mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-22 14:53:56 +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

11
model/models/gemma3/model_text.go
View File

@@ -7,7 +7,6 @@ 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/input"
)
@@ -20,6 +19,10 @@ type TextConfig struct {
largeModelScaling bool
}
func (o TextConfig) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor, base float32) ml.Tensor {
return nn.RoPE(ctx, states, positions, o.attnKeyLen, base, 1./o.ropeScale, rope.WithTypeNeoX())
}
type TextModel struct {
TokenEmbedding *nn.Embedding `gguf:"token_embd"`
Layers []TextLayer `gguf:"blk"`
@@ -87,7 +90,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
q := sa.Query.Forward(ctx, hiddenState)
q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
q = sa.QueryNorm.Forward(ctx, q, opts.eps)
q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
q = opts.applyRotaryPositionEmbeddings(ctx, q, positionIDs, ropeBase)
if opts.largeModelScaling {
q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -98,7 +101,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
k := sa.Key.Forward(ctx, hiddenState)
k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
k = sa.KeyNorm.Forward(ctx, k, opts.eps)
k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
k = opts.applyRotaryPositionEmbeddings(ctx, k, positionIDs, ropeBase)
v := sa.Value.Forward(ctx, hiddenState)
v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -116,7 +119,7 @@ func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.T
ropeBase = m.TextConfig.ropeGlobalBase
}
return fast.RoPE(ctx, key, shift, m.TextConfig.attnKeyLen, ropeBase, 1/m.TextConfig.ropeScale, rope.WithTypeNeoX()), nil
return m.applyRotaryPositionEmbeddings(ctx, key, shift, ropeBase), nil
}
type TextMLP struct {