vicuna-13b-v1.5-medusa1-3head / medusa_model_demo.py
charlescao2023's picture
Upload folder using huggingface_hub
58cb504 verified
from medusa.model.medusa_model_legacy import MedusaConfig
from medusa.model.medusa_model_new import MedusaModel
import torch
import os
import json
import argparse
def read_json(file_path):
with open(file_path, 'r') as file:
data = json.load(file)
return data
def write_json(data, file_path):
with open(file_path, 'w') as file:
json.dump(data, file, indent=4)
os.environ["CUDA_VISIBLE_DEVICES"] = "7"
def main():
parser = argparse.ArgumentParser(description='Example extract medusa head')
parser.add_argument('-m', '--model-path', type=str, required=True, help='Path to new medusa models')
parser.add_argument('-o', '--output-dir', type=str, required=True, help='Path to save medusa head')
parser.add_argument('--use-full-key', action='store_true', help='medusa head keys with medusa_head. or not')
parser.add_argument('--verbose', action='store_true', help='Enable verbose output')
args = parser.parse_args()
config = read_json(os.path.join(args.model_path, "config.json"))
new_config = {}
for k, v in config.items():
if k.startswith("medusa"):
new_config[k] = v
new_config["transformers_version"] = config["transformers_version"]
new_config["base_model_name_or_path"] = config["_name_or_path"]
model = MedusaModel.from_pretrained(pretrained_model_name_or_path=args.model_path)
if args.verbose:
print(model)
medusa_lm_head = dict()
for k, v in model.state_dict().items():
if k.startswith("medusa"):
if args.verbose: print(k)
if args.use_full_key:
medusa_lm_head[k] = v
else:
medusa_lm_head[k[len("medusa_head."):]] = v
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
write_json(new_config, f"{args.output_dir}/config.json")
model_save_path = f"{args.output_dir}/medusa_lm_head.pt"
torch.save(medusa_lm_head, model_save_path)
print(f"Model saved to {model_save_path}")
if __name__ == "__main__":
main()