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

llama : implement YaRN RoPE scaling (#2268)

Browse source 
Co-authored-by: cebtenzzre <cebtenzzre@gmail.com>
Co-authored-by: Jeffrey Quesnelle <jquesnelle@gmail.com>
This commit is contained in:
cebtenzzre 2023-11-01 18:04:33 -04:00 • committed by GitHub
parent c43c2da8af
commit 898aeca90a
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
15 changed files with 763 additions and 257 deletions
Download patch file Download diff file Expand all files Collapse all files

3
convert-baichuan-hf-to-gguf.py
View file

@ -163,7 +163,8 @@ gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
if "type" in hparams["rope_scaling"]:
if hparams["rope_scaling"]["type"] == "linear":
gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
# TOKENIZATION