mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-24 23:48:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(llama): other llama flavours (#12308)

Browse Source 
* fix(llama): rope scale

* spm llama

* skip moe models

* cleanup
This commit is contained in:
Michael Yang
2025-09-17 12:12:21 -07:00
committed by GitHub
parent a417ac97ee
commit 564b558c92
10 changed files with 75 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
model/models/gemma3n/model_text.go
View File

@@ -95,7 +95,7 @@ func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.T
ropeBase = m.ropeBaseLocal
}
return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, 1./m.ropeScale, rope.WithTypeNeoX()), nil
}
type TextScaledWordEmbedding struct {
@@ -256,14 +256,14 @@ func (attn TextAttention) Forward(ctx ml.Context, hiddenStates, positions ml.Ten
query := attn.Query.Forward(ctx, hiddenStates)
query = query.Reshape(ctx, opts.headDim(), opts.numHeads, batchSize)
query = attn.QueryNorm.Forward(ctx, query, opts.eps)
query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
var key, value ml.Tensor
if !sharedKV {
key = attn.Key.Forward(ctx, hiddenStates)
key = key.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
key = attn.KeyNorm.Forward(ctx, key, opts.eps)
key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
value = attn.Value.Forward(ctx, hiddenStates)
value = value.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
@@ -349,7 +349,7 @@ func newTextModel(c fs.Config) *TextModel {
eps: c.Float("attention.layer_norm_rms_epsilon", 1e-06),
ropeBase: c.Float("rope.freq_base", 1_000_000),
ropeBaseLocal: c.Float("rope.freq_base_local", 10_000),
ropeScale: c.Float("rope.freq_scale", 1.0),
ropeScale: c.Float("rope.scaling.factor", 1.0),
slidingWindowPattern: c.Bools("attention.sliding_window_pattern"),
activationSparsityScale: c.Floats("activation_sparsity_scale"),