Pretrain CPM - 大规模预训练语言模型的预训练代码

Overview

CPM-Pretrain

版本更新记录

为了促进中文自然语言处理研究的发展,本项目提供了大规模预训练语言模型的预训练代码。项目主要基于DeepSpeed、Megatron实现,可以支持数据并行、模型加速、流水并行的代码。

安装

1、首先安装pytorch等基础依赖,再安装APEX以支持fp16。

pip install torch
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

2、安装DeepSpeed,这部分更多细节请参照DeepSpeed项目安装说明。

pip install deepspeed

3、本项目也需要TDS方可运行,此部分更多细节可以见TDS。TDS插件只需要将代码拷贝到当前项目的根目录下,再使用下列代码载入DeepSpeed即可(当前项目里此步骤已经完成)。

import tds as deepspeed

模型

当前项目主要支持GPT-2、T5两类模型,未来会有更多模型放出。

预训练与设置脚本

1、多机多卡设置

#每台机器的显卡数量
GPUS_PER_NODE=4

#支持机器之间通信的IP地址
MASTER_ADDR=localhost

#支持机器之间通信的端口
MASTER_PORT=8888

#本次训练一共涉及几台机器
NNODES=1

#当前机器的序号,编号从0开始
NODE_RANK=0

2、数据并行、模型并行、流水并行设置

# 整体模型划分为几层,进行流水并行
pp_size=2

# 每层模型划分为几块,进行模型并行
mp_size=2

3、设置模型尺寸

#层数
NLAYERS=2

#输入层大小
NHIDDEN=128

#单张卡每次接触到的batch大小
BATCHSIZE=4

#梯度累积步数
GAS=16

4、DeepSpeed脚本设置,为DeepSpeed设置训练的流水并行、混合精度等细节,只有在流水并行被启用时,此脚本方被使用。如果pp_size=0或pp_size=1,意味着流水并行不被激活,此时无论脚本里的设置为何,均不会发生作用,代码只会使用Megatron-LM进行模型并行。

config_json="ds_config_t5.json"

脚本形式:
{
    "train_batch_size": 128, 
    "train_micro_batch_size_per_gpu": 4, 
    "steps_per_print": 10, 
    "gradient_clipping": 1.0, 
    "fp16": {
        "enabled": true, 
        "loss_scale": 0, 
        "loss_scale_window": 1000, 
        "hysteresis": 2, 
        "initial_scale_power": 16
    }, 
    "zero_allow_untested_optimizer": true, 
    "wall_clock_breakdown": true
}

注意:脚本里的 train_batch_size = BATCHSIZE * GAS * GPUS_PER_NODE * NNODES / pp_size / mp_size

5、也可以自己自由设置训练参数,以run_t5.sh为例

    --model-parallel-size 2 \
    --pipe-parallel-size 2 \
    --num-layers 12\
    --hidden-size 128 \
    --kv-hidden-size 16 \
    --ff-hidden-size 256 \
    --num-attention-heads 8 \
    --enc-seq-length 1024\
    --dec-seq-length 384\
    --max-position-embeddings 1024 \
    --batch-size 4\
    --gas 16 \
    --lr 1.5e-4 \
    --lr-decay-style cosine \
    --min-lr 1.0e-5 \
    --weight-decay 1e-2 \
    --clip-grad 1.0 \
    --warmup 0.01 \

6、运行脚本

bash run_t5.sh 或 bash run_gpt2.sh

实现自己的模型

我们以pretrain_gpt2.py为例子,介绍下如何实现自己的模型

  • 使用Megatron-LM实现模型并行

    Megatron-LM的较为核心的代码是 megatron/mpu/layers.py,这里面定义了

    • VocabParallelEmbedding旨在将embedding层切割到不同的GPU上,
    • ColumnParallelLinear、RowParallelLinear则是将矩阵按照行列切割到不同的GPU上,完整的输出结果需要通过一次all-reduce来从各个GPU上汇总。

    这里我们以GPT-2中的Transformer的线性层来举例。Transformer的线性层分为两部分:

    • 第一部分是将$\text{hidden_size}$维的向量映射为$4 \times \text{hidden_size}$维的向量,并且通过GeLU激活;
    • 第二部分是将激活后的$4 \times \text{hidden_size}$维的向量映射回$\text{hidden_size}$维,并且施加Dropout。

    这里,我们对第一部分的线性层矩阵按列切分,第二部分按行切分。

    • 对于第一部分$\mathbf{Y}=\text{GeLU}(\mathbf{X}\mathbf{A})$,我们将矩阵$\mathbf{A}$切分为$\mathbf{A}_1,\mathbf{A}_2$两个部分,那么整体的计算也被相应分为两部分$\mathbf{Y}_1=\text{GeLU}(\mathbf{X}\mathbf{A}_1)$和$\mathbf{Y}_2=\text{GeLU}(\mathbf{X}\mathbf{A}_2)$。此时$\mathbf{Y}_1$和$\mathbf{Y}_2$分别为$\mathbf{Y}$的前$2 \times \text{hidden_size}$结果和后$2 \times \text{hidden_size}$结果。

    • 对于$\mathbf{Z}=\mathbf{Y}\mathbf{B}$,我们将矩阵$\mathbf{B}$按行切分为$\mathbf{B}_1, \mathbf{B}_2$。将第一部分的结果$\mathbf{Y}_1$和$\mathbf{Y}_2$传入后,可计算$\mathbf{Z}_1=\mathbf{Y}_1\mathbf{B}_1$与$\mathbf{Z}_1=\mathbf{Y}_2\mathbf{B}_2$,此时有$\mathbf{Z}=\mathbf{Z}_1+\mathbf{Z}_1$。在Dropout之前,我们需要对$\mathbf{Z}_1$与$\mathbf{Z}_2$进行一次同步,加和后的结果再过Dropout。

    总的来看,整个线性层计算过程中,每张显卡都只负责了一半的计算量,整体结果需要一次额外的all_reduce来对输出结果进行汇总。这里也给出具体的代码实现,与我们上述的过程是一致的。此部分代码可见 megatron/model/transformer.py

    class ParallelMLP(MegatronModule):
    
        def __init__(self, ...):
            super(ParallelMLP, self).__init__()
            args = get_args()
            # Project to 4h.
            self.dense_h_to_4h = mpu.ColumnParallelLinear(
                args.hidden_size,
                args.hidden_size * 4,...)
                
            ...
            # Project back to h.
            self.dense_4h_to_h = mpu.RowParallelLinear(
                args.args.hidden_size * 4,
                args.hidden_size,...)
    
        def forward(self, hidden_states):
            # [s, b, 4hp]
            intermediate_parallel, bias_parallel = self.dense_h_to_4h(hidden_states)
            ...
            # [s, b, h]
            output, output_bias = self.dense_4h_to_h(intermediate_parallel)
            return output, output_bias

    同样的方法我们也可以实现Attention层、Embedding层的模型并行。

  • 使用DeepSpeed实现流水并行

    使用DeepSpeed进行流水并行也并不复杂,只需要将模型的各个层加入队列中,并用队列实例化DeepSpeed的PipelineModule类即可。

    但是受限于DeepSpeed框架本身的设计,队列中的各个层输入的tensor数量和输出的tensor数量必须保持一致,以确保各层之间可以无缝衔接。

    例如实现一个流水的Transformer版本,模型的每一层都需要输出hidden_states和mask,与下一层的输入相衔接。

    from deepspeed.pipe import PipelineModule, LayerSpec
    ...
    class TransformerBlockPipe(TransformerBlock)
        def forward(self, ...):
            hidden, mask = ...
            ...
            output = super().forward(hidden, mask)
            return (output, mask)
    ...
    
    class GPT2(PipelineModule):
        def __init__(self, num_layers, ...):
            ...
            specs = [TransformerBlockPipe() for _ in range(num_layers) ]
            ...
            super().__init__(layers=specs, ...)

    上述代码有一个LayerSpec的实现。LayerSpec有点类似一层懒惰标记的封装。

    具体而言,LayerSpec的参数分别为神经网络层的class以及class的初始化参数。LayerSpec将class和初始化参数记忆下来,但不会立刻实例化,而是在流水线构建过程中再将class和初始化参数结合起来进行实例化。

    使用DeepSpeed进行混合加速也十分方便,只需要让加入流水队列里的模型实现是基于模型并行的即可,如使用Megatron-LM来实现上述代码的TransformerBlockPipe和TransformerBlock。

  • 使用TDS实现训练流程

    TDS重新实现了Deepspeed的流水线,通过适配器模式把DeepSpeed的其他功能封装进来,所以用起来也非常简单。在安装完DeepSpeed之后,只需要将TDS的代码拷贝到工程中,然后用以下的方式加载库即可。

    import tds as deepspeed

    使用TDS的训练流程代码与之前的DeepSpeed和Megatron中的范例代码是一致的,首先需要一个获取数据iterator的模块。

    def train_valid_test_datasets_provider(...):
        ...
        train_ds, valid_ds, test_ds = build_train_valid_test_datasets(...)
        ...
        return train_ds, valid_ds, test_ds

    然后,需要一个从data_iterator中获取数据,并且分发到各个GPU上的模块。

    def get_batch(data_iterator):
        ...
        # 定义传输的数据类型
        keys = ['text']
        datatype = torch.int64
        # 将数据压缩之后广播到同一个数据并行group中的每个GPU上
        if data_iterator is not None:
            data = next(data_iterator)
        else:
            data = None
        data_b = mpu.broadcast_data(keys, data, datatype)
        # 将数据解压
        tokens_ = data_b['text'].long()
        # 从数据中提取token, label, mask信息
        tokens, labels, loss_mask, attention_mask, position_ids = (...)
        ...
        return tokens, labels, loss_mask, attention_mask, position_ids

    流水并行需要一个单独的数据获取分发模块,整体结构与get_batch是类似的,但有两个细微差别。一个是DeepSpeed中将流水并行的data_iterator交给了后台管理,所以get_batch_pipe只需要负责分发数据即可。第二个是DeepSpeed对给流水线使用的data_iterator是有格式限制的,必须是返回两个tuple,前一个是用来进行forward的输入,后一个是用来进行loss的计算的输入。

    def get_batch_pipe(data):
        ...
        # 定义传输的数据类型
        keys = ['text']
        datatype = torch.int64
        # 将数据压缩之后广播到同一个数据并行group中的每个GPU上
        data_b = mpu.broadcast_data(keys, data, datatype)
        # 将数据解压
        tokens_ = data_b['text'].long()
        # 从数据中提取token, label, mask信息
        tokens, labels, loss_mask, attention_mask, position_ids = (...)
        ...
        return (tokens, position_ids, attention_mask), (labels, loss_mask)

    在实现了数据获取的各项模块后,模型获取模块就十分简单,根据是否使用流水并行返回不同模型。与DeepSpeed相比,TDS中需要制定流水线中间的输入输出类型,是否需要保存梯度,以及是否需要将中间结果切割到多个GPU来减少显存使用。

    def model_provider():
        """Build the model."""
        args = get_args()
        if args.pipe_parallel_size == 0 :
            model = GPT2Model(...)
        else:
            model = GPT2ModelPipe(...)
            model._megatron_batch_fn = get_batch_pipe
            # TDS中需要制定流水线中间的输入输出类型,是否需要保存梯度,以及是否需要将中间结果切割到多个GPU来减少显存使用
            model._input_grad = [True, False]
            model._input_type = ['float', 'bool']
            model._input_pipe_partitioned = [True, False]
        return model

    最后只要基于上述模块,启动pretrain就能开始计算。

    from megatron.training import pretrain
    
    if __name__ == "__main__":
        pretrain(train_valid_test_datasets_provider, model_provider, ...)

引用

如果您使用了我们的代码,请您引用下面的文章。

@article{cpm-v1,
  title={CPM: A Large-scale Generative Chinese Pre-trained Language Model},
  author={Zhang, Zhengyan and Han, Xu, and Zhou, Hao, and Ke, Pei, and Gu, Yuxian and Ye, Deming and Qin, Yujia and Su, Yusheng and Ji, Haozhe and Guan, Jian and Qi, Fanchao and Wang, Xiaozhi and Zheng, Yanan and Zeng, Guoyang and Cao, Huanqi and Chen, Shengqi and Li, Daixuan and Sun, Zhenbo and Liu, Zhiyuan and Huang, Minlie and Han, Wentao and Tang, Jie and Li, Juanzi and Sun, Maosong},
  year={2020}
}
Owner
Tsinghua AI
Tsinghua AI
Python wrapper for Stanford CoreNLP tools v3.4.1

Python interface to Stanford Core NLP tools v3.4.1 This is a Python wrapper for Stanford University's NLP group's Java-based CoreNLP tools. It can eit

Dustin Smith 610 Sep 07, 2022
Code for paper "Role-oriented Network Embedding Based on Adversarial Learning between Higher-order and Local Features"

Role-oriented Network Embedding Based on Adversarial Learning between Higher-order and Local Features Train python main.py --dataset brazil-flights C

wang zhang 0 Jun 28, 2022
AudioCLIP Extending CLIP to Image, Text and Audio

AudioCLIP Extending CLIP to Image, Text and Audio This repository contains implementation of the models described in the paper arXiv:2106.13043. This

458 Jan 02, 2023
BMInf (Big Model Inference) is a low-resource inference package for large-scale pretrained language models (PLMs).

BMInf (Big Model Inference) is a low-resource inference package for large-scale pretrained language models (PLMs).

OpenBMB 377 Jan 02, 2023
PhoNLP: A BERT-based multi-task learning toolkit for part-of-speech tagging, named entity recognition and dependency parsing

PhoNLP is a multi-task learning model for joint part-of-speech (POS) tagging, named entity recognition (NER) and dependency parsing. Experiments on Vietnamese benchmark datasets show that PhoNLP prod

VinAI Research 109 Dec 02, 2022
Twitter-Sentiment-Analysis - Twitter sentiment analysis for india's top online retailers(2019 to 2022)

Twitter-Sentiment-Analysis Twitter sentiment analysis for india's top online retailers(2019 to 2022) Project Overview : Sentiment Analysis helps us to

Balaji R 1 Jan 01, 2022
Implementation of some unbalanced loss like focal_loss, dice_loss, DSC Loss, GHM Loss et.al

Implementation of some unbalanced loss for NLP task like focal_loss, dice_loss, DSC Loss, GHM Loss et.al Summary Here is a loss implementation reposit

121 Jan 01, 2023
CoSENT 比Sentence-BERT更有效的句向量方案

CoSENT 比Sentence-BERT更有效的句向量方案

苏剑林(Jianlin Su) 201 Dec 12, 2022
Text classification on IMDB dataset using Keras and Bi-LSTM network

Text classification on IMDB dataset using Keras and Bi-LSTM Text classification on IMDB dataset using Keras and Bi-LSTM network. Usage python3 main.py

Hamza Rashid 2 Sep 27, 2022
SentAugment is a data augmentation technique for semi-supervised learning in NLP.

SentAugment SentAugment is a data augmentation technique for semi-supervised learning in NLP. It uses state-of-the-art sentence embeddings to structur

Meta Research 363 Dec 30, 2022
Chatbot for the Chatango messaging platform

BroiestBot The baddest bot in the game right now. Uses the ch.py framework for joining Chantango rooms and responding to user messages. Commands If a

Todd Birchard 3 Jan 17, 2022
An open source framework for seq2seq models in PyTorch.

pytorch-seq2seq Documentation This is a framework for sequence-to-sequence (seq2seq) models implemented in PyTorch. The framework has modularized and

International Business Machines 1.4k Jan 02, 2023
Gold standard corpus annotated with verb-preverb connections for Hungarian.

Hungarian Preverb Corpus A gold standard corpus manually annotated with verb-preverb connections for Hungarian. corpus The corpus consist of the follo

RIL Lexical Knowledge Representation Research Group 3 Jan 27, 2022
Weird Sort-and-Compress Thing

Weird Sort-and-Compress Thing A weird integer sorting + compression algorithm inspired by a conversation with Luthingx (it probably already exists by

Douglas 1 Jan 03, 2022
A Python wrapper for simple offline real-time dictation (speech-to-text) and speaker-recognition using Vosk.

Simple-Vosk A Python wrapper for simple offline real-time dictation (speech-to-text) and speaker-recognition using Vosk. Check out the official Vosk G

2 Jun 19, 2022
Shellcode antivirus evasion framework

Schrodinger's Cat Schrodinger'sCat is a Shellcode antivirus evasion framework Technical principle Please visit my blog https://idiotc4t.com/ How to us

idiotc4t 27 Jul 09, 2022
Simple translation demo showcasing our headliner package.

Headliner Demo This is a demo showcasing our Headliner package. In particular, we trained a simple seq2seq model on an English-German dataset. We didn

Axel Springer News Media & Tech GmbH & Co. KG - Ideas Engineering 16 Nov 24, 2022
Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning

GenSen Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning Sandeep Subramanian, Adam Trischler, Yoshua B

Maluuba Inc. 309 Oct 19, 2022
Data and evaluation code for the paper WikiNEuRal: Combined Neural and Knowledge-based Silver Data Creation for Multilingual NER (EMNLP 2021).

Data and evaluation code for the paper WikiNEuRal: Combined Neural and Knowledge-based Silver Data Creation for Multilingual NER. @inproceedings{tedes

Babelscape 40 Dec 11, 2022
A raytrace framework using taichi language

ti-raytrace The code use Taichi programming language Current implement acceleration lvbh disney brdf How to run First config your anaconda workspace,

蕉太狼 73 Dec 11, 2022