matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-09 23:37:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

ollamarunner: Use a separate context per multimodal input

Browse Source 
Currently there is a single context per sequence, shared all by
all multimodal inputs. Since we build a vision encoder graph per
image, with a large number of inputs we can eventually hit the
maximum number of graph nodes per context.

This changes to use a separate context for each image, ensuring
that available resource limits are consistent.
This commit is contained in:
Jesse Gross
2025-03-13 20:32:50 -07:00
committed by Jesse Gross
parent 9679f40146
commit 282bfaaa95
4 changed files with 33 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
model/model.go
View File

@@ -60,7 +60,7 @@ 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) ([]input.Input, error)
PostTokenize([]input.Input) ([]input.Input, error)
}
// Base implements the common fields and methods for all models

2
model/models/gemma3/model.go
View File

@@ -111,7 +111,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
return visionOutputs, nil
}
func (m *Model) PostTokenize(ctx ml.Context, inputs []input.Input) ([]input.Input, error) {
func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
var result []input.Input
for _, inp := range inputs {

11
model/models/mllama/model.go
View File

@@ -106,17 +106,17 @@ 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 []input.Input) ([]input.Input, error) {
func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
var images []input.Input
fnvHash := fnv.New64a()
for i := range inputs {
if inputs[i].Multimodal == nil {
if len(images) > 0 {
inputs[i].Multimodal = images[0].Multimodal
inputs[i].Multimodal = []ml.Tensor{images[0].Multimodal.(ml.Tensor)}
inputs[i].MultimodalHash = images[0].MultimodalHash
for j := 1; j < len(images); j++ {
inputs[i].Multimodal = inputs[i].Multimodal.(ml.Tensor).Concat(ctx, images[j].Multimodal.(ml.Tensor), 3)
inputs[i].Multimodal = append(inputs[i].Multimodal.([]ml.Tensor), images[0].Multimodal.(ml.Tensor))
fnvHash.Reset()
binary.Write(fnvHash, binary.NativeEndian, inputs[i].MultimodalHash)
binary.Write(fnvHash, binary.NativeEndian, inputs[j].MultimodalHash)
@@ -138,7 +138,10 @@ func (m *Model) PostTokenize(ctx ml.Context, inputs []input.Input) ([]input.Inpu
func (m *Model) Forward(ctx ml.Context, opts input.Options) (ml.Tensor, error) {
var crossAttentionStates ml.Tensor
if len(opts.Multimodal) > 0 {
crossAttentionStates = opts.Multimodal[len(opts.Multimodal)-1].Multimodal.(ml.Tensor)
images := opts.Multimodal[len(opts.Multimodal)-1].Multimodal.([]ml.Tensor)
if len(images) > 0 {
crossAttentionStates = images[len(images)-1]
}
}
inputs, err := ctx.Input().FromIntSlice(opts.Inputs, len(opts.Inputs))