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

39
model/models/qwen25vl/model.go
View File

@@ -5,7 +5,6 @@ import (
"fmt"
"image"
"slices"
"sync"
"github.com/ollama/ollama/fs"
"github.com/ollama/ollama/kvcache"
@@ -77,7 +76,7 @@ func (m *Model) PixelValues(ctx ml.Context, multimodalData []byte) (ml.Tensor, *
return pixelValues, grid, 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
}
@@ -88,31 +87,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
}
visionOutputs := m.VisionModel.Forward(ctx, pixels, grid)
return &chunks{Model: m, Tensor: visionOutputs}, nil
}
type chunks struct {
*Model
ml.Tensor
dataOnce sync.Once
data []float32
}
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)]
return []input.Multimodal{{Tensor: visionOutputs}}, nil
}
// PostTokenize arranges Qwen-2.5-VL's inputs for the forward pass
@@ -142,20 +117,16 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
result = append(result, input.Input{Token: pre[i]})
}
// This is an image token with multimodal data
chunksData := inp.Multimodal.(*chunks)
patchesPerChunk := chunksData.Dim(1)
patchesPerChunk := inp.Multimodal[0].Tensor.Dim(1)
// First add the vision start token
result = append(result, input.Input{Token: visionStartToken, SameBatch: patchesPerChunk + 2})
result = append(result, input.Input{Token: visionStartToken, SameBatch: patchesPerChunk + 1})
// Add the image token with the multimodal tensor data at the first position
// Create a chunk with proper s and n values
result = append(result, input.Input{
Token: imageToken,
Multimodal: &chunk{chunks: chunksData, s: 0, n: patchesPerChunk},
Multimodal: inp.Multimodal,
MultimodalHash: inp.MultimodalHash,
SameBatch: patchesPerChunk,
})
// Add the placeholder tokens for the remaining positions (tokensPerGrid-1)