From 522e5349032e1de7950fcd577bc15ca6656925ca Mon Sep 17 00:00:00 2001 From: postmasters Date: Sun, 31 Dec 2023 12:55:33 -0800 Subject: [PATCH] Update default values for n_embd_head_k and n_embd_head_v Co-authored-by: Georgi Gerganov --- llama.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llama.cpp b/llama.cpp index 85fa4ad76..22ff38012 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2681,10 +2681,10 @@ static void llm_load_hparams( // 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); - 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); // arch-specific KVs