matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 07:46:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

ml: Panic rather than return error on tensor allocation failure

Browse Source 
FromFloatSlice and FromIntSlice return an error if the shape doesn't
match the passed data or if memory can't be allocated. Since these
are inputs, the memory being allocated is system memory rather than VRAM.

In many cases, the caller can't really handle the error and panics.

Empty and Zeros directly panic if they can't allocate memory.

This makes things consistent by panicing for the first two cases,
removing a fair amount of error handling code. This is also consistent
with how Go typically handles these situations.
This commit is contained in:
Jesse Gross
2025-05-19 10:43:56 -07:00
committed by Jesse Gross
parent 73d6a82cce
commit 1f371ea92f
20 changed files with 68 additions and 209 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
model/models/gemma2/model.go
View File

@@ -175,15 +175,8 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Ten
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
hiddenState = hiddenState.Scale(ctx, math.Sqrt(float64(m.Options.hiddenSize)))

16
model/models/gemma3/model.go
View File

@@ -101,14 +101,11 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
return nil, err
}
pixelValues, err := ctx.Input().FromFloatSlice(f32s,
pixelValues := ctx.Input().FromFloatSlice(f32s,
m.ImageProcessor.imageSize,
m.ImageProcessor.imageSize,
m.ImageProcessor.numChannels,
)
if err != nil {
return nil, err
}
visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
visionOutputs = m.MultiModalProjector.Forward(ctx, visionOutputs, m.imageSize, m.patchSize, m.VisionModel.eps)
@@ -144,15 +141,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
}

10
model/models/llama/model.go
View File

@@ -142,10 +142,7 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tenso
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
@@ -154,10 +151,7 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
var outputs ml.Tensor
if i == len(m.Layers)-1 {
outputs, err = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
outputs = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
}
hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, m.Options)

22
model/models/llama4/model.go
View File

@@ -77,10 +77,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
return nil, err
}
tilesLocal, err := ctx.Input().FromFloatSlice(pixelsLocal, size.X, size.Y, m.numChannels)
if err != nil {
return nil, err
}
tilesLocal := ctx.Input().FromFloatSlice(pixelsLocal, size.X, size.Y, m.numChannels)
ratioW, ratioH := size.X/m.imageSize, size.Y/m.imageSize
@@ -91,11 +88,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
pixelValues := tilesLocal
if len(pixelsGlobal) > 0 {
tilesGlobal, err := ctx.Input().FromFloatSlice(pixelsGlobal, m.imageSize, m.imageSize, m.numChannels)
if err != nil {
return nil, err
}
tilesGlobal := ctx.Input().FromFloatSlice(pixelsGlobal, m.imageSize, m.imageSize, m.numChannels)
pixelValues = pixelValues.Concat(ctx, tilesGlobal, 3)
}
@@ -182,15 +175,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
}

6
model/models/llama4/model_text.go
View File

@@ -223,11 +223,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
scales[i] = float32(math.Log(math.Floor(((float64(p)+1.0)/float64(m.attentionFloorScale))+1.0))*m.attentionScale + 1.0)
}
var err error
attentionScales, err = ctx.Input().FromFloatSlice(scales, 1, 1, len(scales))
if err != nil {
panic(err)
}
attentionScales = ctx.Input().FromFloatSlice(scales, 1, 1, len(scales))
}
for i, layer := range m.Layers {

5
model/models/llama4/model_vision.go
View File

@@ -245,10 +245,7 @@ func (m *VisionModel) rotaryEmbedding(ctx ml.Context) (ml.Tensor, ml.Tensor) {
}
}
ropeFreqs, err := ctx.Input().FromFloatSlice(freqs, freqDim/2, numPatches, 2)
if err != nil {
panic(err)
}
ropeFreqs := ctx.Input().FromFloatSlice(freqs, freqDim/2, numPatches, 2)
ropeFreqs = ropeFreqs.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
ropeFreqs = ropeFreqs.Reshape(ctx, freqDim, 1, numPatches)

16
model/models/mistral3/model.go
View File

@@ -114,10 +114,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
return nil, err
}
pixelValues, err := ctx.Input().FromFloatSlice(f32s, size.X, size.Y, m.ImageProcessor.numChannels)
if err != nil {
return nil, err
}
pixelValues := ctx.Input().FromFloatSlice(f32s, size.X, size.Y, m.ImageProcessor.numChannels)
visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
features, size := m.MultiModalProjector.Forward(ctx, visionOutputs, size)
@@ -161,15 +158,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
}

16
model/models/mistral3/model_vision.go
View File

@@ -110,15 +110,8 @@ func (m *VisionModel) positionalEmbedding(ctx ml.Context, positionIDs ml.Tensor)
}
}
h, err := ctx.Input().FromFloatSlice(frequenciesHeight, maxPatchesPerSide, frequencies/2)
if err != nil {
panic(err)
}
w, err := ctx.Input().FromFloatSlice(frequenciesWidth, maxPatchesPerSide, frequencies/2)
if err != nil {
panic(err)
}
h := ctx.Input().FromFloatSlice(frequenciesHeight, maxPatchesPerSide, frequencies/2)
w := ctx.Input().FromFloatSlice(frequenciesWidth, maxPatchesPerSide, frequencies/2)
h = h.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
w = w.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
@@ -151,10 +144,7 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor) ml.Tensor {
}
}
positionIDs, err := ctx.Input().FromIntSlice(positions, len(positions))
if err != nil {
panic(err)
}
positionIDs := ctx.Input().FromIntSlice(positions, len(positions))
positionEmbedding := m.positionalEmbedding(ctx, positionIDs)
cos, sin := positionEmbedding.Cos(ctx), positionEmbedding.Sin(ctx)

22
model/models/mllama/model.go
View File

@@ -80,15 +80,8 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
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
}
aspectRatio, err := ctx.Input().FromIntSlice([]int32{int32(ratio.rank)}, 1)
if err != nil {
return nil, err
}
pixelValues := ctx.Input().FromFloatSlice(f32s, m.imageSize, m.imageSize, m.numChannels, m.maxNumTiles)
aspectRatio := ctx.Input().FromIntSlice([]int32{int32(ratio.rank)}, 1)
positionIDs := ctx.Arange(0, 1601, 1, ml.DTypeI32)
crossAttentionStates := m.VisionModel.Forward(ctx, pixelValues, positionIDs, aspectRatio)
@@ -113,15 +106,8 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
crossAttentionStates = batch.Multimodal[len(batch.Multimodal)-1].Multimodal[0].Tensor
}
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
// TODO: attention mask, cross attention mask
return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, crossAttentionStates, nil, m.Cache.(*kvcache.WrapperCache)), nil

10
model/models/qwen2/model.go
View File

@@ -100,10 +100,7 @@ type Model struct {
// Forward implements model.Model.
func (m Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
@@ -112,10 +109,7 @@ func (m Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
var outputs ml.Tensor
if i == len(m.Layers)-1 {
outputs, err = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
outputs = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
}
hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, &m.Options)

16
model/models/qwen25vl/model.go
View File

@@ -69,10 +69,7 @@ func (m *Model) PixelValues(ctx ml.Context, multimodalData []byte) (ml.Tensor, *
m.ImageProcessor.patchSize * m.ImageProcessor.patchSize
numPatches := grid.Temporal * grid.Height * grid.Width
pixelValues, err := ctx.Input().FromFloatSlice(f32s, patchDim, numPatches)
if err != nil {
return nil, nil, fmt.Errorf("failed to create tensor from image: %w", err)
}
pixelValues := ctx.Input().FromFloatSlice(f32s, patchDim, numPatches)
return pixelValues, grid, nil
}
@@ -142,15 +139,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache)
}

22
model/models/qwen25vl/model_vision.go
View File

@@ -1,7 +1,6 @@
package qwen25vl
import (
"fmt"
"math"
"slices"
@@ -44,10 +43,8 @@ func blockDiagonalMask(ctx ml.Context, seqLength int, bounds []int, numHeads int
}
}
mask, err := ctx.Input().FromFloatSlice(flat, seqLength, seqLength)
if err != nil {
panic(err)
}
mask := ctx.Input().FromFloatSlice(flat, seqLength, seqLength)
// Reshape to match [seqLength, seqLength, 1] for broadcasting
mask = mask.Reshape(ctx, seqLength, seqLength, 1)
@@ -303,10 +300,7 @@ func (m *VisionModel) WindowIndex(ctx ml.Context, grid *Grid) (ml.Tensor, []int)
}
}
t, err := ctx.Input().FromIntSlice(index, len(index))
if err != nil {
panic(err)
}
t := ctx.Input().FromIntSlice(index, len(index))
return t, bounds
}
@@ -326,10 +320,7 @@ func (m *VisionModel) PositionalEmbedding(ctx ml.Context, grid *Grid) ml.Tensor
freqVals[i*freq+j] = float32(i) / float32(math.Pow(theta, float64(j*2)/float64(dim)))
}
}
freqs, err := ctx.Input().FromFloatSlice(freqVals, freq, maxGridSize)
if err != nil {
panic(fmt.Errorf("failed to create tensor from frequencies: %w", err))
}
freqs := ctx.Input().FromFloatSlice(freqVals, freq, maxGridSize)
// Create position coordinates (y,x pairs) for the grid
// In PyTorch: Equivalent to generating position ids with torch.arange()
@@ -339,10 +330,7 @@ func (m *VisionModel) PositionalEmbedding(ctx ml.Context, grid *Grid) ml.Tensor
coords = append(coords, int32(y), int32(x))
}
}
pos, err := ctx.Input().FromIntSlice(coords, 2, grid.Width, grid.Height)
if err != nil {
panic(fmt.Errorf("failed to create tensor from positions: %w", err))
}
pos := ctx.Input().FromIntSlice(coords, 2, grid.Width, grid.Height)
// Reshape and permute positions to match spatial merging pattern
pos = pos.Reshape(ctx, 2, grid.Width, merge, grid.Height/merge)

10
model/models/qwen3/model.go
View File

@@ -156,10 +156,7 @@ type Model struct {
// Forward implements model.Model.
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
if err != nil {
return nil, err
}
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
@@ -168,10 +165,7 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
var outputs ml.Tensor
if i == len(m.Layers)-1 {
outputs, err = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
if err != nil {
return nil, err
}
outputs = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
}
hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, m.Options)