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

Add support for new models and fix GitHub issues

Browse Source 
- Add Gemma3n model support with text generation capabilities
- Add new CUDA mean operations for improved performance
- Add macOS documentation and performance tests
- Update LLAMA patches for ROCm/CUDA compatibility
- Fix various model conversion and processing issues
- Update CI workflows and build configurations
- Add library model tests and Shakespeare test data

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
Shang Chieh Tseng
2025-07-20 00:12:36 +08:00
parent 1fa71c2670
commit cbcbc9ae07
81 changed files with 132316 additions and 747 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
convert/convert_mixtral.go
View File

@@ -2,9 +2,6 @@ package convert
import (
"fmt"
"io"
"slices"
"strings"
"github.com/ollama/ollama/fs/ggml"
)
@@ -30,65 +27,38 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
}
func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
oldnew := []string{
"model.layers", "blk",
"w1", "ffn_gate_exps",
"w2", "ffn_down_exps",
"w3", "ffn_up_exps",
}
for i := range p.NumLocalExperts {
oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
}
// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
namer := strings.NewReplacer(oldnew...)
experts := make(map[string]experts)
// merge experts into a single tensor while removing them from ts
ts = slices.DeleteFunc(ts, func(t Tensor) bool {
if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
return false
}
name := namer.Replace(t.Name())
experts[name] = append(experts[name], t)
return true
})
var out []*ggml.Tensor
for n, e := range experts {
// TODO(mxyng): sanity check experts
out = append(out, &ggml.Tensor{
Name: n,
Kind: e[0].Kind(),
Shape: append([]uint64{uint64(len(e))}, e[0].Shape()...),
WriterTo: e,
merges := make([]merge, 0, p.NumHiddenLayers*6)
for i := range p.NumHiddenLayers {
merges = append(merges, merge{
fmt.Sprintf("blk.%d.*.w1.weight", i),
fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
}, merge{
fmt.Sprintf("blk.%d.*.w1.bias", i),
fmt.Sprintf("blk.%d.ffn_gate_exps.bias", i),
}, merge{
fmt.Sprintf("blk.%d.*.w2.weight", i),
fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
}, merge{
fmt.Sprintf("blk.%d.*.w2.bias", i),
fmt.Sprintf("blk.%d.ffn_up_exps.bias", i),
}, merge{
fmt.Sprintf("blk.%d.*.w3.weight", i),
fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
}, merge{
fmt.Sprintf("blk.%d.*.w3.bias", i),
fmt.Sprintf("blk.%d.ffn_down_exps.bias", i),
})
}
out, ts := mergeTensors(ts, merges...)
return append(out, p.llamaModel.Tensors(ts)...)
}
func (p *mixtralModel) Replacements() []string {
return append(
p.llamaModel.Replacements(),
"model.layers", "blk",
"block_sparse_moe.gate", "ffn_gate_inp",
"block_sparse_moe.experts.", ".",
)
}
type experts []Tensor
func (e experts) WriteTo(w io.Writer) (int64, error) {
// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
for _, t := range e {
// the canonical merged experts tensor stacks all experts along a new, 0 axis,
// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
// this accomplishes the same thing by writing each expert tensor in sequence
if _, err := t.WriteTo(w); err != nil {
return 0, err
}
}
return 0, nil
}