mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-23 23:18:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor rope

Browse Source 
change to a flatter directory structure and group the options with the
function

update models to call rope in one place
This commit is contained in:
Michael Yang
2025-11-18 15:17:03 -08:00
committed by Michael Yang
parent e082d60a24
commit 603ceefaa6
21 changed files with 114 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
model/models/mistral3/model_vision.go
View File

@@ -16,8 +16,8 @@ func rotateHalf(ctx ml.Context, t ml.Tensor) ml.Tensor {
return x2.Scale(ctx, -1).Concat(ctx, x1, 0)
}
func applyRotaryPositionalEmbedding(ctx ml.Context, t, cos, sin ml.Tensor) ml.Tensor {
return t.Mul(ctx, cos).Add(ctx, rotateHalf(ctx, t).Mul(ctx, sin))
func applyRotaryPositionEmbeddings(ctx ml.Context, states, cos, sin ml.Tensor) ml.Tensor {
return states.Mul(ctx, cos).Add(ctx, rotateHalf(ctx, states).Mul(ctx, sin))
}
type VisionSelfAttention struct {
@@ -36,8 +36,8 @@ func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenStates, cos, sin ml
key = key.Reshape(ctx, opts.headDim, opts.numHeads, key.Dim(1), batchSize)
value = value.Reshape(ctx, opts.headDim, opts.numHeads, value.Dim(1), batchSize)
query = applyRotaryPositionalEmbedding(ctx, query, cos, sin)
key = applyRotaryPositionalEmbedding(ctx, key, cos, sin)
query = applyRotaryPositionEmbeddings(ctx, query, cos, sin)
key = applyRotaryPositionEmbeddings(ctx, key, cos, sin)
attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(opts.headDim)), nil)
attention = attention.Reshape(ctx, opts.hiddenSize, attention.Dim(2), batchSize)