You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 
zgy b9c66b35f2 patch 0.1.7 3 weeks ago
.github/workflows workflow use pytorch cpu to build 3 months ago
bmtrain FIX: release the parameter at some special case 3 weeks ago
csrc FX: has_inf_or_nan cuda invalid configuration argument 2 months ago
docs update logo 2 months ago
example zero2 1 month ago
tests FX: eval not working in Checkpoint Block 2 months ago
.dockerignore rename distributed 3 months ago
.gitignore FX: gitignore 3 months ago
CONTRIBUTING.md update docs and readme 3 months ago
Dockerfile FX: has_inf_or_nan cuda invalid configuration argument 2 months ago
LICENSE add license and github workflow 3 months ago
MANIFEST.in update setups 3 months ago
README-ZH.md update readme 3 weeks ago
README.md update readme 3 weeks ago
doc_requirements.txt FX: autodoc libc10 3 months ago
other_requirements.txt rename distributed 3 months ago
setup.py patch 0.1.7 3 weeks ago

README.md

BMTrain

Efficient Training for Big Models

OverviewDocumentationInstallationUsagePerformance简体中文

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

What’s New

  • 2022/06/14 BMTrain 0.1.7 released. ZeRO-2 optimization is supported!
  • 2022/03/30 BMTrain 0.1.2 released. Adapted to OpenPromptand OpenDelta.
  • 2022/03/16 BMTrain 0.1.1 has publicly released the first stable version, which fixes many bugs that were in the beta version.
  • 2022/02/11 BMTrain 0.0.15 has publicly released the first beta version.

Overview

BMTrain is an efficient large model training toolkit that can be used to train large models with tens of billions of parameters. It can train models in a distributed manner while keeping the code as simple as stand-alone training.

Documentation

Our documentation provides more information about the package.

Installation

  • From pip (recommend) : pip install bmtrain

  • From source code: download the package and run python setup.py install

Installing BMTrain may take a few to ten minutes, as it requires compiling the c/cuda source code at the time of installation.
We recommend compiling BMTrain directly in the training environment to avoid potential problems caused by the different environments.

Usage

Step 1: Initialize BMTrain

Before you can use BMTrain, you need to initialize it at the beginning of your code. Just like using the distributed module of PyTorch requires the use of init_process_group at the beginning of the code, using BMTrain requires the use of init_distributed at the beginning of the code.

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

NOTE: Do not use PyTorch’s distributed module and its associated communication functions when using BMTrain.

Step 2: Enable ZeRO-3 Optimization

To enable ZeRO-3 optimization, you need to make some simple replacements to the original model’s code.

  • torch.nn.Module -> bmtrain.DistributedModule
  • torch.nn.Parameter -> bmtrain.DistributedParameter

And wrap the transformer blocks with bmtrain.CheckpointBlock.

Here is an example.

Original

import torch
class MyModule(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.param = torch.nn.Parameter(torch.empty(1024))
        self.module_list = torch.nn.ModuleList([
            SomeTransformerBlock(),
            SomeTransformerBlock(),
            SomeTransformerBlock()
        ])
    
    def forward(self):
        x = self.param
        for module in self.module_list:
            x = module(x, 1, 2, 3)
        return x

Replaced

import torch
import bmtrain as bmt
class MyModule(bmt.DistributedModule): # changed here
    def __init__(self):
        super().__init__()
        self.param = bmt.DistributedParameter(torch.empty(1024)) # changed here
        self.module_list = torch.nn.ModuleList([
            bmt.CheckpointBlock(SomeTransformerBlock()), # changed here
            bmt.CheckpointBlock(SomeTransformerBlock()), # changed here
            bmt.CheckpointBlock(SomeTransformerBlock())  # changed here
        ])
    
    def forward(self):
        x = self.param
        for module in self.module_list:
            x = module(x, 1, 2, 3)
        return x
    

Step 3: Enable Communication Optimization

To further reduce the extra overhead of communication and overlap communication with computing time, TransformerBlockList can be used for optimization.

You can enable them by making the following substitutions to the code:

  • torch.nn.ModuleList -> bmtrain.TransformerBlockList
  • for module in self.module_list: x = module(x, ...) -> x = self.module_list(x, ...)

Original

import torch
import bmtrain as bmt
class MyModule(bmt.DistributedModule):
    def __init__(self):
        super().__init__()
        self.param = bmt.DistributedParameter(torch.empty(1024))
        self.module_list = torch.nn.ModuleList([
            bmt.CheckpointBlock(SomeTransformerBlock()),
            bmt.CheckpointBlock(SomeTransformerBlock()),
            bmt.CheckpointBlock(SomeTransformerBlock())
        ])
    
    def forward(self):
        x = self.param
        for module in self.module_list:
            x = module(x, 1, 2, 3)
        return x
    

Replaced

import torch
import bmtrain as bmt
class MyModule(bmt.DistributedModule):
    def __init__(self):
        super().__init__()
        self.param = bmt.DistributedParameter(torch.empty(1024))
        self.module_list = bmt.TransformerBlockList([ # changed here
            bmt.CheckpointBlock(SomeTransformerBlock()),
            bmt.CheckpointBlock(SomeTransformerBlock()),
            bmt.CheckpointBlock(SomeTransformerBlock())
        ])
    
    def forward(self):
        x = self.param
        x = self.module_list(x, 1, 2, 3) # changed here
        return x
    

Step 4: Launch Distributed Training

BMTrain uses the same launch command as the distributed module of PyTorch.

You can choose one of them 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

$ 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

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

For more information, please refer to the documentation.

Example

We provide an example of training GPT-2 based on BMTrain.
The code mainly consists of the following parts.

Part 1: Model Definition

├── layers
│   ├── attention.py
│   ├── embedding.py
│   ├── feedforward.py
│   ├── __init__.py
│   ├── layernorm.py
│   └── linear.py
└── models
    ├── gpt.py
    └── __init__.py

Above is the directory structure of the code in the part of Model Definition.

We defined all the layers needed in GPT-2 and used BMTrain’s DistributedModule and DistributedParameter to enable ZeRO-3 optimization.

Part 2: BMTrain Initialization

bmtrain.init_distributed(seed=0)

model = GPT(
    num_layers=8,
    vocab_size=10240, 
    dim_model=2560,
    dim_head=80,
    num_heads=32,
    dim_ff=8192,
    max_distance=1024,
    bias=True,
    dtype=torch.half
)

bmtrain.init_parameters(model) # or loading checkpoint use `bmtrain.load`

# ... other initialization (dataset) ...

bmtrain.init_distributed(seed=0) is used to initialize the distributed training environment and set the random seed for reproducibility.

bmtrain.init_parameters(model) is used to initialize the distributed parameters of the model.

Part 3: Intialization of the Optimizer and LR Scheduler

loss_func = torch.nn.CrossEntropyLoss(ignore_index=-100)
optimizer = bmtrain.optim.AdamOffloadOptimizer(model.parameters(), weight_decay=1e-2, scale=2**20)
lr_scheduler = bmtrain.lr_scheduler.Noam(optimizer, start_lr=1e-3, warmup_iter=40, end_iter=1000, num_iter=0)

BMTrain supports all the PyTorch native optimizers and loss functions, and you can also use the fused optimizer provided by BMTrain for mixed-precision training.

In addition, BMTrain also provides the common LRScheduler in the bmtrain.lr_scheduler module.

Part 4: Training Loop

for iteration in range(1000):
    # ... load data for each rank ...

    # zero grad
    optimizer.zero_grad()

    # forward
    pos = torch.arange(enc_input.size(1)).long().cuda().repeat(enc_input.size(0), 1)
    logits = model(
        enc_input,
        pos,
        pos < enc_length[:, None]
    )
    batch, seq_len, vocab_out_size = logits.size()

    loss = loss_func(logits.view(batch * seq_len, vocab_out_size), targets.view(batch * seq_len))
    
    global_loss = bmtrain.sum_loss(loss).item() # sum the loss across all ranks

    # loss scale and backward
    loss = optimizer.loss_scale(loss)
    loss.backward()

    # optimizer step
    bmtrain.optim_step(optimizer, lr_scheduler)

    # ... save checkpoint or print logs ...

The training loop part will be slightly longer, but just like a normal training loop, you don’t need to adapt much to distributed training.

You can follow the comments in the code to get an idea of what each section of code is doing.

The only additional note is optimizer.loss_scale, loss scale is the technique that widely used in mixed precision training to prevent gradient underflow. If you are not using the fused optimizer in BMTrain, you can remove this statement.

Performance

We trained a GPT-2 model with 13B parameters using 4 servers with 8 V100s on each server, and measured the throughput of each GPU during the training process (samples per GPU per second).

Model structure:

  • 40 layers
  • 128 attention heads
  • 5120 hidden dimension
  • 512 sequence length
batch size 8 16 24 32
BMTrain 24.15 26.94 29.42 28.28
ZeRO3(mp=1) 14.88 21.69 24.38 -
ZeRO3(mp=4) 15.51 - - -
ZeRO3(mp=8) 15.51 - - -
ZeRO2(mp=1) - - - -
ZeRO2(mp=4) 22.85 - - -
ZeRO2(mp=8) 21.33 - - -

ZeROa(mp=b) means DeepSpeed + Megatron ZeRO stage a and model parallelism = b.

- in the table means out of memory.

Supported Models

We have migrated most of the common models in NLP to the BMTrain. You can find the list of supported models in the repo ModelCenter.

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.

Other Notes

BMTrain makes underlying changes to PyTorch, so if your program outputs unexpected results, you can submit information about it in an issue.