matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-13 01:07:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

chore: update mllama to use ollama engine (#10637)

Browse Source
This commit is contained in:
Michael Yang
2025-05-13 17:36:02 -07:00
committed by GitHub
parent 0478d440f0
commit 23125648b8
67 changed files with 785 additions and 4354 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
model/models/mllama/model_text.go
View File

@@ -18,7 +18,7 @@ type TextSelfAttention struct {
RopeFactors ml.Tensor `gguf:"rope_freqs.weight"`
}
func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
batchSize := hiddenState.Dim(1)
headDim := opts.hiddenSize / opts.numHeads
ropeType := uint32(0)
@@ -69,11 +69,11 @@ type TextSelfAttentionDecoderLayer struct {
MLP *TextMLP
}
func (d *TextSelfAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, positions, outputs, mask, _, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
func (d *TextSelfAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, positions, outputs, _, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
residual := hiddenState
hiddenState = d.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
hiddenState = d.SelfAttention.Forward(ctx, hiddenState, positions, mask, cache, opts)
hiddenState = d.SelfAttention.Forward(ctx, hiddenState, positions, cache, opts)
// In the final layer (outputs != nil), optimize by pruning to just the token positions
// we need logits for.
@@ -151,7 +151,7 @@ type TextCrossAttentionDecoderLayer struct {
MLPGate ml.Tensor `gguf:"cross_attn_mlp_gate"`
}
func (d *TextCrossAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, _, _, _, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
func (d *TextCrossAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, _, _, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
residual := hiddenState
hiddenState = d.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
@@ -167,14 +167,14 @@ func (d *TextCrossAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, _,
}
type TextDecoderLayer interface {
Forward(ctx ml.Context, hiddenState, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor
Forward(ctx ml.Context, hiddenState, positionIDs, outputs, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor
}
type TextDecoder struct {
Layers []TextDecoderLayer
}
func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
for i, layer := range d.Layers {
layerType := selfAttentionLayer
if slices.Contains(opts.crossAttentionLayers, int32(i)) {
@@ -190,7 +190,7 @@ func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs,
lastLayerOutputs = outputs
}
hiddenState = layer.Forward(ctx, hiddenState, positionIDs, lastLayerOutputs, mask, crossAttentionStates, crossAttentionMask, cache, opts)
hiddenState = layer.Forward(ctx, hiddenState, positionIDs, lastLayerOutputs, crossAttentionStates, crossAttentionMask, cache, opts)
}
}
@@ -214,9 +214,9 @@ type TextModel struct {
*TextModelOptions
}
func (m *TextModel) Forward(ctx ml.Context, inputIDs, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache) ml.Tensor {
func (m *TextModel) Forward(ctx ml.Context, inputIDs, positionIDs, outputs, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache) ml.Tensor {
hiddenState := m.TokenEmbedding.Forward(ctx, inputIDs)
hiddenState = m.Transformer.Forward(ctx, hiddenState, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask, cache, m.TextModelOptions)
hiddenState = m.Transformer.Forward(ctx, hiddenState, positionIDs, outputs, crossAttentionStates, crossAttentionMask, cache, m.TextModelOptions)
hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
return m.Output.Forward(ctx, hiddenState)
}