mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-22 23:03:55 +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

13
model/models/qwen2/model.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"
"github.com/ollama/ollama/model/input"
@@ -22,6 +21,10 @@ type Options struct {
eps, ropeBase, ropeScale float32
}
func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
return nn.RoPE(ctx, states, positions, cmp.Or(o.ropeDim, o.headDim, o.hiddenSize/o.numHeads), o.ropeBase, 1./o.ropeScale, rope.WithTypeNeoX())
}
type Attention struct {
Query *nn.Linear `gguf:"attn_q"`
Key *nn.Linear `gguf:"attn_k"`
@@ -32,7 +35,6 @@ type Attention struct {
func (attn Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
batchSize := hiddenStates.Dim(1)
headDim := cmp.Or(opts.headDim, opts.hiddenSize/opts.numHeads)
ropeDim := cmp.Or(opts.ropeDim, headDim)
query := attn.Query.Forward(ctx, hiddenStates)
query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
@@ -43,8 +45,8 @@ func (attn Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor,
value := attn.Value.Forward(ctx, hiddenStates)
value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
@@ -123,8 +125,7 @@ func (m Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
}
func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
ropeDim := cmp.Or(m.ropeDim, m.hiddenSize/m.numHeads)
return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithTypeNeoX()), nil
return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
}
func New(c fs.Config) (model.Model, error) {