highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-08-31 01:01:17 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix pixel values padding (#10718)

Browse Source 
* panic if trying to pad 4d

* fix pixel values padding
This commit is contained in:
Michael Yang
2025-05-15 13:44:44 -07:00
committed by GitHub
parent 55760195e6
commit ef202789fa
2 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ml/backend/ggml/ggml.go
View File

@@ -915,6 +915,8 @@ func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
if len(shape) != 4 {
panic("expected 4 dimensions")
} else if shape[3] != 0 {
panic("cuda does not support 4d tensors")
}
return &Tensor{

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

@@ -3,6 +3,7 @@ package mllama
import (
"bytes"
"image"
"slices"
"github.com/ollama/ollama/fs"
"github.com/ollama/ollama/kvcache"
@@ -73,13 +74,17 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
return nil, err
}
pixelValues, err := ctx.Input().FromFloatSlice(f32s, m.imageSize, m.imageSize, m.numChannels, ratio.numTiles())
if ratio.numTiles() < m.maxNumTiles {
// Pad tiles to maxNumTiles
f32s = slices.Grow(f32s, m.imageSize*m.imageSize*m.numChannels*m.maxNumTiles)
f32s = f32s[:m.imageSize*m.imageSize*m.numChannels*m.maxNumTiles]
}
pixelValues, err := ctx.Input().FromFloatSlice(f32s, m.imageSize, m.imageSize, m.numChannels, m.maxNumTiles)
if err != nil {
return nil, err
}
pixelValues = pixelValues.Pad(ctx, 0, 0, 0, m.ImageProcessor.maxNumTiles-ratio.numTiles())
aspectRatio, err := ctx.Input().FromIntSlice([]int32{int32(ratio.rank)}, 1)
if err != nil {
return nil, err