Skip to content

GitLab

Commit f63e62e5 authored by jmorganca's avatar jmorganca Committed by Michael Yang
Browse files

reduce kernel size, add TODO for loading from config

parent 65b0f329
Hide whitespace changes
Inline Side-by-side
Showing with 5 additions and 1 deletion
model/models/gemma3/model.go
View file @ f63e62e5
...@@ -90,7 +90,11 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er ...@@ -90,7 +90,11 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
visionOutputs := m.VisionModel.Forward(ctx, pixelValues) visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx) visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
patchesPerImage := m.ImageProcessor.imageSize / m.ImageProcessor.patchSize 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.AvgPool1D(ctx, kernelSize, kernelSize, 0)
visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx) visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment