nemo/core/optim/optimizer_with_main_params.py

# Copyright (c) 2021, NVIDIA CORPORATION.  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 contextlib import contextmanager

import torch

from nemo.utils import logging

try:
    import amp_C
    from apex.multi_tensor_apply import multi_tensor_applier
    from apex.transformer.parallel_state import get_data_parallel_group, get_data_parallel_world_size
    from apex.transformer.tensor_parallel import copy_tensor_model_parallel_attributes

    HAVE_APEX = True

except (ImportError, ModuleNotFoundError):

    HAVE_APEX = False


def _zero_grad_group_helper(group, set_to_none):
    """Zero out the gradient for a group of parameters.
    Note: copied from torch.optim.optimizer."""
    for param in group:
        if param.grad is not None:
            if set_to_none:
                param.grad = None
            else:
                if param.grad.grad_fn is not None:
                    param.grad.detach_()
                else:
                    param.grad.requires_grad_(False)
                param.grad.zero_()


def _multi_tensor_copy_this_to_that(this, that, overflow_buf):
    """Use multi-tensor-applier to copy values from one list to another.
    We don't have a blfoat16 implementation so for now if the overflow_buf
    is not provided, we default back to simple loop copy to be compatible
    with bfloat16."""
    if overflow_buf:
        # Scaling with factor `1.0` is equivalent to copy.
        multi_tensor_applier(amp_C.multi_tensor_scale, overflow_buf, [this, that], 1.0)
    else:
        # FIXME: use multi-tensor applier for bf16
        for this_, that_ in zip(this, that):
            that_.copy_(this_)


class GradBucket(object):
    """
    Persistent buffer for main gradients that remains allocated between training iterations
    """

    def __init__(self, numel, chunk_size_mb):
        if not HAVE_APEX:
            raise ImportError(
                "Apex was not found. Please see the NeMo README for installation instructions: https://github.com/NVIDIA/NeMo#megatron-gpt."
            )

        self.numel = numel
        self.data = torch.zeros(self.numel, dtype=torch.float, device=torch.cuda.current_device(), requires_grad=False)

        self.chunk_size_mb = chunk_size_mb
        if self.chunk_size_mb > 0:
            chunk_size_bytes = chunk_size_mb * 1024 * 1024
            self.chunk_size_numel = chunk_size_bytes // 4
            self.num_chunks = self.numel // self.chunk_size_numel
            self.numel_per_chunk = [self.chunk_size_numel] * self.num_chunks
            if self.numel % self.chunk_size_numel != 0:
                self.num_chunks += 1
                self.numel_per_chunk.append(self.numel % self.chunk_size_numel)

            self.start_index_per_chunk = torch.cumsum(torch.tensor([0] + self.numel_per_chunk[:-1]), dim=0)
            self.current_chunk = 0
            self.computed_numel_per_chunk = [0] * self.num_chunks

    def zero(self):
        """Reset the buffer to zero."""
        self.data.zero_()

    def allreduce_buffer(self):
        """Synchronous buffer data allreduce """
        self.data.div_(get_data_parallel_world_size())
        torch.distributed.all_reduce(self.data, group=get_data_parallel_group())

    def get(self, shape, start_index):
        """Return a tensor with the input `shape` as a view into the
        1-D data starting at `start_index`."""
        end_index = start_index + shape.numel()
        assert end_index <= self.numel, 'requested tensor is out of the buffer range.'
        buffer_tensor = self.data[start_index:end_index]
        buffer_tensor = buffer_tensor.view(shape)

        grad_chunk_info = None
        if self.chunk_size_mb > 0:
            grad_chunk_info = {}
            chunk = start_index // self.chunk_size_numel
            chunk_start_index = self.start_index_per_chunk[chunk]
            chunk_end_index = chunk_start_index + self.numel_per_chunk[chunk]
            grad_chunk_info[chunk] = min(chunk_end_index, end_index) - start_index
            while chunk_end_index < end_index:
                chunk += 1
                chunk_start_index = self.start_index_per_chunk[chunk]
                chunk_end_index = chunk_start_index + self.numel_per_chunk[chunk]
                grad_chunk_info[chunk] = min(chunk_end_index, end_index) - chunk_start_index

        return buffer_tensor, grad_chunk_info

    def update_chunk_info(self, grad_chunk_info):
        for chunk in grad_chunk_info.keys():
            self.computed_numel_per_chunk[chunk] += grad_chunk_info[chunk]

    def get_allreduce_tensor(self):
        if self.computed_numel_per_chunk[self.current_chunk] == self.numel_per_chunk[self.current_chunk]:
            chunk_start_index = self.start_index_per_chunk[self.current_chunk]
            chunk_end_index = chunk_start_index + self.numel_per_chunk[self.current_chunk]
            allreduce_tensor = self.data[chunk_start_index:chunk_end_index]

            self.computed_numel_per_chunk[self.current_chunk] = 0
            self.current_chunk += 1
            if self.current_chunk == self.num_chunks:
                self.current_chunk = 0

            return allreduce_tensor

        return None


class MainParamsOptimizerWrapper(torch.optim.Optimizer):
    """
    Float16 optimizer wrapper for half precision (fp16 and bf16) data types.
    This optimizer wrapper holds main parameters and gradients in fp32 to support
    stable convergence.

    Arguments:
        optimizer: base optimizer such as Adam or SGD.
        fp32_grad_accum: to enable the use of fp32 in gradient accumulation and allreduce.
        contiguous_grad_bucket: to enable allocating the master gradients in the 
            contiguous memory space to reduce memory fragmentation.
        async_grad_allreduce: enable asynchronous gradient allreduce that is executed
            along with the training step backprop.
    """

    def __init__(
        self,
        optimizer,
        fp32_grad_accum=False,
        contiguous_grad_bucket=False,
        async_grad_allreduce=False,
        grad_div_ar_fusion=True,
        grad_allreduce_chunk_size_mb=0,
    ):
        if not HAVE_APEX:
            raise ImportError(
                "Apex was not found. Please see the NeMo README for installation instructions: https://github.com/NVIDIA/NeMo#megatron-gpt."
            )

        self.optimizer = optimizer
        assert self.optimizer, 'no optimizer is provided.'
        if contiguous_grad_bucket:
            assert fp32_grad_accum, 'contiguous gradient buffer assumes using fp32 grad.'
        if async_grad_allreduce:
            assert fp32_grad_accum, (
                'async allreduce applies to master gradients only, '
                'which is supposed to be accumulated after grad op.'
            )
            assert contiguous_grad_bucket, (
                'currently async_grad_allreduce is supported only ' 'with contiguous_grad_bucket.'
            )

        self._fp32_grad_accum = fp32_grad_accum
        self._contiguous_grad_bucket = contiguous_grad_bucket

        # used with tensor parallel only (no pipeline parallelism)
        # be careful, weight update cannot start until all async grad AR works are done
        self._async_grad_allreduce = async_grad_allreduce and get_data_parallel_world_size() > 1
        self._grad_divisor = 1 / get_data_parallel_world_size()

        if self._async_grad_allreduce:
            # use @no_sync to disable backward grad sync during gradient accumulation
            self._require_backward_grad_sync = True
            self._grad_div_ar_fusion = grad_div_ar_fusion
            self._grad_allreduce_chunk_size_mb = grad_allreduce_chunk_size_mb
        else:
            self._require_backward_grad_sync = False
            self._grad_div_ar_fusion = False
            self._grad_allreduce_chunk_size_mb = 0

        # Dummy tensor needed for apex multi-apply tensor.
        self._dummy_overflow_buf = None

        # Create persistent buffers for main gradients in contiguous memory space
        # - Chunked element-wise and allreduce ops without creating a temporary buffer for merged operation
        # - Low memory fragmentation
        self._main_grad_buffers = None
        if self._contiguous_grad_bucket:
            self._main_grad_buffers = {}
            # get the size of buffers
            num_elements = {}
            for i, param_group in enumerate(self.optimizer.param_groups):
                for param in param_group['params']:
                    if param.requires_grad:
                        num_elements[i] = num_elements.get(i, 0) + param.data.nelement()

                # Allocate gradient memory buffers for each data type
                if any(param.requires_grad for param in param_group['params']):
                    self._main_grad_buffers[i] = GradBucket(num_elements[i], self._grad_allreduce_chunk_size_mb)

        # Three groups of parameters:
        self.float16_groups = []  # original float16 parameters
        self.fp32_from_float16_groups = []  # fp32 copy of float16 parameters
        self.fp32_from_fp32_groups = []  # original fp32 parameters

        # gradient function hooks
        if self._fp32_grad_accum:
            self.grad_accs = []

        # For all the groups in the original optimizer:
        for i, param_group in enumerate(self.optimizer.param_groups):
            float16_params_this_group = []
            fp32_params_this_group = []
            fp32_from_float16_params_this_group = []
            # For all the parameters in this group:
            for j, param in enumerate(param_group['params']):
                if param.requires_grad:
                    # float16 params:
                    if param.type() in ['torch.cuda.HalfTensor', 'torch.cuda.BFloat16Tensor']:
                        float16_params_this_group.append(param)

                        # Allocate the main parameter
                        main_param = param.detach().clone().float()

                        # Copy tensor model parallel attributes.
                        copy_tensor_model_parallel_attributes(main_param, param)
                        if hasattr(param, 'shared'):
                            main_param.shared = param.shared

                        # Assign the grad buffer offset to main parameters
                        if self._contiguous_grad_bucket:
                            num_elements[i] -= param.data.nelement()
                            main_param.grad, grad_chunk_info = self._main_grad_buffers[i].get(
                                param.data.shape, num_elements[i]
                            )
                            # Add a pointer to main_grad in model param for first-last stage embedding param reduction
                            param.main_grad = main_param.grad

                        # Replace the optimizer params with the new fp32 copy.
                        param_group['params'][j] = main_param
                        fp32_from_float16_params_this_group.append(main_param)
                        # Reset existing state dict key to the new main param.
                        if param in self.optimizer.state:
                            self.optimizer.state[main_param] = self.optimizer.state.pop(param)
                    # fp32 params.
                    elif param.type() == 'torch.cuda.FloatTensor':
                        fp32_params_this_group.append(param)
                        param_group['params'][j] = param

                    else:
                        raise TypeError(
                            'Wrapped parameters must be one of '
                            'torch.cuda.FloatTensor,  '
                            'torch.cuda.HalfTensor, or '
                            'torch.cuda.BFloat16Tensor. '
                            'Received {}'.format(param.type())
                        )

                # Add gradient accumulation hook for fp32 grad accumulation
                if self._fp32_grad_accum and param.requires_grad:
                    # Expand so we get access to grad_fn
                    param_tmp = param.expand_as(param)
                    # Get the gradient accumulator function.
                    grad_acc = param_tmp.grad_fn.next_functions[0][0]
                    grad_acc.register_hook(self._make_param_hook(param, main_param, i, grad_chunk_info))
                    self.grad_accs.append(grad_acc)

            self.float16_groups.append(float16_params_this_group)
            self.fp32_from_float16_groups.append(fp32_from_float16_params_this_group)
            self.fp32_from_fp32_groups.append(fp32_params_this_group)

        # Leverage state_dict() and load_state_dict() to
        # recast preexisting per-param state tensors
        self.optimizer.load_state_dict(self.optimizer.state_dict())

    def _make_param_hook(self, param, main_param, i, grad_chunk_info):
        """Create the grad accumulation and all-reduce hook for backprop."""
        # Hook used for back-prop.
        def param_hook(*unused):
            # Accumulates gradients on main gradients
            if param.grad is not None:
                if main_param.grad is None:
                    main_param.grad = param.grad.float()
                else:
                    main_param.grad.add_(param.grad.data)
                # Deallocate grad memory.
                param.grad = None

            # Asynchronous gradients allreduce accross data_parallel ranks
            if self._require_backward_grad_sync:
                if self._grad_allreduce_chunk_size_mb > 0:
                    self._main_grad_buffers[i].update_chunk_info(grad_chunk_info)
                    while True:
                        allreduce_tensor = self._main_grad_buffers[i].get_allreduce_tensor()
                        if allreduce_tensor is None:
                            break
                        if self._grad_div_ar_fusion:
                            torch.distributed.all_reduce(
                                allreduce_tensor,
                                group=get_data_parallel_group(),
                                async_op=True,
                                op=torch.distributed._make_nccl_premul_sum(self._grad_divisor),
                            )
                        else:
                            allreduce_tensor.div_(get_data_parallel_world_size())
                            torch.distributed.all_reduce(
                                allreduce_tensor, group=get_data_parallel_group(), async_op=True,
                            )
                else:
                    if self._grad_div_ar_fusion:
                        torch.distributed.all_reduce(
                            main_param.grad,
                            group=get_data_parallel_group(),
                            async_op=True,
                            op=torch.distributed._make_nccl_premul_sum(self._grad_divisor),
                        )
                    else:
                        main_param.grad.div_(get_data_parallel_world_size())
                        torch.distributed.all_reduce(
                            main_param.grad, group=get_data_parallel_group(), async_op=True,
                        )

        return param_hook

    def zero_grad(self, set_to_none=True):
        """We only need to zero the model related parameters, i.e.,
        float16_groups & fp32_from_fp32_groups. We additionally zero
        fp32_from_float16_groups as a memory optimization to reduce
        fragmentation; in the case of set_to_none==True, the space
        used by this field can be safely deallocated at this point."""
        for group in self.float16_groups:
            _zero_grad_group_helper(group, set_to_none)
        if self._contiguous_grad_bucket:
            for i in self._main_grad_buffers:
                self._main_grad_buffers[i].zero()
        else:
            for group in self.fp32_from_float16_groups:
                _zero_grad_group_helper(group, set_to_none)
        for group in self.fp32_from_fp32_groups:
            _zero_grad_group_helper(group, set_to_none)

    def copy_model_grads_to_main_grads(self):
        # This only needs to be done for the float16 group.
        for model_group, main_group in zip(self.float16_groups, self.fp32_from_float16_groups):
            for model_param, main_param in zip(model_group, main_group):
                if model_param.grad is not None:
                    main_param.grad = model_param.grad.float()

                # Safe to deallocate model's grad after copying.
                # (If using contiguous buffers, main_grad's memory should
                # persist and therefore should not be deallocated.)
                model_param.grad = None

    def _get_model_and_main_params_data_float16(self):
        model_data = []
        main_data = []
        half_dtype = None
        for model_group, main_group in zip(self.float16_groups, self.fp32_from_float16_groups):
            for model_param, main_param in zip(model_group, main_group):
                if half_dtype is None:
                    half_dtype = model_param.data.dtype
                model_data.append(model_param.data)
                main_data.append(main_param.data)
        return model_data, main_data, half_dtype

    def _set_overflow_buffer(self, half_dtype):
        if half_dtype == torch.float16:
            if self._dummy_overflow_buf is None:
                self._dummy_overflow_buf = torch.cuda.IntTensor([0])
            else:
                self._dummy_overflow_buf.fill_(0)

    def _copy_main_params_to_model_params(self):
        # Only needed for the float16 params.
        model_data, main_data, half_dtype = self._get_model_and_main_params_data_float16()
        self._set_overflow_buffer(half_dtype)
        _multi_tensor_copy_this_to_that(this=main_data, that=model_data, overflow_buf=self._dummy_overflow_buf)

    def _copy_model_params_to_main_params(self):
        # Only needed for the float16 params.
        model_data, main_data, half_dtype = self._get_model_and_main_params_data_float16()
        self._set_overflow_buffer(half_dtype)
        _multi_tensor_copy_this_to_that(this=model_data, that=main_data, overflow_buf=self._dummy_overflow_buf)

    def reload_model_params(self):
        self._copy_model_params_to_main_params()

    @torch.no_grad()
    def step(self, **kwargs):
        # while async grad allreduce is enabled, bprop will keep moving forward without waiting for
        # the finish of async grad AR works. Hence, to guarantee the correctness of grads reduction,
        # we cannot start weight update until all async grad AR works are done.
        if self._async_grad_allreduce:
            torch.cuda.synchronize()

        # Step the optimizer.
        self.optimizer.step(closure=None, **kwargs)

        # Update params from main params.
        with torch.no_grad():
            self._copy_main_params_to_model_params()

        # Successful update.
        return True

    def state_dict(self):
        state_dict = {}
        state_dict['optimizer'] = self.optimizer.state_dict()
        state_dict['fp32_from_fp16_params'] = self.fp32_from_float16_groups
        return state_dict

    def load_state_dict(self, state_dict):
        # Optimizer.
        optimizer_key = 'optimizer'
        if optimizer_key not in state_dict:
            optimizer_key = 'optimizer_state_dict'
            logging.info('***WARNING*** loading optimizer from ' 'an old checkpoint ...')
        self.optimizer.load_state_dict(state_dict[optimizer_key])

        # Copy data for the main params.
        fp32_from_float16_params_key = 'fp32_from_fp16_params'
        if fp32_from_float16_params_key not in state_dict:
            fp32_from_float16_params_key = 'fp32_from_fp16'
        for current_group, saved_group in zip(self.fp32_from_float16_groups, state_dict[fp32_from_float16_params_key]):
            for current_param, saved_param in zip(current_group, saved_group):
                current_param.data.copy_(saved_param.data)

    def allreduce_main_grads(self):
        for i in self._main_grad_buffers:
            self._main_grad_buffers[i].allreduce_buffer()

    @contextmanager
    def no_sync(self):
        """ A context manager to disable gradient synchronizations across
        data-parallel ranks."""
        old_require_backward_grad_sync = self._require_backward_grad_sync
        self._require_backward_grad_sync = False
        try:
            yield
        finally:
            self._require_backward_grad_sync = old_require_backward_grad_sync

    @property
    def async_master_grads_allreudce(self):
        return self._async_grad_allreduce

    @property
    def fp32_grad_accumulation(self):
        return self._fp32_grad_accum

    def get_parameters(self):
        params = []
        for param_group in self.optimizer.param_groups:
            for param in param_group['params']:
                params.append(param)
        return params

    # Promote state so it can be retrieved or set via
    # "optimizer_instance.state"
    def _get_state(self):
        if hasattr(self, 'optimizer'):
            return self.optimizer.state
        else:
            return []

    def _set_state(self, value):
        self.optimizer.state = value

    state = property(_get_state, _set_state)

    # Promote param_groups so it can be retrieved or set via
    # "optimizer_instance.param_groups"
    # (for example, to adjust the learning rate)
    def _get_param_groups(self):
        if hasattr(self, 'optimizer'):
            return self.optimizer.param_groups
        else:
            return []

    def _set_param_groups(self, value):
        self.optimizer.param_groups = value

    param_groups = property(_get_param_groups, _set_param_groups)

    # Promote defaults so it can be retrieved or set via
    # "optimizer_instance.defaults
    def _get_defaults(self):
        if hasattr(self, 'optimizer'):
            return self.optimizer.defaults
        else:
            return []

    def _set_defaults(self, value):
        self.optimizer.defaults = value

    defaults = property(_get_defaults, _set_defaults)