ColossalAI/docs/amp.md

Mixed Precision Training

In Colossal-AI, we have integrated different implementations of mixed precision training:

1. torch.cuda.amp
2. apex.amp
3. tensor-parallel amp

The first two rely on the original implementation of PyTorch (version 1.6 and above) and Nvidia Apex. However, these two methods are not compatible with tensor parallelism. This is because that tensors are split across devices in tensor parallelism, thus, it is needed to communicate among different processes to check if inf or nan occurs throughout the whole model weights. For the mixed precision training with tensor parallel, we adapted this feature from Megatron-LM.

To use mixed precision training, you can easily specify the fp16 field in the configuration file. Currently, torch and apex amp cannot be guaranteed to work with tensor and pipeline parallelism, thus, only the last one is recommended if you are using hybrid parallelism.

Torch AMP

PyTorch provides mixed precision training in version 1.6 and above. It provides an easy way to cast data to fp16 format while keeping some operations such as reductions in fp32. You can configure the gradient scaler in the configuration.

from colossalai.engine import AMP_TYPE

fp16=dict(
    mode=AMP_TYPE.TORCH,
    # below are default values for grad scaler
    init_scale=2.**16,
    growth_factor=2.0,
    backoff_factor=0.5,
    growth_interval=2000,
    enabled=True
)

Apex AMP

For this mode, we rely on the Apex implementation for mixed precision training. We supported this plugin because it allows for finer control on the granularity of mixed precision. For example, O2 level (optimization level 2) will keep batch normalization in fp32.

The configuration is like below.

from colossalai.engine import AMP_TYPE

fp16 = dict(
    mode=AMP_TYPE.APEX,
    # below are the default values
    enabled=True, 
    opt_level='O1', 
    cast_model_type=None, 
    patch_torch_functions=None, 
    keep_batchnorm_fp32=None, 
    master_weights=None, 
    loss_scale=None, 
    cast_model_outputs=None,
    num_losses=1, 
    verbosity=1, 
    min_loss_scale=None, 
    max_loss_scale=16777216.0
)

Tensor Parallel AMP

We leveraged the Megatron-LM implementation to achieve mixed precision training while maintaining compatibility with complex tensor and pipeline parallel.

from colossalai.engine import AMP_TYPE

fp16 = dict(
    mode=AMP_TYPE.PARALLEL,
    # below are the default values
    clip_grad=0,
    log_num_zeros_in_grad=False,
    initial_scale=2 ** 32,
    min_scale=1,
    growth_factor=2,
    backoff_factor=0.5,
    growth_interval=1000,
    hysteresis=2
)