mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 14:26:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

chore: update models to use slice/chunk/chunksections (#12934)

Browse Source 
* use slice/chunks

* bert

* llama4

* gemma3n

* gptoss

* mistral3

* qwen3vl

* qwen25vl

* deepseek2

* remove unused ops
This commit is contained in:
Michael Yang
2025-11-13 15:20:12 -08:00
committed by GitHub
parent c114987523
commit 333203d871
13 changed files with 59 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
convert/convert_gptoss.go
View File

@@ -110,9 +110,12 @@ func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
for name, mxfp4 := range mxfp4s {
dims := mxfp4.blocks.Shape()
if !strings.HasSuffix(name, ".weight") {
name = name + ".weight"
}
if strings.Contains(name, "ffn_down_exps") {
out = append(out, &ggml.Tensor{
Name: name + ".weight",
Name: name,
Kind: uint32(ggml.TensorTypeMXFP4),
Shape: []uint64{dims[0], dims[1], dims[2] * dims[3] * 2},
WriterTo: mxfp4,
@@ -121,12 +124,12 @@ func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
// gate_up_exps is interleaved, need to split into gate_exps and up_exps
// e.g. gate_exps, up_exps = gate_up_exps[:, 0::2, ...], gate_up_exps[:, 1::2, ...]
out = append(out, &ggml.Tensor{
Name: strings.Replace(name, "gate_up", "gate", 1) + ".weight",
Name: strings.Replace(name, "gate_up", "gate", 1),
Kind: uint32(ggml.TensorTypeMXFP4),
Shape: []uint64{dims[0], dims[1] / 2, dims[2] * dims[3] * 2},
WriterTo: mxfp4.slice(1, 0, int(dims[1]), 2),
}, &ggml.Tensor{
Name: strings.Replace(name, "gate_up", "up", 1) + ".weight",
Name: strings.Replace(name, "gate_up", "up", 1),
Kind: uint32(ggml.TensorTypeMXFP4),
Shape: []uint64{dims[0], dims[1] / 2, dims[2] * dims[3] * 2},
WriterTo: mxfp4.slice(1, 1, int(dims[1]), 2),