Efficient Training for Big Models
Overview • Documentation • Installation • Usage • Performance • 简体中文
- 2022/12/15 BMTrain 0.2.0 released. See the update log.
- 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.
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.
Our documentation provides more information about the package.
-
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.
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,
zero_level=3, # support 2 and 3 now
# ...
)NOTE: Do not use PyTorch's distributed module and its associated communication functions when using BMTrain.
To enable ZeRO optimization, you need to make some simple replacements to the original model's code.
torch.nn.Module->bmtrain.DistributedModuletorch.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 xReplaced
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
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.TransformerBlockListfor 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
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.
$ 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 --nnodes=${NNODES} --nproc_per_node=${GPU_PER_NODE} --rdzv_id=1 --rdzv_backend=c10d --rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} train.pyFor more information, please refer to the documentation.
We provide an example of training GPT-2 based on BMTrain. The code mainly consists of the following parts.
├── 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 optimization.
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.
loss_func = torch.nn.CrossEntropyLoss(ignore_index=-100)
optimizer = bmtrain.optim.AdamOffloadOptimizer(model.parameters(), weight_decay=1e-2)
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.
# create a new instance of optimizer manager
op
9299
tim_manager = bmtrain.optim.OptimManager(loss_scale=1024)
# let optim_manager handle all the optimizer and (optional) their corresponding lr_scheduler
optim_manager.add_optimizer(optimizer, lr_scheduler)
# add_optimizer can be called multiple times to add other optimizers.
for iteration in range(1000):
# ... load data for each rank ...
# forward pass and calculate loss
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. This is only used for the training log
# zero grad
optim_manager.zero_grad() # calling zero_grad for each optimizer
# loss scale and backward
optim_manager.backward(loss)
# clip grad norm
grad_norm = optim_manager.clip_grad_norm(optimizer.param_groups, max_norm=1.0)
# optimizer step
optim_manager.step()
# ... 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. After using BMTrain, some details in optimizers should be adjusted. We have implemented all those details needed in optim_manager. What you need is just letting optim_manager to handle all the optimizers by add_optimizer, and letting optim_manager do zero_grad(), backward(), clip_grad_norm() and step() instead.
If you are not using the mixed-precision training, you can train without loss_scale. Just set loss_scale to None in the __init__ function of OptimManager(loss_scale=None), which is also the default.
If you are using mixed-precision training, loss scale is the technique widely used in mixed precision training to prevent gradient underflow. By using optim_manager.backward(loss) to scale the loss before backward and set loss_scale to some floating number in the __init__ function of OptimManager。The loss_scale would be adjusted adaptively based on the gradient during training.
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.
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.
We welcome everyone to contribute codes following our contributing guidelines.
You can also find us on other platforms:
- QQ Group: 735930538
- Website: https://www.openbmb.org
- Weibo: http://weibo.cn/OpenBMB
- Twitter: https://twitter.com/OpenBMB
The package is released under the Apache 2.0 License.
BMTrain makes underlying changes to PyTorch, so if your program outputs unexpected results, you can submit information about it in an issue.