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

use config partial_rotary_factor as rope ratio

Browse source
This commit is contained in:
liyuhang 2024-11-26 05:29:42 +00:00
parent a249dc0fbb
commit 4f696624a4
1 changed files with 1 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

5
convert_hf_to_gguf.py
View file

@ -3987,10 +3987,7 @@ class ChatGLMModel(Model):
rope_dim = self.hparams["attention_dim"]
else:
rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
if "THUDM/glm4-nano" in self.hparams.get("_name_or_path", "") or "THUDM/glm4-mini" in self.hparams.get("_name_or_path", ""):
self.gguf_writer.add_rope_dimension_count(rope_dim)
else:
self.gguf_writer.add_rope_dimension_count(int(rope_dim/2))
self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
self.gguf_writer.add_add_bos_token(False)
rope_freq = 10000
if "rope_ratio" in self.hparams: