mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-11 18:36:41 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

fixes after rebase

Browse source
This commit is contained in:
Bruce MacDonald 2025-04-28 14:19:26 -07:00
parent db5da138c9
commit b2a051375f
4 changed files with 60 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

36
model/models/llama4/model_text.go
View file

@ -31,8 +31,28 @@ func (sa *TextAttention) Forward(ctx ml.Context, hiddenStates, positions, attent
value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
if useRope {
query = query.RoPE(ctx, positions, sa.RopeFactors, uint32(opts.ropeDim), uint32(0), opts.ropeBase, opts.ropeScale)
key = key.RoPE(ctx, positions, sa.RopeFactors, uint32(opts.ropeDim), uint32(0), opts.ropeBase, opts.ropeScale)
query = query.RoPE(
ctx,
positions,
sa.RopeFactors,
ml.RoPEConfig{
Dim: uint32(opts.ropeDim),
Type: ml.RopeTypeNormal,
Base: opts.ropeBase,
Scale: opts.ropeScale,
},
)
key = key.RoPE(
ctx,
positions,
sa.RopeFactors,
ml.RoPEConfig{
Dim: uint32(opts.ropeDim),
Type: ml.RopeTypeNormal,
Base: opts.ropeBase,
Scale: opts.ropeScale,
},
)
}
if opts.useQKNorm {
@ -255,5 +275,15 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
}
func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
return key.RoPE(ctx, shift, m.Layers[layer].Attention.RopeFactors, uint32(0), uint32(m.ropeDim), m.ropeBase, m.ropeScale), nil
return key.RoPE(
ctx,
shift,
m.Layers[layer].Attention.RopeFactors,
ml.RoPEConfig{
Dim: uint32(m.TextOptions.ropeDim),
Type: ml.RopeTypeNormal,
Base: m.TextOptions.ropeBase,
Scale: m.TextOptions.ropeScale,
},
), nil
}

2
model/models/mllama/model_text.go
View file

@ -240,7 +240,7 @@ func newTextModel(c fs.Config) *TextModel {
numHeads: int(c.Uint("attention.head_count")),
numKVHeads: int(c.Uint("attention.head_count_kv")),
eps: c.Float("attention.layer_norm_rms_epsilon"),
crossAttentionLayers: c.Uints("attention.cross_attention_layers"),
crossAttentionLayers: c.Ints("attention.cross_attention_layers"),
ropeConfig: ml.RoPEConfig{
Base: c.Float("rope.freq_base"),
Scale: c.Float("rope.freq_scale", 1),

2
model/models/qwen25vl/model_text.go
View file

@ -35,7 +35,7 @@ func NewTextModel(c fs.Config) *TextModel {
c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
&model.Vocabulary{
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Uints("tokenizer.ggml.token_type"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
BOS: int32(c.Uint("tokenizer.ggml.bos_token_id")),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),

33
server/create.go
View file

@ -514,17 +514,34 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
} else if err != nil {
return nil, err
}
}
// Fallback to creating layer from file copy (either NewLayerFromLayer failed, or digest empty/n != stat.Size())
if layer.Digest == "" {
layer, err = NewLayer(io.NewSectionReader(blob, 0, n), mediatype)
if err != nil {
return nil, err
mediatype := "application/vnd.ollama.image.model"
if f.KV().Kind() == "adapter" {
mediatype = "application/vnd.ollama.image.adapter"
} else if _, ok := f.KV()[fmt.Sprintf("%s.vision.block_count", f.KV().Architecture())]; ok || f.KV().Kind() == "projector" {
mediatype = "application/vnd.ollama.image.projector"
}
}
layers = append(layers, &layerGGML{layer, f})
var layer Layer
if digest != "" && n == stat.Size() && offset == 0 {
layer, err = NewLayerFromLayer(digest, mediatype, blob.Name())
if err != nil {
slog.Debug("could not create new layer from layer", "error", err)
return nil, err
}
}
// Fallback to creating layer from file copy (either NewLayerFromLayer failed, or digest empty/n != stat.Size())
if layer.Digest == "" {
layer, err = NewLayer(io.NewSectionReader(blob, offset, n), mediatype)
if err != nil {
return nil, err
}
}
layers = append(layers, &layerGGML{layer, f})
offset = n
}
return detectChatTemplate(layers)
}