|
--- |
|
license: llama2 |
|
language: |
|
- zh |
|
tags: |
|
- text-generation-inference |
|
--- |
|
This language model was finetuned with a dataset of 48k machine-translated Chinese instructions. For dataset description, inference examples and other details, see: https://github.com/magichub-opensource/CLAM-Conversational-Language-AI-from-MagicData. |
|
|
|
### 模型推理 |
|
|
|
* 单卡加载一个模型需要15G显存。 |
|
* 本地测试环境：py310-torch1.13.1-cuda11.6-cudnn8 |
|
|
|
#### Web Demo |
|
|
|
我们使用 [text-generation-webui](https://github.com/oobabooga/text-generation-webui/tree/main) 开源项目搭建的 demo 进行推理，得到文档中的对比样例。该demo支持在网页端切换模型、调整多种常见参数等。 |
|
|
|
实验环境：py310-torch1.13.1-cuda11.6-cudnn8 |
|
|
|
``` |
|
git clone https://github.com/oobabooga/text-generation-webui.git |
|
cd text-generation-webui |
|
pip install -r requirements.txt |
|
|
|
# 建议使用软链接将模型绝对路径链至 `./models`。也可以直接拷贝进去。 |
|
ln -s ${model_dir_absolute_path} models/${model_name} |
|
|
|
# 启动服务 |
|
python server.py --model ${model_name} --listen --listen-host 0.0.0.0 --listen-port ${port} |
|
``` |
|
如果服务正常启动，就可以通过该端口访问服务了 `${server_ip}:${port}` |
|
|
|
#### Inference script |
|
|
|
See https://github.com/magichub-opensource/CLAM-Conversational-Language-AI-from-MagicData/blob/master/inference.py |
|
|
|
``` |
|
import os,sys,argparse |
|
# os.environ['CUDA_VISIBLE_DEVICES'] = '1' |
|
import torch |
|
import re |
|
import transformers |
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
# modelpath = 'models/Chinese-llama2-alpaca-7b' # local path |
|
modelpath = 'MagicHub/Chinese-llama2-alpaca-7b' # huggingface repo |
|
|
|
print(f'model path: {modelpath}') |
|
model = AutoModelForCausalLM.from_pretrained(modelpath, device_map="cuda:0", torch_dtype=torch.float16) |
|
tokenizer = AutoTokenizer.from_pretrained(modelpath, use_fast=False) |
|
|
|
prompt = "歌剧和京剧的区别是什么？\n" |
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda:0") |
|
generate_ids = model.generate( |
|
inputs.input_ids, do_sample=True, max_new_tokens=1024, top_k=10, top_p=0.1, temperature=0.5, repetition_penalty=1.18, |
|
eos_token_id=2, bos_token_id=1, pad_token_id=0, typical_p=1.0,encoder_repetition_penalty=1, |
|
) |
|
response = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] |
|
cleaned_response = re.sub('^'+prompt,'', response) |
|
print(f'输入：\n{prompt}\n') |
|
print(f"输出：\n{cleaned_response}\n") |
|
``` |
