File size: 5,692 Bytes
23bd7af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
# coding=utf-8
# Copyright (c) 2020, 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.

"""Gradient clipping."""

import torch
from torch._six import inf

from apex.multi_tensor_apply import multi_tensor_applier
import amp_C

from megatron.model.module import param_is_not_shared
from megatron.mpu.layers import param_is_not_tensor_parallel_duplicate


def clip_grad_norm_fp32(parameters, grads_for_norm,
                        max_norm, norm_type=2,
                        model_parallel_group=None):
    """Clips gradient norm of an iterable of parameters whose gradients
       are in fp32.

    This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and
    added functionality to handle model parallel parameters. Note that
    the gradients are modified in place.

    Arguments:
        parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a
            single Tensor that will have gradients normalized
        grads_for_norm (Iterable[Tensor]): an iterable of Tensors or a single
            Tensor that will be used for calculating the grad norm.
        max_norm (float or int): max norm of the gradients
        norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for
            infinity norm.
        model_parallel_group (group): given the nature of the distributed
            optimizer, this is passed as an argument.

    Returns:
        Total norm of the parameters (viewed as a single vector).
    """

    if isinstance(parameters, torch.Tensor):
        parameters = [parameters]
    if isinstance(grads_for_norm, torch.Tensor):
        grads_for_norm = [grads_for_norm]

    # Grads.
    grads = []
    for param in parameters:
        if param.grad is not None:
            assert param.grad.type() == 'torch.cuda.FloatTensor'
            grads.append(param.grad.detach())

    # Norm parameters.
    max_norm = float(max_norm)
    norm_type = float(norm_type)
    total_norm = 0.0

    # Calculate norm.
    if norm_type == inf:
        total_norm = max(grad.abs().max() for grad in grads_for_norm)
        total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)])
        # Take max across all model-parallel GPUs.
        torch.distributed.all_reduce(total_norm_cuda,
                                     op=torch.distributed.ReduceOp.MAX,
                                     group=model_parallel_group)
        total_norm = total_norm_cuda[0].item()

    else:
        if norm_type == 2.0:
            dummy_overflow_buf = torch.cuda.IntTensor([0])
            # Use apex's multi-tensor applier for efficiency reasons.
            # Multi-tensor applier takes a function and a list of list
            # and performs the operation on that list all in one kernel.
            if grads_for_norm:
                grad_norm, _ = multi_tensor_applier(
                    amp_C.multi_tensor_l2norm,
                    dummy_overflow_buf,
                    [grads_for_norm],
                    False # no per-parameter norm
                )
            else:
                grad_norm = torch.cuda.FloatTensor([0])
            # Since we will be summing across data parallel groups,
            # we need the pow(norm-type).
            total_norm = grad_norm ** norm_type

        else:
            for grad in grads_for_norm:
                grad_norm = torch.norm(grad, norm_type)
                total_norm += grad_norm ** norm_type

        # Sum across all model-parallel GPUs.
        torch.distributed.all_reduce(total_norm,
                                     op=torch.distributed.ReduceOp.SUM,
                                     group=model_parallel_group)
        total_norm = total_norm.item() ** (1.0 / norm_type)

    # Scale.
    clip_coeff = max_norm / (total_norm + 1.0e-6)
    if clip_coeff < 1.0:
        dummy_overflow_buf = torch.cuda.IntTensor([0])
        multi_tensor_applier(amp_C.multi_tensor_scale,
                             dummy_overflow_buf,
                             [grads, grads],
                             clip_coeff)

    return total_norm


def count_zeros_fp32(parameters, model_parallel_group):

    if isinstance(parameters, torch.Tensor):
        parameters = [parameters]

    # Filter parameters based on:
    #   - grad should not be none
    #   - parameter should not be shared
    #   - should not be a replica due to tensor model parallelism
    total_num_zeros = torch.cuda.FloatTensor([0.0])
    for param in parameters:
        grad_not_none = param.grad is not None
        is_not_shared = param_is_not_shared(param)
        is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
        if grad_not_none and is_not_shared and is_not_tp_duplicate:
            grad = param.grad.detach()
            num_zeros = grad.numel() - torch.count_nonzero(grad)
            total_num_zeros = num_zeros + total_num_zeros

    # Sum across all model-parallel GPUs.
    torch.distributed.all_reduce(total_num_zeros,
                                 op=torch.distributed.ReduceOp.SUM,
                                 group=model_parallel_group)

    total_num_zeros = total_num_zeros.item()

    return total_num_zeros