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

fix write gguf padding

Browse Source
This commit is contained in:
Michael Yang
2025-04-11 13:39:51 -07:00
committed by Michael Yang
parent 1e7f62cb42
commit 2fec73eef6
3 changed files with 13 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
convert/convert_llama.go
View File

@@ -28,12 +28,12 @@ type llamaModel struct {
NumKeyValueHeads uint32 `json:"num_key_value_heads"`
RopeTheta float32 `json:"rope_theta"`
RopeScaling struct {
Type string `json:"type"`
RopeType string `json:"rope_type"`
Factor float32 `json:"factor"`
LowFrequencyFactor float32 `json:"low_freq_factor"`
HighFrequencyFactor float32 `json:"high_freq_factor"`
OriginalMaxPositionalEmbeddings uint32 `json:"original_max_positional_embeddings"`
Type string `json:"type"`
RopeType string `json:"rope_type"`
Factor float32 `json:"factor"`
LowFrequencyFactor float32 `json:"low_freq_factor"`
HighFrequencyFactor float32 `json:"high_freq_factor"`
OriginalMaxPositionEmbeddings uint32 `json:"original_max_position_embeddings"`
factors ropeFactor
} `json:"rope_scaling"`
@@ -84,7 +84,7 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
original := cmp.Or(p.RopeScaling.OriginalMaxPositionEmbeddings, 8192)
lambdaLow := float32(original) / factorLow
lambdaHigh := float32(original) / factorHigh