matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 15:57:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

Move quantization to new backend (#10363)

Browse Source 
* Move quantization logic to GGML via new backend

This moves the model aware logic to Go code and calls GGMLs quantization code for model creation.

* Remove "add model quantizations"

This is no longer needed now that quantization is implemented in Go+GGML code directly.
This commit is contained in:
Daniel Hiltgen
2025-05-06 11:20:48 -07:00
committed by GitHub
parent 95e744beeb
commit 424810450f
39 changed files with 1854 additions and 440 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
convert/convert_llama4.go
View File

@@ -88,13 +88,13 @@ func (p *llama4Model) Replacements() []string {
}
// Tensors implements ModelConverter.
func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
var out []ggml.Tensor
func (p *llama4Model) Tensors(ts []Tensor) []*ggml.Tensor {
var out []*ggml.Tensor
var textTensors []Tensor
for _, t := range ts {
if strings.HasPrefix(t.Name(), "v.") || strings.HasPrefix(t.Name(), "mm.") {
out = append(out, ggml.Tensor{
out = append(out, &ggml.Tensor{
Name: t.Name(),
Kind: t.Kind(),
Shape: t.Shape(),
@@ -112,7 +112,7 @@ func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
// clone tensor since we need separate repackers
tt := t.Clone()
tt.SetRepacker(p.repack(nil, nil, tensor.S(i*halfDim, (i+1)*halfDim)))
out = append(out, ggml.Tensor{
out = append(out, &ggml.Tensor{
Name: strings.ReplaceAll(tt.Name(), "ffn_gate_up_exps", name),
Kind: tt.Kind(),
Shape: newShape,
@@ -125,7 +125,7 @@ func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
t.SetRepacker(p.repack())
newShape := slices.Clone(t.Shape())
newShape[1], newShape[2] = newShape[2], newShape[1]
out = append(out, ggml.Tensor{
out = append(out, &ggml.Tensor{
Name: t.Name(),
Kind: t.Kind(),
Shape: newShape,