convert: capture head_dim
for mistral (#5818)
This commit is contained in:
parent
5784c05397
commit
d835368eb8
1 changed files with 5 additions and 0 deletions
|
@ -71,6 +71,11 @@ func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
|
|||
"tokenizer.ggml.unknown_token_id": uint32(0),
|
||||
}
|
||||
|
||||
if m.Params.HeadDimension > 0 {
|
||||
kv["llama.attention.key_length"] = uint32(m.Params.HeadDimension)
|
||||
kv["llama.attention.value_length"] = uint32(m.Params.HeadDimension)
|
||||
}
|
||||
|
||||
return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
|
||||
}
|
||||
|
||||
|
|
Loading…
Reference in a new issue