• 快速使用transformers构建鹏程·盘古α 2.6B模型
• • Model Description
  • 快速使用
  • finetune
  • • 环境
    • 依赖
    • 运行
  • CPU推理
  • • 环境
    • 依赖
    • 运行
  • GPU推理（fp16模式下）
  • • 环境(nvidia V100/A100单卡)
    • 依赖
    • 运行
  • GPU推理（int8模式下）
  • • 环境(nvidia A100单卡)
    • 依赖
    • 运行
  • 性能对比结果
  • Usage (Text Generation)

快速使用transformers构建鹏程·盘古α 2.6B模型

非常感谢一名开发者将鹏程·盘古α 2.6B模型迁移至Huggingface Transformers库 huggingface.co/imone/pangu_2_6B ，同步我们也将此项目迁移到 盘古项目下，方便用户使用。

Model Description

PanGu-α is proposed by a joint technical team headed by PCNL. It is the first large-scale Chinese pre-trained language model with 200 billion parameters trained on 2048 Ascend processors using an automatic hybrid parallel training strategy. The whole training process is done on the “Peng Cheng Cloud Brain II” computing platform with the domestic deep learning framework called MindSpore. The PengCheng·PanGu-α pre-training model can support rich applications, has strong few-shot learning capabilities, and has outstanding performance in text generation tasks such as knowledge question and answer, knowledge retrieval, knowledge reasoning, and reading comprehension.

This repository contains PyTorch implementation of PanGu model, with
2.6 billion parameters pretrained weights (FP32 precision), converted from original MindSpore checkpoint.

快速使用

finetune

环境

(说明：因为transformers库暂不支持模型并行训练，故使用V100会显存溢出，您可以使用A100单卡进行微调)
可以直接使用现成的docker镜像

docker pull nvcr.io/nvidia/pytorch:21.03-py3

依赖

pip install transformers=='4.13.0'
pip install jieba
pip install sentencepiece==0.1.94

运行

python ./pangu_2_6B/pangu_train.py

CPU推理

环境

可以直接使用现成的docker镜像

docker pull nvcr.io/nvidia/pytorch:21.03-py3

依赖

pip install transformers=='4.22.0'
pip install jieba
pip install sentencepiece==0.1.94

运行

python ./pangu_2_6B/pangu_inference_cpu.py或
使用jit模式
python ./pangu_2_6B/pangu_inference_cpu.py --no_cuda --use_ipex --jit_mode_eval

GPU推理（fp16模式下）

环境(nvidia V100/A100单卡)

可以直接使用现成的docker镜像

docker pull nvcr.io/nvidia/pytorch:21.03-py3

依赖

pip install transformers=='4.22.0'
pip install jieba
pip install sentencepiece==0.1.94

运行

python ./pangu_2_6B/pangu_inference.py 

GPU推理（int8模式下）

环境(nvidia A100单卡)

可以直接使用现成的docker镜像

docker pull nvcr.io/nvidia/pytorch:22.03-py3

依赖

pip install transformers=='4.22.0'
pip install jieba
pip install sentencepiece==0.1.94
pip install accelerate>=0.12.0
pip install bitsandbytes>=0.31.5

运行

python ./pangu_2_6B/pangu_inference_int8.py 

性能对比结果

Usage (Text Generation)

Currently PanGu model is not supported by transformers,
so trust_remote_code=True is required to load model implementation in this repo.

from transformers import TextGenerationPipeline, AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("imone/pangu_2_6B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("imone/pangu_2_6B", trust_remote_code=True)

text_generator = TextGenerationPipeline(model, tokenizer)
# greedy search
print(text_generator("中国和美国和日本和法国和加拿大和澳大利亚的首都分别是哪里？", max_length=50))

Expected output:

[{'generated_text': '中国和美国和日本和法国和加拿大和澳大利亚的首都分别是哪里？\n中国北京,美国华盛顿,日本东京,法国巴黎,加拿大多伦多,澳大利亚悉尼,新西兰奥克兰,澳大利亚墨尔本,新西兰奥克兰,'}]