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
255
256
257
258
259
260
261
262
263
264
|
# Owner(s): ["oncall: distributed"]
import copy
import functools
import itertools
from typing import List, Union
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
from torch.distributed._composable import checkpoint, replicate
from torch.distributed.fsdp import fully_shard
from torch.distributed.fsdp._fully_shard._fsdp_param_group import (
RegisterPostBackwardFunction,
)
from torch.testing._internal.common_distributed import skip_if_lt_x_gpu
from torch.testing._internal.common_fsdp import (
check_sharded_parity,
FSDPTest,
MLP,
patch_reduce_scatter,
patch_register_post_backward_hook_backward,
reduce_scatter_with_assert,
)
from torch.testing._internal.common_utils import run_tests
class TestFullyShardFrozen(FSDPTest):
@property
def world_size(self) -> int:
return min(4, torch.cuda.device_count())
@skip_if_lt_x_gpu(2)
def test_train_mixed_requires_grad_per_group(self):
"""
Tests training parity with DDP when mixing frozen and non-frozen
parameters in the same FSDP communication group. This checks that
the reduce-scatters reduce the expected numel and that they are called
via the custom autograd function backward (i.e. that they are not
delayed until the end of backward).
"""
self.run_subtests(
{
"reshard_after_forward": [False, True, 2],
"use_activation_checkpointing": [False, True],
"freeze_after_init": [False, True],
},
self._test_train_mixed_requires_grad_per_group,
)
def _test_train_mixed_requires_grad_per_group(
self,
reshard_after_forward: Union[bool, int],
use_activation_checkpointing: bool,
freeze_after_init: bool,
):
torch.manual_seed(42)
num_mlps, lin_dim = (3, 32)
model = nn.Sequential(
*[MLP(lin_dim, torch.device("cpu")) for _ in range(num_mlps)]
)
# Train biases only (e.g. like BitFit)
if not freeze_after_init:
for param_name, param in model.named_parameters():
if "bias" not in param_name:
param.requires_grad_(False)
ref_model = replicate(
copy.deepcopy(model).cuda(),
device_ids=[self.rank],
find_unused_parameters=freeze_after_init,
)
ref_optim = torch.optim.Adam(ref_model.parameters(), lr=1e-2)
for mlp in model:
if use_activation_checkpointing:
checkpoint(mlp)
fully_shard(mlp, reshard_after_forward=reshard_after_forward)
fully_shard(model, reshard_after_forward=reshard_after_forward)
optim = torch.optim.Adam(model.parameters(), lr=1e-2)
orig_reduce_scatter = dist.reduce_scatter_tensor
if freeze_after_init:
for param_name, param in itertools.chain(
model.named_parameters(), ref_model.named_parameters()
):
if "bias" not in param_name:
param.requires_grad_(False)
for mlp in model:
assert isinstance(mlp, MLP), (
"The reduce-scatter numel check assumes the model consists of "
f"only the same MLP class but got {type(mlp)}"
)
expected_numel = sum(
p._local_tensor.numel()
for n, p in model[0].named_parameters()
if "bias" in n
)
def assert_fn(output: torch.Tensor):
self.assertEqual(output.numel(), expected_numel)
reduce_scatter = functools.partial(
reduce_scatter_with_assert, self, orig_reduce_scatter, assert_fn
)
orig_backward = RegisterPostBackwardFunction.backward
backward_count = 0
def backward_with_count(*args, **kwargs):
nonlocal backward_count
backward_count += 1
return orig_backward(*args, **kwargs)
torch.manual_seed(42 + self.rank + 1)
device = torch.device("cuda")
with patch_reduce_scatter(
reduce_scatter
), patch_register_post_backward_hook_backward(backward_with_count):
for iter_idx in range(10):
inp = torch.randn((8, lin_dim), device=device)
losses: List[torch.Tensor] = []
for _model, _optim in ((ref_model, ref_optim), (model, optim)):
_optim.zero_grad(set_to_none=(iter_idx % 2 == 0))
losses.append(_model(inp).sum())
losses[-1].backward()
_optim.step()
check_sharded_parity(self, ref_model, model)
self.assertEqual(losses[0], losses[1])
# Check that the post-backward hooks ran through the autograd
# backward, not the final callback (except possibly that of the
# first MLP, which does not have an input that requires grad)
self.assertTrue(backward_count >= num_mlps - 1)
@skip_if_lt_x_gpu(2)
def test_train_mixed_requires_grad_across_groups(self):
"""
Tests training parity with DDP when mixing frozen and non-frozen
parameters across different FSDP communication groups, including
possibly unfreezing parameters.
"""
self.run_subtests(
{
"reshard_after_forward": [False, True, 2],
"unfreeze_params": [False, True],
},
self._test_train_mixed_requires_grad_across_groups,
)
def _test_train_mixed_requires_grad_across_groups(
self,
reshard_after_forward: Union[bool, int],
unfreeze_params: bool,
):
torch.manual_seed(42)
num_linears, lin_dim = (6, 32)
modules: List[nn.Module] = []
for _ in range(num_linears):
modules += [nn.Linear(lin_dim, lin_dim), nn.ReLU()]
model = nn.Sequential(*modules)
ref_model = replicate(
copy.deepcopy(model).cuda(),
device_ids=[self.rank],
find_unused_parameters=True,
)
for module in model.modules():
if isinstance(module, nn.Linear):
fully_shard(module, reshard_after_forward=reshard_after_forward)
ref_optim = torch.optim.Adam(ref_model.parameters(), lr=1e-2)
optim = torch.optim.Adam(model.parameters(), lr=1e-2)
orig_backward = RegisterPostBackwardFunction.backward
backward_count = 0
def _set_requires_grad(seq: nn.Module, requires_grad: bool):
for i in range(num_linears):
# Interleave frozen -> non-frozen -> ... linears
if i % 2 == 0:
for param in seq[i % 2].parameters():
param.requires_grad_(requires_grad)
def backward_with_count(*args, **kwargs):
nonlocal backward_count
backward_count += 1
return orig_backward(*args, **kwargs)
_set_requires_grad(model, False)
_set_requires_grad(ref_model, False)
num_iters, no_grad_iter_idx = (3, 1)
torch.manual_seed(42 + self.rank)
inp = torch.randn((8, lin_dim), device="cuda")
with patch_register_post_backward_hook_backward(backward_with_count):
for iter_idx in range(num_iters):
losses: List[torch.Tensor] = []
for _model, _optim in ((ref_model, ref_optim), (model, optim)):
# Unfreeze the parameters on the last step to emulate some
# kinds of fine-tuning
if unfreeze_params and iter_idx == num_iters - 1:
_set_requires_grad(model, True)
if iter_idx == no_grad_iter_idx:
with torch.no_grad():
losses.append(_model(inp).sum())
else:
losses.append(_model(inp).sum())
losses[-1].backward()
_optim.step()
_optim.zero_grad(set_to_none=(iter_idx % 2 == 0))
self.assertEqual(losses[0], losses[1])
# Check that the post-backward hooks ran through the autograd
# backward, not the final callback (except possibly that of the
# first linear, which does not have an input that requires grad)
self.assertTrue(backward_count >= num_linears - 1)
@skip_if_lt_x_gpu(2)
def test_multi_forward_mixed_requires_grad(self):
"""
Tests training parity with DDP when having trainable and frozen modules
that participate multiple times in forward.
"""
self.run_subtests(
{"reshard_after_forward": [True, False, 2]},
self._test_multi_forward_mixed_requires_grad,
)
def _test_multi_forward_mixed_requires_grad(
self,
reshard_after_forward: Union[bool, int],
):
class MultiForwardModule(nn.Module):
def __init__(self, device: torch.device):
super().__init__()
self.layer_0 = nn.Linear(5, 5, device=device)
self.layer_no_grad = nn.Linear(5, 5, device=device)
self.layer_with_grad = nn.Linear(5, 5, device=device)
self.layer_no_grad.requires_grad_(False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.layer_0(x)
for _ in range(3):
x = self.layer_no_grad(F.relu(self.layer_with_grad(x)))
# Make sure that calling the same layer multiple times
# works regardless whether gradient is enabled
with torch.no_grad():
x += F.relu(self.layer_with_grad(x))
return x
torch.manual_seed(42)
model = MultiForwardModule(torch.device("cpu"))
ref_model = replicate(copy.deepcopy(model).cuda(), device_ids=[self.rank])
ref_optim = torch.optim.Adam(ref_model.parameters(), lr=1e-2)
for module in model.modules():
if isinstance(module, nn.Linear):
fully_shard(module, reshard_after_forward=reshard_after_forward)
fully_shard(model, reshard_after_forward=reshard_after_forward)
optim = torch.optim.Adam(model.parameters(), lr=1e-2)
for iter_idx in range(10):
inp = torch.randn((8, 5), device="cuda")
losses: List[torch.Tensor] = []
for _model, _optim in ((ref_model, ref_optim), (model, optim)):
_optim.zero_grad(set_to_none=(iter_idx % 2 == 0))
losses.append(_model(inp).sum())
losses[-1].backward()
_optim.step()
self.assertEqual(losses[0], losses[1])
if __name__ == "__main__":
run_tests()
|
