Load Weights

Llama

import torch
import sys
sys.path.append('/path/of/luduan')
from models.configuration_luduan import LuduanConfig
from models.modeling_luduan import LuduanForCausalLM
from transformers import AutoModel, LlamaTokenizer, AutoModelForCausalLM
tokenizer = LlamaTokenizer.from_pretrained('decapoda-research/llama-7b-hf', trust_remote_code=True)
luduan = LuduanForCausalLM.from_pretrained('decapoda-research/llama-7b-hf').to('cuda:0')

text = "I'm a"
encoded_input = tokenizer(text, return_tensors='pt').to('cuda:0')
pred = luduan.generate(**encoded_input, max_new_tokens=64,repetition_penalty=1.1)
print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))

Baichuan

from transformers import AutoModel, AutoTokenizer, AutoModelForCausalLM
import sys
sys.path.append('/path/of/luduan')
from models.configuration_luduan import LuduanConfig
from models.modeling_luduan import LuduanForCausalLM
tokenizer = AutoTokenizer.from_pretrained('baichuan-inc/Baichuan-7B', trust_remote_code=True)

luduan = LuduanForCausalLM.from_pretrained(
    'baichuan-inc/Baichuan-7B',
    config=LuduanConfig(vocab_size=64000, is_baichuan_architecture=True)).to('cuda:0')

text = "'登鹳雀楼->王之涣\n夜雨寄北->'"
encoded_input = tokenizer(text, return_tensors='pt').to('cuda:0')
pred = luduan.generate(**encoded_input, max_new_tokens=64,repetition_penalty=1.1)
print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))

ChangeLog

• <2023-08-03 四 14:49> 采用is_baichuan_architecture来判断是否是baichuan架构，可以直接使用from_pretrained加载权重
• <2023-07-25 二 11:18> 实现加载baihuan权重, copy from state_dict（不优雅）
• <2023-07-21 五 16:39> 实现from_pretrain llama
• <2023-07-18 二 12:02> 使用Huggingface训练框架和nanoGPT，baichuan tokenizer初始化了第一个版本。