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

llama : use attn_factor in mscale calculation to match the rope_yarn() implementation

Browse source
This commit is contained in:
Stanisław Szymczyk 2024-05-27 14:33:31 +02:00
parent 56f70112eb
commit 82cec8b84b
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
llama.cpp
View file

@ -11186,7 +11186,7 @@ struct llm_build_context {
// We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float mscale = 1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale);
const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));