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

merge-hf-and-lora-to-hf.py

Browse source
This commit is contained in:
FNsi 2023-05-22 19:29:24 +08:00 • committed by GitHub
parent 1fd5d10b07
commit 29995194e3
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

10
merge-hf-and-lora-to-hf.py
View file

@ -32,11 +32,11 @@ parser.add_argument(
args = parser.parse_args()
print(f">>> load model from {args.model_path} and lora from {args.lora_path}....")
print(f">>> load model from {args.model} and lora from {args.lora}....")
# transformer loaded. load and save Tokenizer.
tokenizer = LlamaTokenizer.from_pretrained(args.model_path)
tokenizer.save_pretrained(args.out_path)
tokenizer = LlamaTokenizer.from_pretrained(args.model)
tokenizer.save_pretrained(args.out)
# load model.
model = LlamaForCausalLM.from_pretrained(
@ -49,7 +49,7 @@ model = LlamaForCausalLM.from_pretrained(
# peft loaded. load lora.
model = PeftModel.from_pretrained(
model,
args.lora_path,
args.lora,
torch_dtype=torch.float16,
device_map={"": "cpu"},
)
@ -58,4 +58,4 @@ print(f">>> merging lora...")
# Using Peft function to merge Lora.
model = model.merge_and_unload()
model.save_pretrained(args.out_path)
model.save_pretrained(args.out)