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

9
model/models/nomicbert/model.go
View File

@@ -7,7 +7,6 @@ import (
"github.com/ollama/ollama/fs"
"github.com/ollama/ollama/ml"
"github.com/ollama/ollama/ml/nn"
"github.com/ollama/ollama/ml/nn/fast"
"github.com/ollama/ollama/ml/nn/pooling"
"github.com/ollama/ollama/ml/nn/rope"
"github.com/ollama/ollama/model"
@@ -37,6 +36,10 @@ type Options struct {
ropeFreqBase float32
}
func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
return nn.RoPE(ctx, states, positions, o.headDim, o.ropeFreqBase, 1.0, rope.WithTypeNeoX())
}
// Single Encoder Layer
type EncoderLayer struct {
*Attention
@@ -105,8 +108,8 @@ func (a *Attention) Forward(ctx ml.Context, hiddenStates ml.Tensor, positions ml
chunks := qkv.Chunk(ctx, 1, opts.numHeads)
query, key, value := chunks[0], chunks[1], chunks[2]
query = fast.RoPE(ctx, query, positions, opts.headDim, opts.ropeFreqBase, 1.0, rope.WithTypeNeoX())
key = fast.RoPE(ctx, key, positions, opts.headDim, opts.ropeFreqBase, 1.0, 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(opts.headDim)), nil)