mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-23 15:08:27 +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

9
model/models/qwen3vl/model_text.go
View File

@@ -10,7 +10,6 @@ import (
"github.com/ollama/ollama/kvcache"
"github.com/ollama/ollama/ml"
"github.com/ollama/ollama/ml/nn"
"github.com/ollama/ollama/ml/nn/fast"
"github.com/ollama/ollama/ml/nn/rope"
"github.com/ollama/ollama/model"
)
@@ -35,8 +34,8 @@ func (o TextOptions) headDim() int {
return cmp.Or(o.keyLength, o.valueLength, o.hiddenSize/o.numHeads)
}
func (o TextOptions) applyRotaryPositionalEmbedding(ctx ml.Context, t, p ml.Tensor) ml.Tensor {
return fast.RoPE(ctx, t, p, o.headDim(), o.ropeBase, 1/float32(math.Sqrt(float64(o.ropeScale))),
func (o TextOptions) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
return nn.RoPE(ctx, states, positions, o.headDim(), o.ropeBase, 1/float32(math.Sqrt(float64(o.ropeScale))),
rope.WithInterleaveMRoPE(o.mropeSections),
)
}
@@ -64,8 +63,8 @@ func (sa *TextAttention) Forward(ctx ml.Context, hiddenStates, positions ml.Tens
query = sa.QueryNorm.Forward(ctx, query, opts.eps)
key = sa.KeyNorm.Forward(ctx, key, opts.eps)
query = opts.applyRotaryPositionalEmbedding(ctx, query, positions)
key = opts.applyRotaryPositionalEmbedding(ctx, key, positions)
query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(opts.headDim())), cache)
attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)