{
 		ctxs:                ctxs,
+		mmStore:             mmStore,
 		inputs:              inputs,
 		numPromptInputs:     len(inputs),
 		startProcessingTime: startTime,
@@ -174,9 +181,10 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 // inputs processes the prompt and images into a list of inputs
 // by splitting the prompt on [img-<n>] tags, tokenizing text and
 // decoding images
-func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, []ml.Context, error) {
+func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, []ml.Context, multimodalStore, error) {
 	var inputs []input.Input
 	var ctxs []ml.Context
+	var mmStore multimodalStore
 
 	var parts []string
 	var matches [][]string
@@ -187,6 +195,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 		re := regexp.MustCompile(`\[img-(\d+)\]`)
 		parts = re.Split(prompt, -1)
 		matches = re.FindAllStringSubmatch(prompt, -1)
+		mmStore = newMultimodalStore()
 	} else {
 		parts = []string{prompt}
 	}
@@ -196,7 +205,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 		// text - tokenize
 		tokens, err := s.model.(model.TextProcessor).Encode(part, i == 0)
 		if err != nil {
-			return nil, nil, err
+			return nil, nil, nil, err
 		}
 
 		for _, t := range tokens {
@@ -216,7 +225,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 			}
 
 			if imageIndex < 0 {
-				return nil, nil, fmt.Errorf("invalid image index: %d", n)
+				return nil, nil, nil, fmt.Errorf("invalid image index: %d", n)
 			}
 
 			ctx := s.model.Backend().NewContext()
@@ -224,13 +233,15 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 			ctxs = append(ctxs, ctx)
 			imageEmbeddings, err := multimodalProcessor.EncodeMultimodal(ctx, images[imageIndex].Data)
 			if err != nil {
-				return nil, nil, err
+				return nil, nil, nil, err
 			}
 
 			s.multimodalHash.Reset()
 			_, _ = s.multimodalHash.Write(images[imageIndex].Data)
 			imageHash := s.multimodalHash.Sum64()
 
+			mmStore.addMultimodal(imageEmbeddings)
+
 			inputs = append(inputs, input.Input{Multimodal: imageEmbeddings, MultimodalHash: imageHash})
 			postTokenize = true
 		}
@@ -240,11 +251,11 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 		var err error
 		inputs, err = multimodalProcessor.PostTokenize(inputs)
 		if err != nil {
-			return nil, nil, err
+			return nil, nil, nil, err
 		}
 	}
 
-	return inputs, ctxs, nil
+	return inputs, ctxs, mmStore, nil
 }
 
 type Server struct {
@@ -363,6 +374,9 @@ func (s *Server) processBatch() error {
 	}
 	defer s.mu.Unlock()
 
+	ctx := s.model.Backend().NewContext()
+	defer ctx.Close()
+
 	var batchInputs []int32
 	var batch input.Batch
 
@@ -433,7 +447,11 @@ func (s *Server) processBatch() error {
 
 			batchInputs = append(batchInputs, inp.Token)
 			if inp.Multimodal != nil {
-				batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: len(batchInputs) - 1, Multimodal: inp.Multimodal})
+				mm, err := seq.mmStore.getMultimodal(s.model.Backend(), ctx, inp.Multimodal, false)
+				if err != nil {
+					return err
+				}
+				batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: len(batchInputs) - 1, Multimodal: mm})
 			}
 
 			batch.Positions = append(batch.Positions, int32(len(seq.cache.Inputs)+len(seq.pendingInputs)))
@@ -459,9 +477,6 @@ func (s *Server) processBatch() error {
 		return nil
 	}
 
-	ctx := s.model.Backend().NewContext()
-	defer ctx.Close()
-
 	modelOutput, err := model.Forward(ctx, s.model, batchInputs, batch)
 	if err != nil {
 		return fmt.Errorf("failed to decode batch: %w", err)
@@ -720,12 +735,71 @@ func (s *Server) reserveWorstCaseGraph() error {
 	ctx := s.model.Backend().NewContext()
 	defer ctx.Close()
 
+	var err error
+	inputs := make([]input.Input, s.batchSize)
+	mmStore := newMultimodalStore()
+
+	// Multimodal strategy:
+	// - Encode a 2048x2048 image. This assumes that a single image of this
+	//   size is sufficient to trigger the worst case. This is currently true
+	//   because for existing models, only a single image fits in a batch.
+	// - Add the embedding to a full batch of tokens - this is necessary because
+	//   the model may be looking for non-image data, such as <image> tags.
+	// - Run PostTokenize to execute any transformations between generated
+	//   embeddings and what the forward pass expects.
+	// - The result may now be larger than a batch (images may not fit in a
+	//   single batch), so trim based on what will fit and must be grouped together.
+	// - Fill out the rest of the space with text tokens.
+	if multimodalProcessor, ok := s.model.(model.MultimodalProcessor); ok {
+		mmCtx := s.model.Backend().NewContext()
+		defer mmCtx.Close()
+
+		img := image.NewGray(image.Rect(0, 0, 2048, 2048))
+		var buf bytes.Buffer
+		bmp.Encode(&buf, img)
+
+		if inputs[0].Multimodal, err = multimodalProcessor.EncodeMultimodal(mmCtx, buf.Bytes()); err == nil {
+			mmStore.addMultimodal(inputs[0].Multimodal)
+
+			inputs, err = multimodalProcessor.PostTokenize(inputs)
+			if err != nil {
+				return err
+			}
+
+			for i, inp := range inputs {
+				minBatch := 1 + inp.SameBatch
+				if minBatch > s.batchSize {
+					inputs = inputs[i:min(i+minBatch, len(inputs))]
+					break
+				} else if i+minBatch > s.batchSize {
+					inputs = inputs[:i]
+					break
+				}
+			}
+
+			if len(inputs) < s.batchSize {
+				newInputs := make([]input.Input, s.batchSize)
+				copy(newInputs, inputs)
+				inputs = newInputs
+			}
+		}
+	}
+
 	var batch input.Batch
 
-	inputs := make([]int32, s.batchSize)
+	batchInputs := make([]int32, len(inputs))
 	batch.Positions = make([]int32, len(inputs))
 	batch.Sequences = make([]int, len(inputs))
-	for i := range inputs {
+	for i, inp := range inputs {
+		batchInputs[i] = inp.Token
+		if inp.Multimodal != nil {
+			mm, err := mmStore.getMultimodal(s.model.Backend(), ctx, inp.Multimodal, true)
+			if err != nil {
+				return err
+			}
+			batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: i, Multimodal: mm})
+		}
+
 		batch.Positions[i] = int32(i)
 	}
 
@@ -734,8 +808,7 @@ func (s *Server) reserveWorstCaseGraph() error {
 		batch.Outputs[i] = int32(i)
 	}
 
-	var err error
-	batch.Inputs, err = ctx.Input().FromIntSlice(inputs, len(inputs))
+	batch.Inputs, err = ctx.Input().FromIntSlice(batchInputs, len(batchInputs))
 	if err != nil {
 		return err
 	}