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

Fixed incorrectly applying RMS norm twice

Browse source
This commit is contained in:
JohannesGaessler 2023-06-18 14:34:52 +02:00
parent 8596af4277
commit ca4ae7844d
1 changed files with 1 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

6
llama.cpp
View file

@ -1657,11 +1657,7 @@ static bool llama_eval_internal(
{
cur = ggml_rms_norm(ctx0, inpL);
offload_func_nr(cur);
ggml_set_name(cur, "rms_norm_inpL");
cur = ggml_rms_norm(ctx0, cur);
offload_func_nr(cur);
ggml_set_name(cur, "rms_norm_after");
ggml_set_name(cur, "rms_norm_2");
// cur = cur*norm(broadcasted)
cur = ggml_mul(ctx0, cur, model.norm);