File size: 9,821 Bytes
67b96c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/block.py
# Commit id: c94cd09744d20f0ac587a351ff6ff2e8ad11ae1b

# Previously adapted from https://github.com/mlcommons/training_results_v1.1/blob/main/NVIDIA/benchmarks/bert/implementations/pytorch/padding.py

import torch
import torch.nn.functional as F
from einops import rearrange, repeat


class IndexFirstAxis(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input, indices):
        ctx.save_for_backward(indices)
        assert input.ndim >= 2
        ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
        second_dim = other_shape.numel()
        # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
        # return input[indices]
        return torch.gather(
            rearrange(input, "b ... -> b (...)"), 0, repeat(indices, "z -> z d", d=second_dim)
        ).reshape(-1, *other_shape)

    @staticmethod
    def backward(ctx, grad_output):
        (indices,) = ctx.saved_tensors
        assert grad_output.ndim >= 2
        other_shape = grad_output.shape[1:]
        grad_output = rearrange(grad_output, "b ... -> b (...)")
        grad_input = torch.zeros(
            [ctx.first_axis_dim, grad_output.shape[1]],
            device=grad_output.device,
            dtype=grad_output.dtype,
        )
        # TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
        # grad_input[indices] = grad_output
        grad_input.scatter_(0, repeat(indices, "z -> z d", d=grad_output.shape[1]), grad_output)
        return grad_input.reshape(ctx.first_axis_dim, *other_shape), None


index_first_axis = IndexFirstAxis.apply


class IndexPutFirstAxis(torch.autograd.Function):
    @staticmethod
    def forward(ctx, values, indices, first_axis_dim):
        ctx.save_for_backward(indices)
        assert indices.ndim == 1
        assert values.ndim >= 2
        output = torch.zeros(
            first_axis_dim, *values.shape[1:], device=values.device, dtype=values.dtype
        )
        # TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
        output[indices] = values
        # output.scatter_(0, repeat(indices, 'z -> z d', d=values.shape[1]), values)
        return output

    @staticmethod
    def backward(ctx, grad_output):
        (indices,) = ctx.saved_tensors
        # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
        grad_values = grad_output[indices]
        # grad_values = torch.gather(grad_output, 0, repeat(indices, 'z -> z d', d=grad_output.shape[1]))
        return grad_values, None, None


index_put_first_axis = IndexPutFirstAxis.apply


class IndexFirstAxisResidual(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input, indices):
        ctx.save_for_backward(indices)
        assert input.ndim >= 2
        ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
        second_dim = other_shape.numel()
        # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
        output = input[indices]
        # We don't want to reshape input (b ... -> b (...)) since it could change the channel_last
        # memory format to channel_first. In other words, input might not be contiguous.
        # If we don't detach, Pytorch complains about output being a view and is being modified inplace
        return output, input.detach()

    @staticmethod
    def backward(ctx, grad_output, grad_residual):
        (indices,) = ctx.saved_tensors
        assert grad_output.ndim >= 2
        other_shape = grad_output.shape[1:]
        assert grad_residual.shape[1:] == other_shape
        grad_input = grad_residual
        # grad_input[indices] += grad_output
        indices = indices.reshape(indices.shape[0], *((1,) * (grad_output.ndim - 1)))
        indices = indices.expand_as(grad_output)
        grad_input.scatter_add_(0, indices, grad_output)
        return grad_input.reshape(ctx.first_axis_dim, *other_shape), None


index_first_axis_residual = IndexFirstAxisResidual.apply


def unpad_input(hidden_states, attention_mask):
    """
    Arguments:
        hidden_states: (batch, seqlen, ...)
        attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
    Return:
        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
        indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
        cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
        max_seqlen_in_batch: int
    """
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
    # TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
    # bool mask, then call nonzero to get the indices, then index with those. The indices is @dim
    # times larger than it needs to be, wasting memory. It's faster and more memory-efficient to
    # index with integer indices. Moreover, torch's index is a bit slower than it needs to be,
    # so we write custom forward and backward to make it a bit faster.
    return (
        index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


def unpad_input_for_concatenated_sequences(hidden_states, attention_mask_in_length):
    """
    Supports concatenating short samples in one sequence. The attention_mask_in_length is utilized to mask other short samples. It helps efficient training of variant lengths-based samples (e.g., the supervised fine-tuning task in large language model).
    The motivation for this function is explained [here](https://github.com/Dao-AILab/flash-attention/issues/432#issuecomment-1668822286).

    For example, if batch = 3 and seqlen = 6, the attention_mask_in_length is:
        ```
        [
          [2, 3, 0, 0, 0, 0],
          [3, 2, 0, 0, 0, 0],
          [6, 0, 0, 0, 0, 0]
        ]
        ```
    , which refers to the 3D-attention mask:
        ```
        [
          [
            [1, 0, 0, 0, 0, 0],
            [1, 1, 0, 0, 0, 0],
            [0, 0, 1, 0, 0, 0],
            [0, 0, 1, 1, 0, 0],
            [0, 0, 1, 1, 1, 0],
            [0, 0, 0, 0, 0, 1]
          ],
          [
            [1, 0, 0, 0, 0, 0],
            [1, 1, 0, 0, 0, 0],
            [1, 1, 1, 0, 0, 0],
            [0, 0, 0, 1, 0, 0],
            [0, 0, 0, 1, 1, 0],
            [0, 0, 0, 0, 0, 1]
          ],
          [
            [1, 0, 0, 0, 0, 0],
            [1, 1, 0, 0, 0, 0],
            [1, 1, 1, 0, 0, 0],
            [1, 1, 1, 1, 0, 0],
            [1, 1, 1, 1, 1, 0],
            [1, 1, 1, 1, 1, 1]
          ]
        ]
        ```.

    Arguments:
        hidden_states: (batch, seqlen, ...)
        attention_mask_in_length: (batch, seqlen), int, a nonzero number (e.g., 1, 2, 3, etc.) means length of concatenated sequence in b-th batch, and 0 means none.
    Return:
        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
        indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
        cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
        max_seqlen_in_batch: int
    """
    length = attention_mask_in_length.sum(dim=-1)
    seqlen = attention_mask_in_length.size(-1)
    attention_mask_2d = torch.arange(seqlen, device=length.device, dtype=length.dtype).expand(len(length),
                                                                                              seqlen) < length.unsqueeze(
        1)
    real_indices_idx = torch.nonzero(attention_mask_in_length.flatten(), as_tuple=False).flatten()
    seqlens_in_batch = attention_mask_in_length.flatten()[real_indices_idx]
    indices = torch.nonzero(attention_mask_2d.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
    # TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
    # bool mask, then call nonzero to get the indices, then index with those. The indices is @dim
    # times larger than it needs to be, wasting memory. It's faster and more memory-efficient to
    # index with integer indices. Moreover, torch's index is a bit slower than it needs to be,
    # so we write custom forward and backward to make it a bit faster.
    return (
        index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


def pad_input(hidden_states, indices, batch, seqlen):
    """
    Arguments:
        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
        indices: (total_nnz), the indices that represent the non-masked tokens of the original padded input sequence.
        batch: int, batch size for the padded sequence.
        seqlen: int, maximum sequence length for the padded sequence.
    Return:
        hidden_states: (batch, seqlen, ...)
    """
    dim = hidden_states.shape[-1]
    # output = torch.zeros((batch * seqlen), dim, device=hidden_states.device, dtype=hidden_states.dtype)
    # output[indices] = hidden_states
    output = index_put_first_axis(hidden_states, indices, batch * seqlen)
    return rearrange(output, "(b s) ... -> b s ...", b=batch)