Giters

OpenBMB / ModelCenter

Efficient, Low-Resource, Distributed transformer implementation based on BMTrain

Home Page:https://modelcenter.readthedocs.io

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ModelCenter

Efficient Low-Resource Implementations of Big Models

Overview • Documentation • Installation • Quick Start • Supported Models • 简体中文

Documentation Status GitHub release (latest by date including pre-releases) GitHub

What's New

Overview

ModelCenter implements pre-trained language models (PLMs) based on the backend OpenBMB/BMTrain. ModelCenter supports Efficient, Low-Resource, Extendable model usage and distributed training.

Our main advantages are:

  • Easy to use. Compared to Deepspeed and Megatron, we have better and more flexible code-packaging and easy to configure python environments, and the training code is uniform with PyTorch style.
  • More efficient memory utilization. Models with large memory footprints can cause OOM (out of memory) before the computational power of the GPU is fully utilized. Our implementation reduces the memory footprint by several times, allowing more efficient use of the GPU's computational power with a larger batch size.
  • Efficient distributed training with low resources. With the support of OpenBMB/BMTrain, we are able to easily extend the ZeRO optimization to any PLMs, and we optimize communication and time scheduling for faster distributed training.

Documentation

Our documentation provides more information about the package.

Installation

1. From PyPI (Recommend)

$ pip install model-center

2. From Source

$ git clone https://github.com/OpenBMB/ModelCenter.git
$ cd ModelCenter
$ pip install -r requirements.txt
$ python3 setup.py install

Quick Start

In the quick start, you will walk through how to fine-tune a BERT model on a classification task.

1. Initialize bmtrain backend

First, you need to import bmtrain and use bmtrain.init_distributed() at the beginning of your code, which can initialize the distributed environments.

import bmtrain as bmt
bmt.init_distributed(seed=0)

2. Prepare the model

Next, you can simply get a pre-trained BERT model from model_center, e.g., bert-base-uncased. When fine-tuning BERT on the classification task, a feed-forward layer need to be appended to the last layer.

import torch
from model_center.model import Bert, BertConfig
from model_center.layer import Linear

class BertModel(torch.nn.Module):
    def __init__(self, config):
        super().__init__()
        self.bert = Bert.from_pretrained("bert-base-uncased")
        self.dense = Linear(config.dim_model, 2)
        bmt.init_parameters(self.dense)

    def forward(self, input_ids, attention_mask):
        pooler_output = self.bert(input_ids=input_ids, attention_mask=attention_mask).pooler_output
        logits = self.dense(pooler_output)
        return logits

config = BertConfig.from_pretrained("bert-base-uncased")
model = BertModel(config)

If only config is needed instead of pretrained checkpoint, you can initialize a model as the following:

config = BertConfig.from_json_file("your/path/to/config.json")
model = Bert(config)
bmt.init_parameters(model)
# bmt.load(model, "your/path/to/pytorch_model.pt")

3. Perpare the dataset

The next step is to prepare the dataset used for training and evaluation. Here, we use the BoolQ dataset from the SuperGLUE benchmark. You need to download the dataset and put the unzipped folder in your_path_to_dataset.

from model_center.dataset.bertdataset import DATASET
from model_center.dataset import DistributedDataLoader
from model_center.tokenizer import BertTokenizer

tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
splits = ['train', 'dev']
dataset = {}

for split in splits:
    dataset[split] = DATASET['BoolQ']('your_path_to_dataset', split, bmt.rank(), bmt.world_size(), tokenizer, max_encoder_length=512)

batch_size = 64
train_dataloader = DistributedDataLoader(dataset['train'], batch_size=batch_size, shuffle=True)
dev_dataloader = DistributedDataLoader(dataset['dev'], batch_size=batch_size, shuffle=False)

4. Train the model

Now, select optimizer, learning rate scheduler, loss function, and then start training the model! Here, we train BERT for 5 epochs and evaluate it at the end of each epoch.

optimizer = bmt.optim.AdamOffloadOptimizer(model.parameters())

lr_scheduler = bmt.lr_scheduler.Noam(
    optimizer, 
    start_lr = 1e-5,
    warmup_iter = 100, 
    end_iter = -1)

loss_func = bmt.loss.FusedCrossEntropy(ignore_index=-100)

optim_manager = bmt.optim.OptimManager(loss_scale=1024)
optim_manager.add_optimizer(optimizer, lr_scheduler)

for epoch in range(5):
    model.train()
    for data in train_dataloader:
        input_ids = data['input_ids']
        attention_mask = data['attention_mask']
        labels = data['labels']

        # model forward
        logits = model(input_ids, attention_mask)

        # calculate loss
        loss = loss_func(logits.view(-1, logits.shape[-1]), labels.view(-1))

        # use bmt.sum_loss(loss) to gather all loss information from all distributed processes
        global_loss = bmt.sum_loss(loss).item()

        # zero grad
        optim_manager.zero_grad()

        # scale loss before backward to avoid precision underflow of fp16
        optim_manager.backward(loss)

        # clip gradient norm
        grad_norm = optim_manager.clip_grad_norm(optimizer.param_groups, max_norm=10.0, scale = optimizer.scale, norm_type = 2)

        # step for all optimizer inside optim_manager
        optim_manager.step()

        # print information only on rank 0 when distributed training
        bmt.print_rank(
            "loss: {:.4f} | lr: {:.4e}, scale: {:10.4f} | grad_norm: {:.4f} |".format(
                global_loss,
                lr_scheduler.current_lr,
                int(optimizer.scale),
                grad_norm,
            )
        )

    # evaluate model
    model.eval()
    with torch.no_grad():
        pd = [] # prediction
        gt = [] # ground_truth
        for data in dev_dataloader:
            input_ids = data["input_ids"]
            attention_mask = data["attention_mask"]
            labels = data["labels"]

            logits = model(input_ids, attention_mask)
            loss = loss_func(logits.view(-1, logits.shape[-1]), labels.view(-1))

            logits = logits.argmax(dim=-1)

            pd.extend(logits.cpu().tolist())
            gt.extend(labels.cpu().tolist())

        # gather results from all distributed processes
        pd = bmt.gather_result(torch.tensor(pd).int()).cpu().tolist()
        gt = bmt.gather_result(torch.tensor(gt).int()).cpu().tolist()

        # calculate metric
        from sklearn.metrics import accuracy_score
        acc = accuracy_score(gt, pd)
        bmt.print_rank(f"accuracy: {acc*100:.2f}")

5. Run your code

You can run the above code using the same launch command as the distributed module of PyTorch.

Choose one of the following commands depending on your version of PyTorch.

  • ${MASTER_ADDR} means the IP address of the master node.
  • ${MASTER_PORT} means the port of the master node.
  • ${NNODES} means the total number of nodes.
  • ${GPU_PER_NODE} means the number of GPUs per node.
  • ${NODE_RANK} means the rank of this node.

torch.distributed.launch (more suitable for torch < 1.10)

$ python3 -m torch.distributed.launch --master_addr ${MASTER_ADDR} \
                                      --master_port ${MASTER_PORT} \
                                      --nproc_per_node ${GPU_PER_NODE} \
                                      --nnodes ${NNODES} \
                                      --node_rank ${NODE_RANK} \
                                      train.py

torchrun (more suitable for torch >= 1.10)

$ torchrun --nnodes=${NNODES} \
           --nproc_per_node=${GPU_PER_NODE} \
           --rdzv_id=1 \
           --rdzv_backend=c10d \
           --rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} \
           train.py

More information can be found from the documentation.

Supported Models

  • CPM-1[paper]. We currently support loading the following checkpoint via CPM1.from_pretrained(identifier) as follows,

    • cpm1-large

  • CPM-2[paper]. We currently support loading the following checkpoint via CPM2.from_pretrained(identifier) as follows,

    • cpm2-large

  • BERT[paper]. We currently support loading the following checkpoint via Bert.from_pretrained(identifier) as follows,

    • bert-base-cased
    • bert-base-uncased
    • bert-large-cased
    • bert-large-uncased
    • bert-base-chinese
    • bert-base-multilingual-cased
    • kv-plm

  • RoBERTa[paper]. We currently support loading the following checkpoint via Roberta.from_pretrained(identifier) of the following:

    • roberta-base
    • roberta-large

  • T5[paper]. We currently support loading the following checkpoint via T5.from_pretrained(identifier) of the following:

    • t5-small
    • t5-base
    • t5-large
    • t5-3b
    • t5-11b
    • t5-v1_1-small
    • t5-v1_1-base
    • t5-v1_1-large
    • t5-v1_1-xl
    • t5-v1_1-xxl
    • mt5-small
    • mt5-base
    • mt5-large
    • mt5-xl
    • mt5-xxl
    • mengzi-t5-base
    • flan-t5-small
    • flan-t5-base
    • flan-t5-large
    • flan-t5-xl
    • flan-t5-xxl

  • GPT-2[paper]. We currently support loading the following checkpoint via GPT2.from_pretrained(identifier) of the following:

    • gpt2-base
    • gpt2-medium
    • gpt2-large
    • gpt2-xl
    • wenzhong-gpt2-3.5b

  • GPT-J[paper]. We currently support loading the following checkpoint via GPTj.from_pretrained(identifier) of the following:

    • gptj-6b

  • Longformer[paper]. we currently support loading the following checkpoint via Longformer.from_pretrained(identifier) of the following:

    • lawformer

  • GLM[paper]. we currently support loading the following checkpoint via GLM.from_pretrained(identifier) of the following:

    • glm-10b-zh

  • ViT[paper]. we currently support loading the following checkpoint via ViT.from_pretrained(identifier) of the following:

    • vit-base-patch16-224

  • LLaMA[paper]. convert checkpoint via transfer/hugLLaMa_bmtrainLLaMa.py.

Performance

You can find more performance metrics in the repo OpenBMB/BMTrain.

Community

We welcome everyone to contribute codes following our contributing guidelines.

You can also find us on other platforms:

License

The package is released under the Apache 2.0 License.

About

Efficient, Low-Resource, Distributed transformer implementation based on BMTrain

https://modelcenter.readthedocs.io

License:Apache License 2.0

Languages

Language:Python 99.7%Language:Shell 0.3%