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

17
model/models/gptoss/model.go
View File

@@ -9,7 +9,6 @@ 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"
"github.com/ollama/ollama/model/input"
@@ -52,7 +51,7 @@ func (m *Transformer) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, err
}
func (m *Transformer) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
return fast.RoPE(ctx, key, shift, m.headDim(), m.ropeBase, 1./m.ropeScale, m.RoPEOptions()...), nil
return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
}
type Options struct {
@@ -70,14 +69,14 @@ type Options struct {
ropeScale float32
}
func (o Options) RoPEOptions() []func(*rope.Options) {
return []func(*rope.Options){
func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
return nn.RoPE(ctx, states, positions, o.headDim(), o.ropeBase, 1./o.ropeScale,
rope.WithTypeNeoX(),
rope.WithOriginalContextLength(o.originalContextLength),
rope.WithExtrapolationFactor(1.),
// NOTE: ggml sets this implicitly so there's no need to set it here
// rope.WithAttentionFactor(0.1*float32(math.Log(float64(o.ropeScale))) + 1.0),
}
// NOTE: ggml sets this implicitly so there's no need to set it here
// rope.WithAttentionFactor(0.1*float32(math.Log(float64(o.ropeScale))) + 1.0),
)
}
func (o Options) headDim() int {
@@ -135,8 +134,8 @@ func (attn *AttentionBlock) Forward(ctx ml.Context, hiddenStates, positions ml.T
value = value.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
}
query = fast.RoPE(ctx, query, positions, opts.headDim(), opts.ropeBase, 1./opts.ropeScale, opts.RoPEOptions()...)
key = fast.RoPE(ctx, key, positions, opts.headDim(), opts.ropeBase, 1./opts.ropeScale, opts.RoPEOptions()...)
query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
attention := nn.AttentionWithSinks(ctx, query, key, value, attn.Sinks, 1/math.Sqrt(float64(opts.headDim())), cache)
attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)