model: Update encoder cache to use multimodal input processing handler

The encoder cache needs to know the position of images in the input
stream so that it knows when to delete them. Previously images didn't
have a position, so we implied one by breaking batches before an
image and then assuming the image was in the first position. However,
multimodal objects are now given explicit positions in the input
stream, so we can use that instead.

Breaking batches was also a way to simulate a cross attention mask
for mllama. However, given that it only supports a single sequence
and a single image, this mask doesn't serve any real purpose.
Removing the batch break does not appear to affect the quality of
the output.

Most of this is simply moving the input data structures to a new
package to avoid import cycles.

model/input/input.go

@@ -0,0 +1,37 @@
+package input
+
+// 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
+
+	// MultimodalHash is a unique representation of the data
+	// stored in Multimodal, used for caching and comparing
+	// equality.
+	MultimodalHash uint64
+}
+
+// MultimodalIndex is a multimodal element (such as an image)
+// together with an index into the slice of Inputs with the
+// corresponding token. Note that the index is not the same
+// as the position - to find that use the index with the
+// Positions slice.
+type MultimodalIndex struct {
+	Index      int
+	Multimodal any
+}
+
+// Options contains the inputs for a model forward pass
+type Options struct {
+	Inputs     []int32
+	Multimodal []MultimodalIndex
+	Positions  []int32
+	Sequences  []int
+	Outputs    []int32
+}

model/model.go

@@ -19,66 +19,12 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	_ "github.com/ollama/ollama/ml/backend"
+	"github.com/ollama/ollama/model/input"
 )
 
-// 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
-
-	// MultimodalHash is a unique representation of the data
-	// stored in Multimodal, used for caching and comparing
-	// equality.
-	MultimodalHash uint64
-}
-
-// MultimodalIndex is a multimodal element (such as an image)
-// together with an index into the slice of Inputs with the
-// corresponding token. Note that the index is not the same
-// as the position - to find that use the index with the
-// Positions slice.
-type MultimodalIndex struct {
-	Index      int
-	Multimodal any
-}
-
-// Options contains the inputs for a model forward pass
-type Options struct {
-	Inputs     []int32
-	Multimodal []MultimodalIndex
-	Positions  []int32
-	Sequences  []int
-	Outputs    []int32
-}
-
-type config struct {
-	Cache kvcache.Cache
-}
-
-// Base implements the common fields and methods for all models
-type Base struct {
-	b ml.Backend
-	config
-}
-
-// Backend returns the underlying backend that will run the model
-func (m *Base) Backend() ml.Backend {
-	return m.b
-}
-
-func (m *Base) Config() config {
-	return m.config
-}
-
 // Model implements a specific model architecture, defining the forward pass and any model-specific configuration
 type Model interface {
-	Forward(ml.Context, Options) (ml.Tensor, error)
+	Forward(ml.Context, input.Options) (ml.Tensor, error)
 
 	Backend() ml.Backend
 	Config() config
@@ -112,7 +58,26 @@ type MultimodalProcessor interface {
 	// This function is also responsible for updating MultimodalHash for any Multimodal
 	// that is modified to ensure that there is a unique hash value that accurately
 	// represents the contents.
-	PostTokenize(ml.Context, []Input) ([]Input, error)
+	PostTokenize(ml.Context, []input.Input) ([]input.Input, error)
+}
+
+// Base implements the common fields and methods for all models
+type Base struct {
+	b ml.Backend
+	config
+}
+
+type config struct {
+	Cache kvcache.Cache
+}
+
+// Backend returns the underlying backend that will run the model
+func (m *Base) Backend() ml.Backend {
+	return m.b
+}
+
+func (m *Base) Config() config {
+	return m.config
 }
 
 var models = make(map[string]func(ml.Config) (Model, error))
@@ -313,7 +278,7 @@ func canNil(t reflect.Type) bool {
 		t.Kind() == reflect.Slice
 }
 
-func Forward(ctx ml.Context, m Model, opts Options) (ml.Tensor, error) {
+func Forward(ctx ml.Context, m Model, opts input.Options) (ml.Tensor, error) {
 	if len(opts.Positions) != len(opts.Sequences) {
 		return nil, fmt.Errorf("length of positions (%v) must match length of seqs (%v)", len(opts.Positions), len(opts.Sequences))
 	}
@@ -324,7 +289,7 @@ func Forward(ctx ml.Context, m Model, opts Options) (ml.Tensor, error) {
 
 	cache := m.Config().Cache
 	if cache != nil {
-		err := cache.StartForward(ctx, opts.Positions, opts.Sequences)
+		err := cache.StartForward(ctx, opts)
 		if err != nil {
 			return nil, err
 		}

model/model_test.go

@@ -11,6 +11,7 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/backend/ggml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model/input"
 )
 
 func TestParseTags(t *testing.T) {
@@ -162,7 +163,7 @@ func TestGetTextProcessor(t *testing.T) {
 
 type notTextProcessorModel struct{}
 
-func (notTextProcessorModel) Forward(ml.Context, Options) (ml.Tensor, error) {
+func (notTextProcessorModel) Forward(ml.Context, input.Options) (ml.Tensor, error) {
 	panic("unimplemented")
 }
 

model/models/llama/model.go

@@ -9,6 +9,7 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
 )
 
 type Options struct {
@@ -137,7 +138,7 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Ten
 	return hiddenState.Add(ctx, residual)
 }
 
-func (m *Model) Forward(ctx ml.Context, opts model.Options) (ml.Tensor, error) {
+func (m *Model) Forward(ctx ml.Context, opts input.Options) (ml.Tensor, error) {
 	inputs, err := ctx.Input().FromIntSlice(opts.Inputs, len(opts.Inputs))
 	if err != nil {
 		return nil, err

model/models/mllama/model.go

@@ -12,6 +12,7 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
 )
 
 type Model struct {
@@ -101,8 +102,8 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 	return m.Projector.Forward(ctx, crossAttentionStates), nil
 }
 
-func (m *Model) PostTokenize(ctx ml.Context, inputs []model.Input) ([]model.Input, error) {
-	var images []model.Input
+func (m *Model) PostTokenize(ctx ml.Context, inputs []input.Input) ([]input.Input, error) {
+	var images []input.Input
 	fnvHash := fnv.New64a()
 
 	for i := range inputs {
@@ -125,15 +126,15 @@ func (m *Model) PostTokenize(ctx ml.Context, inputs []model.Input) ([]model.Inpu
 		}
 	}
 
-	inputs = slices.DeleteFunc(inputs, func(input model.Input) bool { return input.Token == -1 })
+	inputs = slices.DeleteFunc(inputs, func(input input.Input) bool { return input.Token == -1 })
 
 	return inputs, nil
 }
 
-func (m *Model) Forward(ctx ml.Context, opts model.Options) (ml.Tensor, error) {
+func (m *Model) Forward(ctx ml.Context, opts input.Options) (ml.Tensor, error) {
 	var crossAttentionStates ml.Tensor
-	if opts.Multimodal != nil {
-		crossAttentionStates = opts.Multimodal[0].Multimodal.(ml.Tensor)
+	if len(opts.Multimodal) > 0 {
+		crossAttentionStates = opts.Multimodal[len(opts.Multimodal)-1].Multimodal.(ml.Tensor)
 	}
 
 	inputs, err := ctx.Input().FromIntSlice(opts.Inputs, len(opts.Inputs))