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

llama4

Browse Source
This commit is contained in:
Michael Yang
2025-04-03 15:18:29 -07:00
committed by Michael Yang
parent 54055a6dae
commit f0c66e6dea
13 changed files with 833 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
convert/convert_llama.go
View File

@@ -42,6 +42,8 @@ type llamaModel struct {
LayerNormEpsilon float32 `json:"layer_norm_epsilon"`
NormEpsilon float32 `json:"norm_epsilon"`
HeadDim uint32 `json:"head_dim"`
skipRepack bool
}
var _ ModelConverter = (*llamaModel)(nil)
@@ -70,6 +72,10 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
kv["llama.rope.dimension_count"] = p.HiddenSize / headCount
}
if p.HeadDim > 0 {
kv["llama.attention.head_dim"] = p.HeadDim
}
if p.RopeTheta > 0 {
kv["llama.rope.freq_base"] = p.RopeTheta
}
@@ -133,9 +139,10 @@ func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
}
for _, t := range ts {
if strings.HasSuffix(t.Name(), "attn_q.weight") ||
strings.HasSuffix(t.Name(), "attn_k.weight") {
t.SetRepacker(p.repack)
if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
if !p.skipRepack {
t.SetRepacker(p.repack)
}
}
out = append(out, ggml.Tensor{