Source code for composer.core.precision

# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0

"""Enum class for the numerical precision to be used by the model."""

import contextlib
import os
import textwrap
from typing import Generator, Union

import torch

from composer.utils import StringEnum

try:
    import transformer_engine.pytorch as te
    te_installed = True
except ImportError:
    te_installed = False

__all__ = ['Precision', 'get_precision_context']


class Precision(StringEnum):
    """Enum class for the numerical precision to be used by the model.

    Attributes:
        FP32: Use 32-bit floating-point precision. Compatible with CPUs and GPUs.
        AMP_FP16: Use :mod:`torch.cuda.amp` with 16-bit floating-point precision. Only compatible
            with GPUs.
        AMP_BF16: Use :mod:`torch.cuda.amp` with 16-bit BFloat precision.
        AMP_FP8: Use :mod:`transformer_engine.pytorch.fp8_autocast` with 8-bit FP8 precison.
    """
    FP32 = 'fp32'
    AMP_FP16 = 'amp_fp16'
    AMP_BF16 = 'amp_bf16'
    AMP_FP8 = 'amp_fp8'


@contextlib.contextmanager
def get_precision_context(precision: Union[str, Precision]) -> Generator[None, None, None]:
    """Returns a context manager to automatically cast to a specific precision.

    Args:
        precision (str | Precision): Precision for the context
    """
    precision = Precision(precision)
    if precision == Precision.FP32:
        if torch.cuda.is_available():
            with torch.cuda.amp.autocast(False):
                yield
        else:
            # Yield here to avoid warnings about cuda not being available
            yield
    elif precision == Precision.AMP_FP16:
        # Retain compatibility with PyTorch < 1.10
        with torch.cuda.amp.autocast(True):
            yield
    elif precision == Precision.AMP_BF16:
        if torch.cuda.is_available():
            with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
                yield
        else:
            os.environ['XLA_USE_BF16'] = '1'
            yield
    elif precision == Precision.AMP_FP8:
        if te_installed and torch.cuda.get_device_capability()[0] > 8:
            from transformer_engine.common.recipe import DelayedScaling, Format

            # These default values for fp8_recipe are taken from NVidia's docs. We may want to change
            # these once we get a chance to do more convergence experiments.
            # https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/examples/fp8_primer.html#id1
            fp8_format = Format.HYBRID  # E4M3 during forward pass, E5M2 during backward pass
            fp8_recipe = DelayedScaling(fp8_format=fp8_format, amax_history_len=16, amax_compute_algo='max')
            with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
                yield
        else:
            if te_installed:
                raise RuntimeError('AMP_FP8 precision is used but current device does not support it.')
            else:
                raise ImportError(
                    textwrap.dedent("""\
                        AMP_FP8 precision is used but TransformerEngine is not installed.
                        After making sure torch is already installed, please install it using
                        pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable"""))
    else:
        raise ValueError(f'Unsupported precision: {precision}')