test_pipe.py 30.3 KB
Newer Older
Tom Birch's avatar
Tom Birch committed
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
# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
#
# This source code is licensed under the BSD license found in the
# LICENSE file in the root directory of this source tree.

# Copyright 2019 Kakao Brain
#
# 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 collections import OrderedDict
from copy import deepcopy
import os
import time

import pytest
import torch
from torch import nn

29
30
31
32
33
from fairscale.nn.model_parallel.initialize import (
    destroy_model_parallel,
    get_pipeline_parallel_group,
    initialize_model_parallel,
)
34
35
from fairscale.nn.pipe import AsyncPipe, LazyModule, MultiProcessPipe
from fairscale.nn.pipe.types import PipelineStyle
36
from fairscale.utils.testing import get_worker_map, set_random_seed, torch_spawn, torch_version
Tom Birch's avatar
Tom Birch committed
37
38
39


@torch_spawn([2])
40
41
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def parameters(pipe_class):
Tom Birch's avatar
Tom Birch committed
42
    model = nn.Sequential(nn.Linear(1, 1))
43
    pipe = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=1)
Tom Birch's avatar
Tom Birch committed
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
    if torch.distributed.get_rank() == 0:
        assert list(pipe.parameters()) != []
    else:
        assert list(pipe.parameters()) == []


@torch_spawn([2])
@pytest.mark.skipif(not torch.cuda.is_available(), reason="cuda required")
def infiniband():
    if torch.distributed.get_rank() == 0:
        t = torch.Tensor(range(100)).cuda()
        torch.distributed.broadcast(t, 0)
    else:
        t = torch.empty(100).cuda()
        torch.distributed.broadcast(t, 0)

    assert torch.equal(t, torch.Tensor(range(100)).cuda())
    print(f"t on {torch.distributed.get_rank()} is {t}")


@torch_spawn([2])
@pytest.mark.skipif("OMPI_COMM_WORLD_RANK" not in os.environ, reason="mpi required")
@pytest.mark.skipif(not torch.cuda.is_available(), reason="cuda required")
def infiniband2():
    if torch.distributed.get_rank() == 0:
        t = torch.Tensor(range(100)).cuda()
70
        torch.distributed.send(t, 1, group=get_pipeline_parallel_group())
Tom Birch's avatar
Tom Birch committed
71
72
    else:
        t = torch.empty(100).cuda()
73
        torch.distributed.recv(t, 0, group=get_pipeline_parallel_group())
Tom Birch's avatar
Tom Birch committed
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

    assert torch.equal(t, torch.Tensor(range(100)).cuda())
    print(f"t on {torch.distributed.get_rank()} is {t}")


@torch_spawn([2])
@pytest.mark.skipif(not torch.cuda.is_available(), reason="cuda required")
def infiniband3():
    t = torch.Tensor(range(100)).cuda()
    torch.distributed.all_reduce(t, op=torch.distributed.ReduceOp.SUM)
    assert torch.equal(t, torch.Tensor(range(0, 200, 2)).cuda())


@torch_spawn([2])
@pytest.mark.skipif("OMPI_COMM_WORLD_RANK" not in os.environ, reason="mpi required")
def mpi():
    seed = 1234
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)

    torch.distributed.barrier()
    tensor_size = (1024, 1024, 10)
    torch.cuda.set_device(torch.distributed.get_rank())  # need to pin device or ucx gets unhappy

    if torch.distributed.get_rank() == 0:
        # t = torch.Tensor(range(10)).cuda(0)
        t = torch.rand(*tensor_size).cuda(0)
        torch.distributed.send(t, 1, tag=1234)
    else:
        t = torch.empty(*tensor_size).cuda(1)
        torch.distributed.recv(t, 0, tag=1234)
        t2 = torch.rand(*tensor_size).cuda(1)

        assert torch.equal(t, t2)


@torch_spawn([1])
111
112
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def public_attrs(pipe_class):
Tom Birch's avatar
Tom Birch committed
113
114
115
116
117
118
119
120
121
    class MyString:
        def __init__(self, value):
            self.value = value

        def __str__(self):
            return self.value

    model = nn.Sequential(nn.Linear(1, 1))

122
    pipe = pipe_class(model, balance=(1,), worker_map=get_worker_map(), chunks=42.000, checkpoint=MyString("always"),)
Tom Birch's avatar
Tom Birch committed
123
124
125
126
127
128
129
130
131
132

    assert pipe.balance == [1]
    assert pipe.chunks == 42
    assert isinstance(pipe.chunks, int)
    assert pipe.checkpoint == "always"
    assert isinstance(pipe.checkpoint, str)


@torch_spawn([2])
@pytest.mark.parametrize("balance", [[2], [1, 1]])
133
134
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def sequential_like(balance, pipe_class):
Tom Birch's avatar
Tom Birch committed
135
136
137
138
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)

    model = nn.Sequential(a, b)
139
    model = pipe_class(model, balance, worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
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

    if balance == [2]:
        if torch.distributed.get_rank() == 0:
            assert len(model) == 2
            assert list(model) == [a, b]

            assert model[0] is a
            assert model[1] is b
            with pytest.raises(IndexError):
                _ = model[2]

            assert model[-1] is b
            assert model[-2] is a
        else:
            assert len(model) == 0
            assert list(model) == []
    else:
        assert len(model) == 1
        if torch.distributed.get_rank() == 0:
            assert list(model) == [a]
            assert model[0] is a
            assert model[-1] is a
        else:
            assert list(model) == [b]
            assert model[0] is b
            assert model[-1] is b

        with pytest.raises(IndexError):
            _ = model[1]


@torch_spawn([1])
172
173
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def balance_wrong_length(pipe_class):
Tom Birch's avatar
Tom Birch committed
174
175
176
177
178
179
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)

    model = nn.Sequential(a, b)

    with pytest.raises(ValueError):
180
        pipe_class(model, balance=[1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
181
182

    with pytest.raises(ValueError):
183
        pipe_class(model, balance=[3], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
184
185
186


@torch_spawn([2])
187
188
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def balance_less_than_1(pipe_class):
Tom Birch's avatar
Tom Birch committed
189
190
191
192
193
194
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)

    model = nn.Sequential(a, b)

    with pytest.raises(ValueError):
195
        pipe_class(model, balance=[0, 2], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
196
197

    with pytest.raises(ValueError):
198
        pipe_class(model, balance=[-1, 3], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
199
200
201


@torch_spawn([1])
202
203
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def chunks_less_than_1(pipe_class):
Tom Birch's avatar
Tom Birch committed
204
205
206
    model = nn.Sequential(nn.Linear(1, 1))

    with pytest.raises(ValueError):
207
        pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=0)
Tom Birch's avatar
Tom Birch committed
208
209

    with pytest.raises(ValueError):
210
        pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=-1)
Tom Birch's avatar
Tom Birch committed
211
212
213


@torch_spawn([1])
214
215
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def too_few_devices(pipe_class):
Tom Birch's avatar
Tom Birch committed
216
217
218
219
    model = nn.Sequential(nn.Linear(1, 1), nn.Linear(1, 1), nn.Linear(1, 1), nn.Linear(1, 1))

    with pytest.raises(IndexError):
        # len(balance) > len(group.size())
220
        model = pipe_class(model, balance=[1, 1, 1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
221
222
223


@torch_spawn([1])
224
225
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def batch_size_indivisible(pipe_class):
Tom Birch's avatar
Tom Birch committed
226
    model = nn.Sequential(nn.Linear(1, 1))
227
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=4)
Tom Birch's avatar
Tom Birch committed
228
229
230
231
232
233
234
235
236

    with pytest.warns(None) as record:
        model(torch.rand(7, 1))

    # Indivisible batch size is legal.
    assert not record


@torch_spawn([1])
237
238
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def batch_size_small(pipe_class):
Tom Birch's avatar
Tom Birch committed
239
    model = nn.Sequential(nn.Linear(1, 1))
240
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=4)
Tom Birch's avatar
Tom Birch committed
241
242
243
244
245
246
247
248
249

    with pytest.warns(None) as record:
        model(torch.rand(2, 1))

    # Batch size smaller than chunks is legal.
    assert not record


@torch_spawn([1])
250
251
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def checkpoint_mode(pipe_class):
Tom Birch's avatar
Tom Birch committed
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
    def count_grad_fn(grad_fn, name, visited=set()):
        if grad_fn in visited:
            return 0
        visited.add(grad_fn)

        if grad_fn is None:
            return 0
        if grad_fn.__class__.__name__ == name:
            return 1

        counter = 0
        for next_grad_fn, _ in grad_fn.next_functions:
            counter += count_grad_fn(next_grad_fn, name, visited=visited)
        return counter

    model = nn.Sequential(nn.Linear(1, 1))
    input = torch.rand(2, 1)

270
271
    always = pipe_class(
        model, balance=[1], worker_map=get_worker_map(), chunks=2, checkpoint="always", pipelined_backward=False,
Tom Birch's avatar
Tom Birch committed
272
    )
273
274
    except_last = pipe_class(
        model, balance=[1], worker_map=get_worker_map(), chunks=2, checkpoint="except_last", pipelined_backward=False,
Tom Birch's avatar
Tom Birch committed
275
    )
276
277
    never = pipe_class(
        model, balance=[1], worker_map=get_worker_map(), chunks=2, checkpoint="never", pipelined_backward=False,
Tom Birch's avatar
Tom Birch committed
278
279
280
281
282
283
284
285
286
287
288
289
    )

    always_output = always(input)
    except_last_output = except_last(input)
    never_output = never(input)

    assert count_grad_fn(always_output.grad_fn, "CheckpointBackward") == 2
    assert count_grad_fn(except_last_output.grad_fn, "CheckpointBackward") == 1
    assert count_grad_fn(never_output.grad_fn, "CheckpointBackward") == 0


@torch_spawn([1])
290
291
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def checkpoint_mode_invalid(pipe_class):
Tom Birch's avatar
Tom Birch committed
292
293
294
    model = nn.Sequential(nn.Linear(1, 1))

    with pytest.raises(ValueError, match="checkpoint is not one of 'always', 'except_last', or 'never'"):
295
296
        pipe_class(
            model, balance=[1], worker_map=get_worker_map(), chunks=2, checkpoint="INVALID_CHECKPOINT",
Tom Birch's avatar
Tom Birch committed
297
298
299
300
        )


@torch_spawn([1])
301
302
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def checkpoint_mode_when_chunks_1(pipe_class):
Tom Birch's avatar
Tom Birch committed
303
304
305
    model = nn.Sequential(nn.Linear(1, 1))

    # All checkpoint modes are fine.
306
307
    pipe_class(
        model, balance=[1], worker_map=get_worker_map(), chunks=1, checkpoint="except_last",
308
    )
309
310
    pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=1, checkpoint="always")
    pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=1, checkpoint="never")
Tom Birch's avatar
Tom Birch committed
311
312
313


@torch_spawn([1])
314
315
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def checkpoint_eval(pipe_class):
Tom Birch's avatar
Tom Birch committed
316
    model = nn.Sequential(nn.Linear(1, 1))
317
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=2, pipelined_backward=False,)
Tom Birch's avatar
Tom Birch committed
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
    input = torch.rand(2, 1)

    def find_grad_fn(grad_fn, name):
        if grad_fn is None:
            return False
        if grad_fn.__class__.__name__ == name:
            return True
        for next_grad_fn, _ in grad_fn.next_functions:
            if find_grad_fn(next_grad_fn, name):
                return True
        return False

    model.train()
    train_output = model(input)
    assert find_grad_fn(train_output.grad_fn, "CheckpointBackward")
    assert find_grad_fn(train_output.grad_fn, "RecomputeBackward")

    model.eval()
    eval_output = model(input)
    assert not find_grad_fn(eval_output.grad_fn, "CheckpointBackward")
    assert not find_grad_fn(eval_output.grad_fn, "RecomputeBackward")


@torch_spawn([2])
342
@pytest.mark.xfail(torch_version() < (1, 6, 0), reason="Doesn't work on torch < 1.6.0", strict=True)
343
344
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def checkpoint_non_float_input(pipe_class):
Tom Birch's avatar
Tom Birch committed
345
346
347
348
349
350
351
352
353
    class ForkNonFloat(nn.Module):
        def forward(self, input):
            return (input * 2, torch.tensor([False]))

    class JoinNonFloat(nn.Module):
        def forward(self, input):
            return input[0] * 2

    model = nn.Sequential(ForkNonFloat(), JoinNonFloat())
354
355
    model = pipe_class(
        model, balance=[1, 1], worker_map=get_worker_map(), chunks=1, checkpoint="always", pipelined_backward=False,
Tom Birch's avatar
Tom Birch committed
356
357
358
359
360
361
362
    )

    input = torch.rand(1, requires_grad=True)
    output = model(input)
    if model.group.rank() == 1:
        # with torch.autograd.detect_anomaly():
        output.backward()
363
    elif pipe_class == MultiProcessPipe:
Tom Birch's avatar
Tom Birch committed
364
365
        model.back_helper(output)

366
367
    torch.distributed.barrier()

Tom Birch's avatar
Tom Birch committed
368
369

@torch_spawn([1])
370
371
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def no_grad(pipe_class):
Tom Birch's avatar
Tom Birch committed
372
    model = nn.Sequential(nn.Linear(1, 1))
373
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=2)
Tom Birch's avatar
Tom Birch committed
374
375
376
377
378
379
380
381
382
383
384
    input = torch.rand(2, 1)

    latent = None

    def hook(module, input, output):
        _ = module
        _ = input

        nonlocal latent
        latent = output

385
    partition = model.partitions[0]
386
    partition.module.register_forward_hook(hook)
Tom Birch's avatar
Tom Birch committed
387
388
389
390
391
392
393
394

    with torch.no_grad():
        model(input)

    assert latent.grad_fn is None


@torch_spawn([1])
395
396
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def exception(pipe_class):
Tom Birch's avatar
Tom Birch committed
397
398
399
400
401
402
403
404
    class ExpectedException(Exception):
        pass

    class Raise(nn.Module):
        def forward(self, *_):
            raise ExpectedException()

    model = nn.Sequential(Raise())
405
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=1)
Tom Birch's avatar
Tom Birch committed
406
407
408
409
410
411
412

    with pytest.raises(ExpectedException):
        model(torch.rand(1))


# FIXME(tom) should probably signal to all hosts in group to stop
@torch_spawn([4])
413
@pytest.mark.skipif(torch.cuda.is_available() and torch.cuda.device_count() < 4, reason="Not enough GPUs")
Tom Birch's avatar
Tom Birch committed
414
@pytest.mark.xfail(strict=True)
415
416
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def exception_early_stop_asap(pipe_class):
Tom Birch's avatar
Tom Birch committed
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
    """Even the first partitions have finished to process, the partition before
    the failed partition hould be killed as soon as possible.
    """

    class ExpectedExceptio(Exception):
        pass

    class Pass(nn.Module):
        def forward(self, x):
            return x

    counter = 0

    class Counter(nn.Module):
        def forward(self, x):
            time.sleep(0.1)

            nonlocal counter
            counter += 1

            return x

    class Raise(nn.Module):
        def forward(self, x):
            raise ExpectedException()

    model = nn.Sequential(Pass(), Pass(), Counter(), Raise())
444
    model = pipe_class(model, [1, 1, 1, 1], worker_map=get_worker_map(), chunks=3)
Tom Birch's avatar
Tom Birch committed
445
446
447
448
449
450
451
452
453

    with pytest.raises(ExpectedException):
        model(torch.rand(3))

    # If the early stop doesn't work, it would be 3 instead.
    assert counter == 2


@torch_spawn([1])
454
455
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def input_pair(pipe_class):
Tom Birch's avatar
Tom Birch committed
456
457
458
459
460
461
462
463
464
465
466
    class Two(nn.Module):
        def __init__(self):
            super().__init__()
            self.fc_a = nn.Linear(1, 1)
            self.fc_b = nn.Linear(1, 1)

        def forward(self, a_and_b):
            a, b = a_and_b
            return (self.fc_a(a), self.fc_b(b))

    model = nn.Sequential(Two())
467
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=2, pipelined_backward=False,)
Tom Birch's avatar
Tom Birch committed
468
469
470
471
472
473
474
475
476
477
478
479
480

    a = torch.rand(10, 1, requires_grad=True)
    b = torch.rand(10, 1, requires_grad=True)

    a_out, b_out = model((a, b))
    loss = (a_out + b_out).mean()
    loss.backward()

    assert a.grad is not None
    assert b.grad is not None


@torch_spawn([1])
481
482
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def input_singleton(pipe_class):
Tom Birch's avatar
Tom Birch committed
483
484
485
486
487
488
489
490
491
492
    class One(nn.Module):
        def __init__(self):
            super().__init__()
            self.fc = nn.Linear(1, 1)

        def forward(self, only_a):
            (a,) = only_a
            return (self.fc(a),)

    model = nn.Sequential(One())
493
    model = pipe_class(model, balance=[1], worker_map=get_worker_map(), chunks=2, pipelined_backward=False,)
Tom Birch's avatar
Tom Birch committed
494
495
496
497
498
499
500
501
502
503
504
505

    a = torch.rand(10, 1, requires_grad=True)

    (a_out,) = model((a,))
    loss = a_out.mean()
    loss.backward()

    assert all(p.grad is not None for p in model.parameters())
    assert a.grad is not None


@torch_spawn([1])
506
507
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def input_varargs(pipe_class):
Tom Birch's avatar
Tom Birch committed
508
    model = nn.Sequential(nn.Linear(1, 1))
509
    model = pipe_class(model, balance=[1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
510
511
512
513
514
515
516
517
518
519

    a = torch.rand(1)
    b = torch.rand(1)

    # TypeError: forward() takes 2 positional arguments but 3 were given
    with pytest.raises(TypeError):
        model(a, b)


@torch_spawn([1])
520
521
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def non_tensor(pipe_class):
Tom Birch's avatar
Tom Birch committed
522
523
524
525
526
    class NonTensor(nn.Module):
        def forward(self, _):
            return "hello"

    model = nn.Sequential(NonTensor())
527
    model = pipe_class(model, balance=[1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
528
529
530
531
532
533
534
535
536
537
538
539
    x = torch.rand(1)

    # TypeError: expected Tensor as element 0 in argument 0, but got str
    with pytest.raises(TypeError):
        model(x)

    # TypeError: expected Tensor to scatter, but got str
    with pytest.raises(TypeError):
        model("hello")


@torch_spawn([1])
540
541
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def non_tensor_tuple(pipe_class):
Tom Birch's avatar
Tom Birch committed
542
543
544
545
546
    class NonTensorTuple(nn.Module):
        def forward(self, x):
            return (x, "hello")

    model = nn.Sequential(NonTensorTuple())
547
    model = pipe_class(model, balance=[1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
548
549
550
551
552
553
554
555
556
557
558
559
560
561
    x = torch.rand(1)

    # TypeError: CheckpointBackward.forward: expected Variable (got str) for return value 1
    with pytest.raises(TypeError):
        model(x)

    # TypeError: expected Tensor to scatter, but got str
    with pytest.raises(TypeError):
        model((x, "hello"))


@torch_spawn([1])
@pytest.mark.parametrize("checkpoint", ["never", "always", "except_last"])
@pytest.mark.parametrize("lazy", [True, False])
562
563
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def deferred_batch_norm(checkpoint, lazy, pipe_class):
Tom Birch's avatar
Tom Birch committed
564
565
566
567
    bn = nn.BatchNorm2d(3)
    pipe_bn = deepcopy(bn)
    pipe_fn = lambda: pipe_bn  # noqa: E731
    if lazy:
568
        model = [LazyModule(pipe_fn)]
Tom Birch's avatar
Tom Birch committed
569
570
    else:
        model = nn.Sequential(pipe_bn)
571
572
    pipe = pipe_class(
        model, balance=[1], worker_map=get_worker_map(), chunks=2, checkpoint=checkpoint, deferred_batch_norm=True,
Tom Birch's avatar
Tom Birch committed
573
574
575
576
577
578
579
580
581
582
583
584
585
    )

    x = torch.rand(4, 3, 10, 10)
    pipe(x).mean().backward()
    bn(x).mean().backward()

    assert torch.allclose(pipe[0].running_mean, bn.running_mean, atol=1e-4)
    assert torch.allclose(pipe[0].running_var, bn.running_var, atol=1e-4)


@torch_spawn([1])
@pytest.mark.parametrize("checkpoint", ["never", "always"])
@pytest.mark.parametrize("lazy", [True, False])
586
587
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def deferred_batch_norm_params(checkpoint, lazy, pipe_class):
Tom Birch's avatar
Tom Birch committed
588
589
590
591
    bn = nn.BatchNorm2d(3)
    pipe_bn = deepcopy(bn)
    pipe_fn = lambda: pipe_bn  # noqa: E731
    if lazy:
592
        model = [LazyModule(pipe_fn)]
Tom Birch's avatar
Tom Birch committed
593
594
    else:
        model = nn.Sequential(pipe_bn)
595
596
    pipe = pipe_class(
        model, balance=[1], worker_map=get_worker_map(), chunks=1, checkpoint=checkpoint, deferred_batch_norm=True,
Tom Birch's avatar
Tom Birch committed
597
598
599
600
601
602
603
604
605
606
607
608
609
    )

    x = torch.rand(4, 3, 10, 10)
    pipe(x).mean().backward()
    bn(x).mean().backward()

    assert pipe[0].weight.grad is not None
    assert pipe[0].bias.grad is not None

    assert torch.allclose(pipe[0].weight.grad, bn.weight.grad, atol=1e-4)
    assert torch.allclose(pipe[0].bias.grad, bn.bias.grad, atol=1e-4)


610
@torch_spawn([4])
611
612
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def devices(pipe_class):
Tom Birch's avatar
Tom Birch committed
613
614
615
616
617
618
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)
    c = nn.Linear(1, 1)

    # There are extra two ranks.
    model = nn.Sequential(a, b, c)
619
    model = pipe_class(model, [1, 1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
620
621

    # Extra devices must be discarded.
622
    if model.group.rank() == 3:
Tom Birch's avatar
Tom Birch committed
623
624
625
626
        assert model.pipeline is None


@torch_spawn([2])
627
628
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def partitions(pipe_class):
Tom Birch's avatar
Tom Birch committed
629
630
631
632
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)

    model = nn.Sequential(a, b)
633
    model = pipe_class(model, [1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
634

635
    assert isinstance(model.partitions, list)
Tom Birch's avatar
Tom Birch committed
636
    assert len(model) == 1
637
    assert isinstance(model.partitions[0].module, nn.Sequential)
Tom Birch's avatar
Tom Birch committed
638

639
640
641
642
    if model.group.rank() == 0:
        assert "0.0.weight" in model.state_dict()
    else:
        assert "0.1.weight" in model.state_dict()
Tom Birch's avatar
Tom Birch committed
643
644
645
646


@torch_spawn([2])
@pytest.mark.skipif(not torch.cuda.is_available(), reason="cuda required")
647
648
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def deny_moving(pipe_class):
Tom Birch's avatar
Tom Birch committed
649
650
651
652
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)

    model = nn.Sequential(a, b)
653
    model = pipe_class(model, [1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670

    model.cuda()
    model.cpu()
    model.to(torch.device("cuda"))
    model.to(0)
    model.to("cuda")
    model.to(device=0)
    model.to(torch.rand(1))
    model.to(tensor=torch.rand(1))

    # Casting is allowed.
    model.half()
    model.to(torch.double)
    model.to(dtype=torch.float)


@torch_spawn([1])
671
672
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def empty_module(pipe_class):
Tom Birch's avatar
Tom Birch committed
673
674
    # Empty sequential module is not illegal.
    model = nn.Sequential()
675
    model = pipe_class(model, [], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
676
677
678
679

    assert model(torch.tensor([42])) == torch.tensor([42])
    assert model((torch.tensor([42]),)) == (torch.tensor([42]),)

680
    # But only tensor or tensors is legal in MultiProcessPipe.
Tom Birch's avatar
Tom Birch committed
681
682
683
684
685
686

    with pytest.raises(TypeError):
        model(42)


@torch_spawn([2])
687
688
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def named_children(pipe_class):
Tom Birch's avatar
Tom Birch committed
689
690
691
692
    a = nn.Linear(1, 1)
    b = nn.Linear(1, 1)

    model = nn.Sequential(OrderedDict([("a", a), ("b", b)]))
693
    model = pipe_class(model, [1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
694
695

    names = set(n for n, _ in model.named_modules())
696
697
698
699
    if model.group.rank() == 0:
        assert "0.a" in names
    else:
        assert "0.b" in names
Tom Birch's avatar
Tom Birch committed
700

701
    # MultiProcessPipe doesn't support __getattr__. Unlike nn.Sequential, MultiProcessPipe requires
Tom Birch's avatar
Tom Birch committed
702
703
704
705
706
707
    # several methods in its namespace.
    with pytest.raises(AttributeError):
        model.a


@torch_spawn([1])
708
709
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def recommend_auto_balance(pipe_class):
Tom Birch's avatar
Tom Birch committed
710
711
    with pytest.raises(ValueError, match="fairscale.nn.pipe.balance"):
        # balance is required
712
        pipe_class(nn.Sequential())
Tom Birch's avatar
Tom Birch committed
713
714
715

    with pytest.raises(ValueError, match="fairscale.nn.pipe.balance"):
        # module and sum of balance have differen length (module: 0, sum of balance: 1)
716
        pipe_class(nn.Sequential(), [1])
Tom Birch's avatar
Tom Birch committed
717
718
719

    with pytest.raises(ValueError, match="fairscale.nn.pipe.balance"):
        # module and sum of balance have different length (module: 2, sum of balance: 1)
720
        pipe_class(nn.Sequential(nn.Linear(1, 1), nn.Linear(1, 1)), [1])
Tom Birch's avatar
Tom Birch committed
721
722
723


@torch_spawn([2])
724
725
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def lazy_construction(pipe_class):
Tom Birch's avatar
Tom Birch committed
726
727
728
729
730
731
732
733
734
735
736
737
    init_count = 0

    class Custom(nn.Module):
        def __init__(self):
            super(Custom, self).__init__()
            nonlocal init_count
            init_count += 1

        def forward(self, x):
            return x

    model = [
738
739
740
741
        LazyModule(lambda: Custom()),
        LazyModule(lambda: Custom()),
        LazyModule(lambda: Custom()),
        LazyModule(lambda: Custom()),
Tom Birch's avatar
Tom Birch committed
742
743
    ]

744
    pipe = pipe_class(model, balance=[2, 2], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
745
746
747
748
749
750
751
752

    assert isinstance(pipe[0], Custom)
    assert isinstance(pipe[1], Custom)
    assert len(pipe) == 2
    assert init_count == 2


@torch_spawn([2])
753
@pytest.mark.skipif("OMPI_COMM_WORLD_RANK" in os.environ, reason="doesn't apply to mpi")
754
755
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def missing_worker_map(pipe_class):
Tom Birch's avatar
Tom Birch committed
756
757
    model = nn.Sequential(nn.ReLU(), nn.ReLU())

758
    with pytest.raises(ValueError, match="'RpcTransport' requires 'worker_map' to be set"):
759
        pipe_class(model, [1, 1])
Tom Birch's avatar
Tom Birch committed
760
761
762
763


@torch_spawn([2])
@pytest.mark.skip(reason="currently broken")
764
765
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def verify_module_duplicate_parameters_on_distinct_partitions(pipe_class):
Tom Birch's avatar
Tom Birch committed
766
767
768
769
770
771
772
773
774
775
    class Surrogate(nn.Module):
        def __init__(self, module):
            super().__init__()
            self.module = module

    conv = nn.Conv2d(3, 3, 1)
    model = nn.Sequential(Surrogate(conv), Surrogate(conv))

    # FIXME(tom) can't have duplicate params with separate processes
    with pytest.raises(ValueError, match="module with duplicate parameters on distinct devices is not supported"):
776
        pipe_class(model, [1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
777
778
779


@torch_spawn([4])
780
781
@pytest.mark.parametrize("pipe_class", [MultiProcessPipe, AsyncPipe])
def pipelined_backward(pipe_class):
Tom Birch's avatar
Tom Birch committed
782
783
784
785
    model = nn.Sequential(nn.ReLU(), nn.ReLU())

    destroy_model_parallel()
    initialize_model_parallel(1, 4)
786
    pipe = pipe_class(model, [1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
787
788
789
790
791

    assert pipe.pipelined_backward is False

    destroy_model_parallel()
    initialize_model_parallel(2, 2)
792
    pipe = pipe_class(model, [1, 1], worker_map=get_worker_map())
Tom Birch's avatar
Tom Birch committed
793
794

    assert pipe.pipelined_backward is True
795
796
797
798
799
800


@torch_spawn([4])
def async_event_loop():

    model = nn.Sequential(nn.Linear(10, 10), nn.ReLU(), nn.Linear(10, 10), nn.ReLU())
801
    pipe = AsyncPipe(model, [1, 1, 1, 1], worker_map=get_worker_map(), chunks=10)
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816

    inputs = torch.rand(100, 10)

    output = pipe(inputs)
    if pipe.final_stage:
        loss = output.mean()
        loss.backward()


@torch_spawn([4])
def reuse_lazy():
    if False:  # speed
        reused = LazyModule(lambda: nn.Linear(10, 10))
        model = [reused, nn.Linear(10, 10), nn.ReLU(), reused, nn.ReLU(), reused, nn.ReLU()]
        # model = [reused, reused, nn.Linear(10, 10), nn.ReLU(), reused, reused, nn.ReLU(), reused, reused, nn.ReLU()]
817
        pipe = AsyncPipe(model, [3, 1, 1], worker_map=get_worker_map())
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
        pipe.eval()
        output = pipe(torch.rand(10))

        print(f"output on {pipe.group.rank()}, {output}")
        torch.distributed.barrier()

    set_random_seed(1234)
    # test both foward
    reused = nn.Linear(10, 10)
    layers = [reused, nn.Linear(10, 10), nn.ReLU(), reused, nn.ReLU(), reused, nn.ReLU()]
    model = nn.Sequential(*layers)
    model.eval()

    set_random_seed(1234)
    # ensure identical weights but no sharing between model and pipe
    reused = nn.Linear(10, 10)
    layers = [reused, nn.Linear(10, 10), nn.ReLU(), reused, nn.ReLU(), reused, nn.ReLU()]
835
    pipe = AsyncPipe(layers, [3, 1, 1], worker_map=get_worker_map())
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
    pipe.eval()
    model_optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
    pipe_optimizer = torch.optim.SGD(pipe.parameters(), lr=0.01, momentum=0.9) if len(list(pipe.parameters())) else None
    inputs = torch.rand(10)
    if False:  # speed
        model_out = model(inputs)
        pipe_out = pipe(inputs)

        torch.distributed.barrier()

        if pipe.final_stage:
            assert torch.equal(model_out, pipe_out)

    model.train()
    pipe.train()
    model_out = model(inputs)
    pipe_out = pipe(inputs)
    if pipe.final_stage:
        pipe_loss = pipe_out.mean()
        pipe_loss.backward()

    model_loss = model_out.mean()
    model_loss.backward()

    model_optimizer.step()
    if pipe_optimizer:
        pipe_optimizer.step()

    model.eval()
    pipe.eval()
    model_out = model(inputs)
    pipe_out = pipe(inputs)

    print(f"before barrier on {torch.distributed.get_rank()}")
    torch.distributed.barrier()
    print(f"after barrier on {torch.distributed.get_rank()}")

    if pipe.final_stage:
        assert torch.equal(model_out, pipe_out)


def test_instantiate_partition():
    from fairscale.nn.pipe.async_schedule import Location
879
    from fairscale.nn.pipe.multiprocess_pipe import instantiate_partition
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894

    class FakeGroup:
        def __init__(self, rank, size):
            self._rank = rank
            self._size = size

        def rank(self):
            return self._rank

        def size(self):
            return self._size

    def check_partitions(model, balance, expected_order, expected_ranks):
        """Check the instantiated model matches expectation of order and rank
        model: a list of modules or an nn.Sequential
895
        balance: the balance argument to MultiProcessPipe
896
897
898
899
900
901
902
903
904
905
906
        expected_order: the index of modules in `model` in the order they will
            be executed, grouped by nn.Sequential
        expected_rank: the rank that each module will be executed on
        """

        invocations = []
        invocation_wrapper = dict()

        # Collect `Invocation` and `Invocation` -> `ModuleWrapper` mapping from
        # instantiated model
        for rank in range(len(balance)):
907
            instantiated = instantiate_partition(
908
                model, balance, FakeGroup(rank, len(balance)), PipelineStyle.AsyncSchedule
909
            )
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
            for part in instantiated:
                assert isinstance(part.module, nn.Sequential)
                for inv in part.invocations:
                    invocations.append(inv)
                    invocation_wrapper[inv] = part

        modules = []
        prev = None
        current = Location(0, 0)
        ranks = []

        for order, inv in enumerate(sorted(invocations, key=lambda x: x.order)):
            # Check integrity of Location chain
            assert inv.order == order
            assert inv.source == prev
            assert inv.this == current
            prev = inv.this
            current = inv.dest
            modules.append(list(invocation_wrapper[inv].module.children()))
            ranks.append(inv.this.stage)

        # assert len(modules) == len(expected_order)
        for left, right in zip(modules, expected_order):
            assert len(left) == len(right), f"{right}"
            assert list(map(id, left)) == list(map(id, (model[e] for e in right))), f"{right}"

        assert ranks == expected_ranks

    reused = nn.Linear(20, 20)
    model = [reused, nn.Linear(10, 10), nn.ReLU(), reused, nn.ReLU(), reused, nn.ReLU()]
    balance = [3, 1, 1]

    check_partitions(
        model, balance, expected_order=[[0], [1, 2], [0], [4], [0], [6]], expected_ranks=[0, 0, 0, 1, 0, 2]
    )

    reused2 = nn.Linear(5, 5)
    model = [reused, reused2, nn.Linear(10, 10), nn.ReLU(), reused, reused2, nn.ReLU(), reused, reused2, nn.ReLU()]
    balance = [4, 1, 1]

    check_partitions(
        model,
        balance,
        expected_order=[[0], [1], [2, 3], [0], [1], [6], [0], [1], [9]],
        expected_ranks=[0, 0, 0, 0, 0, 1, 0, 0, 2],
    )

    reused2 = nn.Linear(5, 5)
    model = [
        nn.Linear(10, 10),
        reused,
        nn.Linear(10, 10),
        nn.ReLU(),
        reused,
        reused2,
        nn.ReLU(),
        reused,
        reused2,
        nn.ReLU(),
    ]
    # 0 1 2 3 1 5 6 1 5 9
    balance = [4, 2, 1]

    check_partitions(
        model,
        balance,
        expected_order=[[0], [1], [2, 3], [1], [5], [6], [1], [5], [9]],
        expected_ranks=[0, 0, 0, 0, 1, 1, 0, 1, 2],
    )