matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 15:57:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

ollamarunner: Separate text and multimodal graphs

Browse Source 
For some multimodal models (such as gemma3), we create a single
graph that generates the image embedding and then use this in the
text model. The embedding tensor is completely opaque to the runner.

However, this doesn't work if we need to use the embedding in multiple
batches. This can arise if the embedding is larger than the batch size.
In these cases (as with llama4), we would like to create views that
are more appropriately sized. However, if we do this then the original
source tensor is used in multiple graphs, which isn't allowed. To
avoid that problem, models with this pattern compute the embedding
tensor on first use and recreate the individual views. There is no
longer a single vision and text graph.

This codifies the pattern of separating vision and text graphs. The
logic of computing tensors on demand is moved to the runner, so models
no longer have to worry about this. It also gives the runner visibility
into the multimodal tensors, which is important for memory management.
This commit is contained in:
Jesse Gross
2025-05-05 13:32:11 -07:00
committed by Jesse Gross
parent 499ae7311f
commit 3c14461d5d
14 changed files with 241 additions and 186 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
model/models/gemma3/model.go
View File

@@ -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{Token: 108, SameBatch: inputMultimodal.Dim(1) + 3}, // "\n\n"
input.Input{Token: 255999}, // "<start_of_image>""
input.Input{Multimodal: []input.Multimodal{{Tensor: inputMultimodal}}, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
)
// add image token placeholders

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

@@ -165,7 +165,7 @@ 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) {