auto_cast.py

#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from paddle.fluid.wrapped_decorator import (
    signature_safe_contextmanager,
    wrap_decorator,
)
from paddle.fluid import core
import contextlib
from paddle.fluid.framework import (
    Variable,
    OpProtoHolder,
    Parameter,
    _dygraph_tracer,
    dygraph_only,
    set_flags,
    get_flags,
)
import warnings
import copy
import functools
import paddle
import operator
import types

AMP_LEVEL = core.AmpLevel

__all__ = ['amp_guard', 'amp_decorate']

# The set of ops that support fp16 calculation and are considered numerically-
# safe and performance-critical. These ops are always converted to fp16.
WHITE_LIST = {
    'conv2d',
    'matmul',
    'matmul_v2',
    'mul',
    'fake_quantize_dequantize_abs_max',
    'fake_quantize_dequantize_moving_average_abs_max',
}

# The set of ops that support fp16 calculation and are considered numerically-
# dangerous and whose effects may also be observed in downstream ops.
BLACK_LIST = {
    'exp',
    'square',
    'log',
    'mean',
    'sum',
    'cos_sim',
    'softmax',
    'softmax_with_cross_entropy',
    'sigmoid_cross_entropy_with_logits',
    'c_softmax_with_cross_entropy',
    'cross_entropy',
    'cross_entropy2',
    # default fp32 can avoid return inf when the sum value large than 65504
    'reduce_sum',
    # FP16 performance of grad op is worse than that of FP32. Use FP32 by default.
    'linear_interp_v2',
    'nearest_interp_v2',
    'bilinear_interp_v2',
    'bicubic_interp_v2',
    'trilinear_interp_v2',
}

AMP_RELATED_FLAGS = [
    'FLAGS_cudnn_exhaustive_search',
    'FLAGS_conv_workspace_size_limit',
    'FLAGS_cudnn_batchnorm_spatial_persistent',
]

AMP_RELATED_FLAGS_SETTING = {
    'FLAGS_cudnn_exhaustive_search': 1,
    'FLAGS_conv_workspace_size_limit': 1000,
    'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
}

PURE_FP16_WHITE_LIST = set()
PURE_FP16_BLACK_LIST = {
    'lookup_table',
    'lookup_table_v2',
    'scatter',
    'scatter_grad',
    # FP16 performance of grad op is worse than that of FP32. Use FP32 by default.
    'linear_interp_v2',
    'nearest_interp_v2',
    'bilinear_interp_v2',
    'bicubic_interp_v2',
    'trilinear_interp_v2',
}

BF16_WHITE_LIST = {'conv2d', 'matmul_v2'}
BF16_BLACK_LIST = set()

PURE_BF16_WHITE_LIST = set()
PURE_BF16_BLACK_LIST = set()

_g_amp_state_ = None


def low_precision_op_list():
    op_list = paddle.fluid.core.get_low_precision_op_list()
    op_count = 0
    print('<---------------- low precision op list ------------------->')
    print('<---- op name ------|------- op count---------------------->')
    for x in op_list:
        print('  %-18s| %4d' % (x, op_list[x]))
        op_count += 1
    print(
        '<------------- low precision op num:{:5d} ----------------->'.format(
            op_count
        )
    )


def amp_state():
    global _g_amp_state_
    return _g_amp_state_


# NOTE(zhiqiu): similar as paddle.fluid.contrib.mixed_precision.fp16_lists.AutoMixedPrecisionLists._update_list
# The reason why not use AutoMixedPrecisionLists is that custom_black_varnames is not suitable for imperative mode.
def _update_list(
    custom_white_list, custom_black_list, level='O1', dtype='float16'
):
    """
    Update black and white list according to users' custom list.
    """
    if dtype == 'float16':
        if level == 'O1':
            _white_list = copy.copy(WHITE_LIST)
            _black_list = copy.copy(BLACK_LIST)
        else:
            _white_list = copy.copy(PURE_FP16_WHITE_LIST)
            _black_list = copy.copy(PURE_FP16_BLACK_LIST)
    else:
        if level == 'O1':
            _white_list = copy.copy(BF16_WHITE_LIST)
            _black_list = copy.copy(BF16_BLACK_LIST)
        else:
            _white_list = copy.copy(PURE_BF16_WHITE_LIST)
            _black_list = copy.copy(PURE_BF16_BLACK_LIST)
    if custom_white_list and custom_black_list:
        for op_name in custom_white_list:
            if op_name in custom_black_list:
                raise ValueError(
                    "Custom white list overlap " "custom black list"
                )
    if custom_white_list:
        for op_name in custom_white_list:
            if op_name in _black_list:
                _black_list.remove(op_name)
            _white_list.add(op_name)
    if custom_black_list:
        for op_name in custom_black_list:
            if op_name in _white_list:
                _white_list.remove(op_name)
            _black_list.add(op_name)
    return _white_list, _black_list


def _in_amp_guard():
    """
    Judge whether current code block is in `amp_guard` context.
    """
    tracer = _dygraph_tracer()
    if tracer:
        if tracer._amp_level == core.AmpLevel.O1:
            return True
        else:
            return False
    else:
        return False


def _in_pure_fp16_guard():
    tracer = _dygraph_tracer()
    return tracer and tracer._amp_level == core.AmpLevel.O2


def _is_gpu_float16_supported():
    """
    Judge whether current gpu support float16 amp.
    """
    prop = paddle.device.cuda.get_device_capability()
    return prop[0] >= 7


def _is_gpu_bfloat16_supported():
    """
    Judge whether current gpu support bfloat16 amp.
    """
    prop = paddle.device.cuda.get_device_capability()
    cuda_version = paddle.version.cuda()
    if cuda_version is not None and cuda_version != 'False':
        cuda_version_check = int(cuda_version.split('.')[0]) >= 11
    else:
        cuda_version_check = False
    return prop[0] >= 8 and cuda_version_check


@dygraph_only
def pure_fp16_initialize(models):
    for idx in range(len(models)):
        for layer in models[idx].sublayers(include_self=True):
            layer._casted_by_pure_fp16 = True
            if (layer._dtype == 'float16') or isinstance(
                layer,
                (
                    paddle.nn.BatchNorm,
                    paddle.nn.BatchNorm1D,
                    paddle.nn.BatchNorm2D,
                    paddle.nn.BatchNorm3D,
                    paddle.nn.LayerNorm,
                    paddle.nn.SyncBatchNorm,
                ),
            ):
                continue
            if isinstance(
                layer,
                (
                    paddle.incubate.nn.FusedFeedForward,
                    paddle.incubate.nn.FusedMultiHeadAttention,
                ),
            ):
                layer._amp_decorate(dtype='float16')
                continue
            layer._to_impl(
                dtype='float16', include_sublayers=False, floating_only=True
            )
    return models


@dygraph_only
def pure_bf16_initialize(models):
    for idx in range(len(models)):
        for layer in models[idx].sublayers(include_self=True):
            layer._to_impl(
                dtype='bfloat16', include_sublayers=False, floating_only=True
            )
    return models


def check_models(models):
    for model in models:
        if not isinstance(model, paddle.nn.Layer):
            raise RuntimeError(
                "Current train mode is pure fp16, models should be paddle.nn.Layer, but receive {}.".format(
                    type(model)
                )
            )
        if isinstance(model, paddle.DataParallel):
            raise RuntimeError(
                "For distributed AMP training, you should first use paddle.amp.decorate() to decotate origin model, and then call paddle.DataParallel get distributed model."
            )


def _is_valid_optimizer(optimizer):
    from paddle.distributed.fleet.meta_optimizers.dygraph_optimizer.dygraph_sharding_optimizer import (
        DygraphShardingOptimizer,
    )

    return isinstance(
        optimizer,
        (
            paddle.optimizer.Optimizer,
            paddle.fluid.optimizer.Optimizer,
            DygraphShardingOptimizer,
        ),
    )


def check_optimizers(optimizers):
    for optimizer in optimizers:
        if not _is_valid_optimizer(optimizer):
            raise RuntimeError(
                "Current train mode is pure fp16, optimizers should be paddle.optimizer.Optimizer or paddle.fluid.optimizer.Optimizer or DygraphShardingOptimizer, but receive {}.".format(
                    type(optimizer)
                )
            )


@signature_safe_contextmanager
@dygraph_only
def amp_guard(
    enable=True,
    custom_white_list=None,
    custom_black_list=None,
    level='O1',
    dtype='float16',
):
    """
    :api_attr: imperative

    Create a context which enables auto-mixed-precision(AMP) of operators executed in dynamic graph mode.
    If enabled, the input data type (float32 or float16) of each operator is decided
    by autocast algorithm for better performance.

    Commonly, it is used together with `GradScaler` to achieve Auto-Mixed-Precision in
    imperative mode. It is used together with `decorator` to achieve Pure fp16 in imperative mode.

    Args:
        enable(bool, optional): Enable auto-mixed-precision or not. Default is True.
        custom_white_list(set|list|tuple, optional): The custom white_list. It's the set of ops that support
             fp16 calculation and are considered numerically-safe and performance-critical. These ops
             will be converted to fp16.
        custom_black_list(set|list|tuple, optional): The custom black_list. The set of ops that support fp16
             calculation and are considered numerically-dangerous and whose effects may also be
             observed in downstream ops. These ops will not be converted to fp16.
        level(str, optional): Auto mixed precision level. Accepted values are "O1" and "O2": O1 represent mixed precision, the input data type of each operator will be casted by white_list and black_list;
             O2 represent Pure fp16, all operators parameters and input data will be casted to fp16, except operators in black_list, don't support fp16 kernel and batchnorm. Default is O1(amp)
        dtype(str, optional): Whether to use 'float16' or 'bfloat16'. Default is 'float16'.


    Examples:

     .. code-block:: python

        import numpy as np
        import paddle

        data = np.random.uniform(-1, 1, [10, 3, 32, 32]).astype('float32')
        with paddle.fluid.dygraph.guard():
            conv2d = paddle.fluid.dygraph.Conv2D(3, 2, 3)
            data = paddle.fluid.dygraph.to_variable(data)
            with paddle.fluid.dygraph.amp_guard():
                conv = conv2d(data)
                print(conv.dtype) # FP16
            with paddle.fluid.dygraph.amp_guard(enable=False):
                conv = conv2d(data)
                print(conv.dtype) # FP32

    """
    amp_state = locals()
    global _g_amp_state_
    original_state = _g_amp_state_
    _g_amp_state_ = amp_state

    # check amp_level: O0-O2
    level = level.upper()
    if not (level in ['O0', 'O1', 'O2']):
        raise ValueError(
            "level should be O0, O1 or O2. O0 represents fp32 train mode, O1 represents AMP train mode, O2 represents pure fp16/bf16 train mode."
        )

    # check amp_dtype: float16 or bfloat16
    dtype = dtype.lower()
    if not (dtype in ['float16', 'bfloat16']):
        raise ValueError("dtype should be 'float16' or 'bfloat16'.")

    # check tracer
    tracer = _dygraph_tracer()
    if not tracer:
        raise ValueError(
            "current_tracer is None, maybe it is not in imperative mode."
        )

    # check device_type:
    # NOTE: Now, amp only support gpu for float16 and bfloat16, xpu for float16, mlu for float16, npu for float16.
    # Maybe we will support cpu for bfloat16.
    if enable and not (
        tracer._expected_place.is_gpu_place()
        or tracer._expected_place.is_xpu_place()
        or tracer._expected_place.is_mlu_place()
        or tracer._expected_place.is_npu_place()
        or tracer._expected_place.is_custom_place()
    ):
        warnings.warn(
            'amp_guard can only be enabled on CUDAPlace, XPUPlace, MLUPlace, NPUPlace, and CustomPlace, current place is %s, so it makes no effect.'
            % tracer._expected_place
        )
        enable = False
    # For npu:
    if tracer._expected_place.is_npu_place() and (dtype == 'bfloat16'):
        warnings.warn('NPUPlace only support float16 amp.')
        enable = False
    # For xpu:
    if tracer._expected_place.is_xpu_place() and (dtype == 'bfloat16'):
        warnings.warn('XPUPlace only support float16 amp.')
        enable = False
    # For mlu:
    if tracer._expected_place.is_mlu_place() and (dtype == 'bfloat16'):
        warnings.warn('MLUPlace only support float16 amp.')
        enable = False
    # For custom device:
    if tracer._expected_place.is_custom_place() and (dtype == 'bfloat16'):
        warnings.warn('CustomPlace only support float16 amp.')
        enable = False
    # For gpu float16: Compute Capability should >= 7.
    # For gpu bfloat16: Compute Capability should >= 8 & CUDA Version should >= 11.
    if tracer._expected_place.is_gpu_place():
        if (dtype == 'float16') and not _is_gpu_float16_supported():
            prop = paddle.device.cuda.get_device_capability()
            warnings.warn(
                "For float16, amp only support NVIDIA GPU with Compute Capability 7.0 or higher, current GPU is: %s, with Compute Capability: %d.%d."
                % (paddle.device.cuda.get_device_name(), prop[0], prop[1])
            )
        elif (dtype == 'bfloat16') and not _is_gpu_bfloat16_supported():
            prop = paddle.device.cuda.get_device_capability()
            cuda_version = paddle.version.cuda()
            warnings.warn(
                "For bfloat16, amp only support NVIDIA GPU with Compute Capability 8.0 or higher and CUDA Version 11.0 or higher, current GPU is: %s, with Compute Capability: %d.%d, current CUDA Version is: %s."
                % (
                    paddle.device.cuda.get_device_name(),
                    prop[0],
                    prop[1],
                    cuda_version,
                )
            )

    amp_dtype = dtype

    if level == 'O1':
        amp_level = AMP_LEVEL.O1
        if dtype == 'float16':
            _white_list = WHITE_LIST
            _black_list = BLACK_LIST
        elif dtype == 'bfloat16':
            _white_list = BF16_WHITE_LIST
            _black_list = BF16_BLACK_LIST

    elif level == 'O2':
        amp_level = AMP_LEVEL.O2
        if dtype == 'float16':
            _white_list = PURE_FP16_WHITE_LIST
            _black_list = PURE_FP16_BLACK_LIST
        elif dtype == 'bfloat16':
            _white_list = BF16_WHITE_LIST
            _black_list = BF16_BLACK_LIST
    elif level == 'O0':
        amp_level = AMP_LEVEL.O0
        if dtype == 'float16':
            _white_list = WHITE_LIST
            _black_list = BLACK_LIST
        elif dtype == 'bfloat16':
            _white_list = BF16_WHITE_LIST
            _black_list = BF16_BLACK_LIST

    if custom_white_list or custom_black_list:
        _white_list, _black_list = _update_list(
            custom_white_list, custom_black_list, level, dtype
        )

    if not enable:
        amp_level = AMP_LEVEL.O0
        amp_dtype = "float32"

    if tracer:
        # enable auto_cast
        original_amp_level = tracer._amp_level
        tracer._amp_level = amp_level

        # set amp op list
        original_white_list, original_black_list = tracer._get_amp_op_list()
        tracer._set_amp_op_list(_white_list, _black_list)

        # TODO(zhiqiu) set amp related flags automatically in this guard
        # Currently, if FLAGS_cudnn_batchnorm_spatial_persistent is set True in amp_guard,
        # batch_norm can run in fast mode, but batch_norm_grad can not if backward if not executed insise amp_guard.
        # So, users need to set related flags manually.

        # original_flags = get_flags(AMP_RELATED_FLAGS)
        # set_flags(AMP_RELATED_FLAGS_SETTING)

        # set amp dtype
        original_amp_dtype = tracer._amp_dtype
        tracer._amp_dtype = amp_dtype

    # restore status
    try:
        yield
    finally:
        if tracer:
            _g_amp_state_ = original_state
            tracer._amp_level = original_amp_level
            tracer._set_amp_op_list(original_white_list, original_black_list)
            # set_flags(original_flags)
            tracer._amp_dtype = original_amp_dtype


class StateDictHook:
    def __init__(self, save_dtype):
        self._save_dtype = save_dtype

    def __call__(self, state_dict):
        for key in state_dict:
            param = state_dict[key]
            with paddle.fluid.dygraph.guard():
                if paddle.is_floating_point(param):
                    param_applied = paddle.cast(param, self._save_dtype)
                    param_applied.name = param.name
                    state_dict[key] = param_applied


def _set_multi_precision(optimizer, multi_precision):
    from paddle.distributed.fleet.meta_optimizers.dygraph_optimizer.dygraph_sharding_optimizer import (
        DygraphShardingOptimizer,
    )

    optimizer = (
        optimizer._inner_optimizer
        if isinstance(optimizer, DygraphShardingOptimizer)
        else optimizer
    )
    if hasattr(optimizer, "_multi_precision"):
        optimizer._multi_precision = multi_precision


@dygraph_only
def amp_decorate(
    models,
    optimizers=None,
    level='O1',
    dtype='float16',
    master_weight=None,
    save_dtype=None,
):
    """
    Decorate models and optimizers for auto-mixed-precision. When level is O1(amp), the decorate will do nothing.
    When level is O2(pure fp16), the decorate will cast all parameters of models to FP16, except BatchNorm and LayerNorm.

    Commonly, it is used together with `amp_guard` to achieve Pure fp16 in imperative mode.

    Args:
        models(Layer|list of Layer, optional): The defined models by user, models must be either a single model or a list of models. Default is None.
        optimizers(Optimizer|list of Optimizer, optional): The defined optimizers by user, optimizers must be either a single optimizer or a list of optimizers. Default is None.
        level(str, optional): Auto mixed precision level. Accepted values are "O1" and "O2": O1 represent mixed precision, the decorator will do nothing;
             O2 represent Pure fp16/bf16, the decorator will cast all parameters of models to FP16/BF16, except BatchNorm and LayerNorm. Default is O1(amp)
        dtype(str, optional): Whether to use 'float16' or 'bfloat16'. Default is 'float16'.
        master_weight(bool, optinal): For level='O2', whether to use multi-precision during weight updating. If master_weight is None, in O2 level optimizer will use multi-precision. Default is None.
        save_dtype(float, optional): The save model parameter dtype when use `paddle.save` or `paddle.jit.save`,it should be float16, bfloat16, float32, float64 or None.
             The save_dtype will not change model parameters dtype, it just change the state_dict dtype. When save_dtype is None, the save dtype is same as model dtype. Default is None.

    Examples:

     .. code-block:: python

        # required: gpu
        # Demo1: single model and optimizer:
        import paddle

        model = paddle.nn.Conv2D(3, 2, 3, bias_attr=False)
        optimizer = paddle.optimizer.SGD(parameters=model.parameters())

        model, optimizer = paddle.fluid.dygraph.amp_decorate(models=model, optimizers=optimizer, level='O2')

        data = paddle.rand([10, 3, 32, 32])

        with paddle.fluid.dygraph.amp_guard(enable=True, custom_white_list=None, custom_black_list=None, level='O2'):
            output = model(data)
            print(output.dtype) # FP16

        # required: gpu
        # Demo2: multi models and optimizers:
        model2 = paddle.nn.Conv2D(3, 2, 3, bias_attr=False)
        optimizer2 = paddle.optimizer.Adam(parameters=model2.parameters())

        models, optimizers = paddle.fluid.dygraph.amp_decorate(models=[model, model2], optimizers=[optimizer, optimizer2], level='O2')

        data = paddle.rand([10, 3, 32, 32])

        with paddle.fluid.dygraph.amp_guard(enable=True, custom_white_list=None, custom_black_list=None, level='O2'):
            output = models[0](data)
            output2 = models[1](data)
            print(output.dtype) # FP16
            print(output2.dtype) # FP16

        # required: gpu
        # Demo3: optimizers is None:
        model3 = paddle.nn.Conv2D(3, 2, 3, bias_attr=False)
        optimizer3 = paddle.optimizer.Adam(parameters=model2.parameters())

        model = paddle.fluid.dygraph.amp_decorate(models=model3, level='O2')

        data = paddle.rand([10, 3, 32, 32])

        with paddle.fluid.dygraph.amp_guard(enable=True, custom_white_list=None, custom_black_list=None, level='O2'):
            output = model(data)
            print(output.dtype) # FP16
    """
    if not (level in ['O1', 'O2']):
        raise ValueError(
            "level should be O1 or O2, O1 represent AMP train mode, O2 represent Pure fp16 train mode."
        )

    if level == 'O1':
        if optimizers is None:
            return models
        else:
            return models, optimizers

    models_is_list = False
    if isinstance(models, paddle.nn.Layer):
        models_is_list = False
        models = [models]
        check_models(models)
    elif isinstance(models, list):
        check_models(models)
        models_is_list = True
    else:
        raise TypeError(
            "models must be either a single model or a list of models."
        )
    if dtype == 'float16':
        models = pure_fp16_initialize(models=models)
    elif dtype == 'bfloat16':
        models = pure_bf16_initialize(models=models)
    else:
        raise TypeError("dtype only support float16 or bfloat16.")

    if optimizers is not None:
        # check optimizers
        optimizers_is_list = False
        if _is_valid_optimizer(optimizers):
            optimizers_is_list = False
            optimizers = [optimizers]
            check_optimizers(optimizers)
        elif isinstance(optimizers, list):
            check_optimizers(optimizers)
            optimizers_is_list = True
        else:
            raise TypeError(
                "optimizers must be either a single optimizer or a list of optimizers."
            )
        # support master_weight
        use_multi_precision = not (master_weight is False)
        for opt in optimizers:
            _set_multi_precision(opt, use_multi_precision)

    if save_dtype is not None:
        if not (save_dtype in ['float16', 'bfloat16', 'float32', 'float64']):
            raise ValueError(
                "save_dtype can only be float16 float32 or float64, but your input save_dtype is %s."
                % save_dtype
            )
        for idx in range(len(models)):
            for layer in models[idx].sublayers(include_self=True):
                layer.register_state_dict_hook(StateDictHook(save_dtype))

    if models_is_list:
        if optimizers is not None:
            if optimizers_is_list:
                return models, optimizers
            else:
                return models, optimizers[0]
        else:
            return models
    else:
        if optimizers is not None:
            if optimizers_is_list:
                return models[0], optimizers
            else:
                return models[0], optimizers[0]
        else:
            return models[0]