PaddleNLP使用Vicuna

LLaMA 模型

LLaMa 是一个大型语言模型，由 Meta 开源。它的全称是 Large Language Model Meta AI，参数量从 70 亿到 650 亿不等。例如，130 亿参数的 LLaMA 模型在大多数基准上可以胜过参数量达 1750 亿的 GPT-3，而且可以在单块 V100 GPU 上运行。而最大的 650 亿参数的 LLaMA 模型可以媲美谷歌的 Chinchilla-70B 和 PaLM-540B。

Vicuna 模型

Vicuna 是一个由 UC 伯克利、CMU、斯坦福等机构的学者联手发布的最新开源大模型。基于 Meta 开源的 LLaMA 大模型，使用 ShareGPT 平台上的用户共享对话数据微调而来。包含 7B 和 13B 两个型号的开源预训练模型。

下载模型

# 下载 Vicuna 7B
# !git lfs clone http://git.aistudio.baidu.com/180581/vicuna-7b-v1.1.git# 下载 Vicuna 13B
!git lfs clone http://git.aistudio.baidu.com/180581/vicuna-13b-v1.1.git

开发环境

!pip install --pre --upgrade paddlenlp -f https://www.paddlepaddle.org.cn/whl/paddlenlp.html --user
!pip install paddlepaddle-gpu==0.0.0.post112 -f https://www.paddlepaddle.org.cn/whl/linux/gpu/develop.html --user

代码

import os
import glob
import paddlefrom tqdm import tqdm
from paddlenlp.transformers import LlamaForCausalLM, LlamaConfig, LlamaTokenizerpattern = 'paddle-model-?????-of-?????.pdparams'# Vicuna 7B
# ckpt_dir = 'vicuna-7b-v1.1'
# config_dict =  {
#     "hidden_size": 4096,
#     "initializer_range": 0.02,
#     "intermediate_size": 11008,
#     "max_position_embeddings": 2048,
#     "model_type": "llama",
#     "num_attention_heads": 32,
#     "num_hidden_layers": 32,
#     "rms_norm_eps": 1e-06,
#     "vocab_size": 32000,
#     "bos_token_id": 1,
#     "eos_token_id": 2,
#     "pad_token_id": 0,
#     "use_cache": True,
#     "use_recompute": False,
#     "use_flash_attention": False,
# }# Vicuna 13B
ckpt_dir = 'vicuna-13b-v1.1'
config_dict =  {"hidden_size": 5120,"initializer_range": 0.02,"intermediate_size": 13824,"max_position_embeddings": 2048,"model_type": "llama","num_attention_heads": 40,"num_hidden_layers": 40,"rms_norm_eps": 1e-06,"vocab_size": 32000,"bos_token_id": 1,"eos_token_id": 2,"pad_token_id": 0,"use_cache": True,"use_recompute": False,"use_flash_attention": False,
}paddle.set_default_dtype('float16')tokenizer = LlamaTokenizer.from_pretrained(ckpt_dir)config = LlamaConfig(**config_dict)model = LlamaForCausalLM(config)
model.eval()for name, layer in model.named_sublayers():if 'rotary_emb' in name:layer.inv_freq = layer.inv_freq.cast(paddle.float32)paddle.device.cuda.empty_cache()for file_path in tqdm(glob.glob(os.path.join(ckpt_dir, pattern))):params = paddle.load(file_path)assert model.set_dict(params)[1] == [], 'Load error.'del paramspaddle.device.cuda.empty_cache()input_text = input('USER: ')
prompt = f'''USER: {input_text}\n\nASSISTANT: '''
with paddle.no_grad():with paddle.amp.auto_cast(False, level='O2', dtype='float16'):while True:if input_text == 'exit':breakinputs = tokenizer(prompt, return_tensors="pd", return_attention_mask=True,return_position_ids=True)outputs = model.generate(input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, position_ids=inputs.position_ids, max_length=2048-inputs.input_ids.shape[1], min_length=0, decode_strategy="sampling",temperature=0.8, top_k=40, top_p=0.95, repetition_penalty=1.1,bos_token_id=tokenizer.bos_token_id,eos_token_id=tokenizer.eos_token_id,pad_token_id=tokenizer.pad_token_id,use_cache=True, use_fast=True, use_fp16_decoding=True)response = tokenizer.decode(outputs[0][0], skip_special_tokens=True)print('ASSISTANT: ' + response)input_text = input('USER: ')prompt += f'''{response}\n\nUSER: {input_text}\n\nASSISTANT: '''del inputsdel outputsdel responsepaddle.device.cuda.empty_cache()