File size: 9,185 Bytes
5238467
 
 
 
 
 
 
 
 
 
 
6ec60d5
 
5238467
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ec60d5
 
5238467
6ec60d5
 
 
 
 
 
 
 
 
 
 
 
 
5238467
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ec60d5
 
 
 
 
 
 
 
 
 
5238467
6ec60d5
5238467
6ec60d5
 
 
 
5238467
6ec60d5
 
 
 
5238467
6ec60d5
 
 
5238467
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ec60d5
5238467
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

from itertools import product

import pytest
import torch

from audiocraft.modules.transformer import (
    StreamingMultiheadAttention, StreamingTransformer, set_efficient_attention_backend)


def test_transformer_causal_streaming():
    torch.manual_seed(1234)

    for context, custom in product([None, 10], [False, True]):
        # Test that causality and receptive fields are properly handled.
        # looking at the gradients
        tr = StreamingTransformer(
            16, 4, 1 if context else 2,
            causal=True, past_context=context, custom=custom,
            dropout=0.)
        steps = 20
        for k in [0, 10, 15, 19]:
            x = torch.randn(4, steps, 16, requires_grad=True)
            y = tr(x)
            y[:, k].abs().sum().backward()
            if k + 1 < steps:
                assert torch.allclose(x.grad[:, k + 1:], torch.tensor(0.)), x.grad[:, k + 1:].norm()
            assert not torch.allclose(x.grad[:, :k + 1], torch.tensor(0.)), x.grad[:, :k + 1].norm()
            if context is not None and k > context:
                limit = k - context - 1
                assert torch.allclose(x.grad[:, :limit],
                                      torch.tensor(0.)), x.grad[:, :limit].norm()

        # Now check that streaming gives the same result at batch eval.
        x = torch.randn(4, steps, 16)
        y = tr(x)
        ys = []
        with tr.streaming():
            for k in range(steps):
                chunk = x[:, k:k + 1, :]
                ys.append(tr(chunk))
        y_stream = torch.cat(ys, dim=1)
        delta = torch.norm(y_stream - y) / torch.norm(y)
        assert delta < 1e-6, delta


def test_transformer_vs_pytorch():
    torch.manual_seed(1234)
    # Check that in the non causal setting, we get the same result as
    # PyTorch Transformer encoder.
    for custom in [False, True]:
        tr = StreamingTransformer(
            16, 4, 2,
            causal=False, custom=custom, dropout=0., positional_scale=0.)
        layer = torch.nn.TransformerEncoderLayer(16, 4, dropout=0., batch_first=True)
        tr_ref = torch.nn.TransformerEncoder(layer, 2)
        tr.load_state_dict(tr_ref.state_dict())

        x = torch.randn(4, 20, 16)
        y = tr(x)
        y2 = tr_ref(x)
        delta = torch.norm(y2 - y) / torch.norm(y)
        assert delta < 1e-6, delta


def test_streaming_api():
    tr = StreamingTransformer(16, 4, 2, causal=True, dropout=0.)
    tr.eval()
    steps = 12
    x = torch.randn(1, steps, 16)

    with torch.no_grad():
        with tr.streaming():
            _ = tr(x[:, :1])
            state = {k: v.clone() for k, v in tr.get_streaming_state().items()}
            y = tr(x[:, 1:2])
            tr.set_streaming_state(state)
            y2 = tr(x[:, 1:2])
            assert torch.allclose(y, y2), (y - y2).norm()
            assert tr.flush() is None


def test_memory_efficient():
    torch.manual_seed(1234)
    for backend in ['torch', 'xformers']:
        set_efficient_attention_backend(backend)

        tr = StreamingTransformer(
            16, 4, 2, custom=True, dropout=0., layer_scale=0.1)
        tr_mem_efficient = StreamingTransformer(
            16, 4, 2, dropout=0., memory_efficient=True, layer_scale=0.1)
        tr_mem_efficient.load_state_dict(tr.state_dict())
        tr.eval()
        steps = 12
        x = torch.randn(3, steps, 16)

        with torch.no_grad():
            y = tr(x)
            y2 = tr_mem_efficient(x)
            assert torch.allclose(y, y2), ((y - y2).norm(), backend)


def test_attention_as_float32():
    torch.manual_seed(1234)
    cases = [
        {'custom': True},
        {'custom': False},
    ]
    for case in cases:
        tr = StreamingTransformer(16, 4, 2, dropout=0., dtype=torch.bfloat16, **case)
        tr_float32 = StreamingTransformer(
            16, 4, 2, dropout=0., attention_as_float32=True, dtype=torch.bfloat16, **case)
        if not case['custom']:
            # we are not using autocast here because it doesn't really
            # work as expected on CPU, so we have to manually cast the weights of the MHA.
            for layer in tr_float32.layers:
                layer.self_attn.mha.to(torch.float32)
        tr_float32.load_state_dict(tr.state_dict())
        steps = 12
        x = torch.randn(3, steps, 16, dtype=torch.bfloat16)

        with torch.no_grad():
            y = tr(x)
            y2 = tr_float32(x)
            assert not torch.allclose(y, y2), (y - y2).norm()


@torch.no_grad()
def test_streaming_memory_efficient():
    torch.manual_seed(1234)
    for backend in ['torch', 'xformers']:
        set_efficient_attention_backend(backend)
        tr = StreamingTransformer(16, 4, 2, causal=True, dropout=0., custom=True)
        tr_mem_efficient = StreamingTransformer(
            16, 4, 2, dropout=0., memory_efficient=True, causal=True)
        tr.load_state_dict(tr_mem_efficient.state_dict())
        tr.eval()
        tr_mem_efficient.eval()
        steps = 12
        x = torch.randn(3, steps, 16)

        ref = tr(x)

        with tr_mem_efficient.streaming():
            outs = []
            # frame_sizes = [2] + [1] * (steps - 2)
            frame_sizes = [1] * steps

            for frame_size in frame_sizes:
                frame = x[:, :frame_size]
                x = x[:, frame_size:]
                outs.append(tr_mem_efficient(frame))

        out = torch.cat(outs, dim=1)
        delta = torch.norm(out - ref) / torch.norm(out)
        assert delta < 1e-6, delta


def test_cross_attention():
    torch.manual_seed(1234)
    for norm_first in [True, False]:
        m = StreamingTransformer(
            16, 4, 2, cross_attention=False, norm_first=norm_first, dropout=0., custom=True)
        m_cross = StreamingTransformer(
            16, 4, 2, cross_attention=True, norm_first=norm_first, dropout=0., custom=True)
        m_cross.load_state_dict(m.state_dict(), strict=False)
        x = torch.randn(2, 5, 16)
        cross_x = torch.randn(2, 3, 16)
        y_ref = m(x)
        y_cross_zero = m_cross(x, cross_attention_src=0 * cross_x)
        # With norm_first, the two should be exactly yhe same,
        # but with norm_first=False, we get 2 normalization in a row
        # and the epsilon value leads to a tiny change.
        atol = 0. if norm_first else 1e-6
        print((y_ref - y_cross_zero).norm() / y_ref.norm())
        assert torch.allclose(y_ref, y_cross_zero, atol=atol)

        # We now expect a difference even with a generous atol of 1e-2.
        y_cross = m_cross(x, cross_attention_src=cross_x)
        assert not torch.allclose(y_cross, y_cross_zero, atol=1e-2)

        with pytest.raises(AssertionError):
            _ = m_cross(x)
            _ = m(x, cross_attention_src=cross_x)


def test_cross_attention_compat():
    torch.manual_seed(1234)
    num_heads = 2
    dim = num_heads * 64
    with pytest.raises(AssertionError):
        StreamingMultiheadAttention(dim, num_heads, causal=True, cross_attention=True)

    cross_attn = StreamingMultiheadAttention(
        dim, num_heads, dropout=0, cross_attention=True, custom=True)
    ref_attn = torch.nn.MultiheadAttention(dim, num_heads, dropout=0, batch_first=True)

    # We can load the regular attention state dict
    # so we have compat when loading old checkpoints.
    cross_attn.load_state_dict(ref_attn.state_dict())

    queries = torch.randn(3, 7, dim)
    keys = torch.randn(3, 9, dim)
    values = torch.randn(3, 9, dim)

    y = cross_attn(queries, keys, values)[0]
    y_ref = ref_attn(queries, keys, values)[0]
    assert torch.allclose(y, y_ref, atol=1e-7), (y - y_ref).norm() / y_ref.norm()

    # Now let's check that streaming is working properly.
    with cross_attn.streaming():
        ys = []
        for step in range(queries.shape[1]):
            ys.append(cross_attn(queries[:, step: step + 1], keys, values)[0])
    y_streaming = torch.cat(ys, dim=1)
    assert torch.allclose(y_streaming, y, atol=1e-7)


def test_repeat_kv():
    torch.manual_seed(1234)
    num_heads = 8
    kv_repeat = 4
    dim = num_heads * 64
    with pytest.raises(AssertionError):
        mha = StreamingMultiheadAttention(
            dim, num_heads, causal=True, kv_repeat=kv_repeat, cross_attention=True)
        mha = StreamingMultiheadAttention(
            dim, num_heads, causal=True, kv_repeat=kv_repeat)
    mha = StreamingMultiheadAttention(
        dim, num_heads, causal=True, kv_repeat=kv_repeat, custom=True)
    x = torch.randn(4, 18, dim)
    y = mha(x, x, x)[0]
    assert x.shape == y.shape


def test_qk_layer_norm():
    torch.manual_seed(1234)
    tr = StreamingTransformer(
        16, 4, 2, custom=True, dropout=0., qk_layer_norm=True, bias_attn=False)
    steps = 12
    x = torch.randn(3, steps, 16)
    y = tr(x)

    tr = StreamingTransformer(
        16, 4, 2, custom=True, dropout=0., qk_layer_norm=True, cross_attention=True)
    z = torch.randn(3, 21, 16)
    y = tr(x, cross_attention_src=z)
    assert y.shape == x.shape