vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update default values for n_embd_head_k and n_embd_head_v

Browse source 
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
postmasters 2023-12-31 12:55:33 -08:00 • committed by GitHub
parent f56cbce6b7
commit 522e534903
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
llama.cpp
View file

@ -2681,10 +2681,10 @@ static void llm_load_hparams(
// gpt-j n_rot = rotary_dim // gpt-j n_rot = rotary_dim
} }
hparams.n_embd_head_k = hparams.n_embd / hparams.n_head_kv; hparams.n_embd_head_k = hparams.n_embd / hparams.n_head;
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false);
hparams.n_embd_head_v = hparams.n_embd / hparams.n_head_kv; hparams.n_embd_head_v = hparams.n_embd / hparams.n_head;
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
// arch-specific KVs // arch-specific KVs