highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-30 02:59:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

reduce kernel size, add TODO for loading from config

Browse Source
This commit is contained in:
jmorganca 2025-03-11 12:01:57 +01:00 committed by Michael Yang
parent 65b0f329d1
commit f63e62e546
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
model/models/gemma3/model.go
View File

@ -90,7 +90,11 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
patchesPerImage := m.ImageProcessor.imageSize / m.ImageProcessor.patchSize
kernelSize := patchesPerImage * patchesPerImage / 256
// TODO (jmorganca): read this from the model config
// it should instead be math.Sqrt(tokens per image)
tokensPerSide := 8
kernelSize := patchesPerImage / tokensPerSide
visionOutputs = visionOutputs.AvgPool1D(ctx, kernelSize, kernelSize, 0)
visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)