matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-11 00:07:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

gemma2 impl

Browse Source
This commit is contained in:
Patrick Devine
2025-02-07 15:58:15 -08:00
committed by Michael Yang
parent 4dcf80167a
commit 5f74d1fd47
18 changed files with 1057 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
model/models/llama/model.go
View File

@@ -76,14 +76,15 @@ type SelfAttention struct {
func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
batchSize := hiddenState.Dim(1)
headDim := opts.hiddenSize / opts.numHeads
ropeType := uint32(0)
q := sa.Query.Forward(ctx, hiddenState)
q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
k := sa.Key.Forward(ctx, hiddenState)
k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
v := sa.Value.Forward(ctx, hiddenState)
v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -96,7 +97,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
}
func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, m.ropeDim, m.ropeBase, m.ropeScale), nil
return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, uint32(0), m.ropeDim, m.ropeBase, m.ropeScale), nil
}
type MLP struct {