zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-08-31 11:54:37 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Update convert.py

Browse Source
This commit is contained in:
Anchen 2023-12-30 19:35:25 +11:00 committed by GitHub
parent 581a5733a1
commit c687a4e02e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 4 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
llms/deepseek-coder/convert.py
View File

@ -44,7 +44,7 @@ def convert(args):
config = model.config.to_dict()
state_dict = model.state_dict()
tokenizer = AutoTokenizer.from_pretrained(str(hf_path), trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(str(hf_path), trust_remote_code=True, use_fast=False)
# things to change
# 1. there's no "model." in the weight names
@ -84,7 +84,7 @@ def convert(args):
weights = {k: v.numpy() for k, v in state_dict.items()}
config["rope_scaling_factor"] = config["rope_scaling"]["factor"]
config["rope_scaling_factor"] = config["rope_scaling"]["factor"] if config["rope_scaling"] is not None else 1.0
keep_keys = set(
[
"vocab_size",
@ -96,6 +96,7 @@ def convert(args):
"rms_norm_eps",
"intermediate_size",
"rope_scaling_factor",
"rope_theta"
]
)
for k in list(config.keys()):
@ -151,4 +152,4 @@ if __name__ == "__main__":
tokenizer.save_pretrained(mlx_path)
with open(mlx_path / "config.json", "w") as f:
config["model_type"] = "deepseek_coder"
json.dump(config, f, indent=4)
json.dump(config, f, indent=4)%