CHANGELOG.md 15 KB
Newer Older
1
2
3
4
5
6
# Changelog
All notable changes to this project will be documented in this file.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
7
## NEXT - TBD
8
### Fixed
9
- FSDP: Fixed an pre-backward hook bug for certain type of models and FSDP config. [#833]
tmarkstrum's avatar
tmarkstrum committed
10
11

### Added
12
- LayerwiseMemoryTracker[feature][experimental] - This is a new experimental tool to help track, visualize and suggest fix for memory issues occurring during the forward/backward pass of your models. [#808]
13
14
15
- [FSDP]: limited support of shared weights between FSDP wrappers. This allows large parameter
          and gradient memory to be sharded despite being needed from different layers due to
          weight sharing. [#836]
tmarkstrum's avatar
tmarkstrum committed
16
17
18

## [0.4.1] - 2021-09-17
### Fixed
anj-s's avatar
anj-s committed
19
20
21
22
23
- FSDP: We don't attach post backward hooks for params that don't require grad. However in the hook
        triggered after the post backward hook, we assert on the POST_BACKWARD state which can only
        be set in the post backward hook. Modified the assert to account for the fact that the root
        FSDP module can have child modules with params that require grad and it can contain params
        that don't require grad and hence can fail the previous assert. [#761]
24
25
- FSDP: Fixed a bug when multiple backward pass is called within an iteration, parameters' sharding
        state might be incorrect. [#775]
26
27
- activation checkpoint: Ensure outputs of checkpointed modules only require grad if either
                         the input requires grad or if the parameters require grad. [#787]
Min Xu's avatar
Min Xu committed
28

29
30
31
32
- OSS: fix the broadcast_fp16 option, broken after a refactor, this flag was doing nothing (bugfix).[#795]
- OSS: update default device when refreshing the params, meaning that moving the model to GPU after
       the OSS wrap will not trigger warnings and slow the jobs (ease of use). [#786]

Min Xu's avatar
Min Xu committed
33
### Added
anj-s's avatar
anj-s committed
34
- FSDP: Added support for returning the original names of parameters when `named_parameters` is called on
35
        the module. To retrieve the orginal names of the parameters along with the params, you need to
anj-s's avatar
anj-s committed
36
37
38
        call `named_parameters` under the `summon_full_params` context when using flattened params or original
        params. If you are using original params (i.e flatten_params=False), calling `named_parameters` outside
        of the `summon_full_params` context will still return the original param names along with the local shards. [#755]
39
40
41
42
43
44
45
46
- FSDP: Ensure gradient reduction accumulates into the unsharded gradient tensor
        within a backwards pass. This matters when an FSDP module is called
        multiple times within a forward pass, and reduction is not deferred
        using activation checkpoint forward counters, bucketing or some other
        mechanism. [#784]
- activation checkpoint: Added a context manager to disable checkpoint in case the same wrapped module
                         needs to be checkpointed and not checkpointed in different parts of
                         the module forward pass. [#772]
tmarkstrum's avatar
tmarkstrum committed
47
48
49
- FSDP: Added a toggle with an environment variable ENABLE_NCCL_BASE_COLLECTIVES=[0,1] to allow users
        enable/disable using new nccl base collecectives. By default, using new nccl base collectives
        is enabled. [#801]
Min Xu's avatar
Min Xu committed
50
51
52

## [0.4.0] - 2021-07-31
### Fixed
53
54
55
56
57
58
59
- FSDP: fixed final backward callback in certain activation checkpointed cases. Before this fix,
        if a model is activation checkpointed in a certain way, the final backward
        callback can fire incorrectly. That's due to autograd and reentrant backward
        graphs. With this fix, the final callback is always registered on the outer
        most root FSDP instance (i.e. the outer most backward graph), which result
        in reliably firing. This makes FSDP much more robust with respect to different
        models and activation checkpoints. [#753]
Min Xu's avatar
Min Xu committed
60
61

### Added
62
63
64
65
- FSDP: support gradient accumulation without the `no_sync` context. This is useful
        in training with smaller number of GPU with same overall batch size as large
        number of GPUs. Compared with the `no_sync` context, this mode consumes less
        GPU memory but uses more networking bandwidth. [#752]
Min Xu's avatar
Min Xu committed
66

Min Xu's avatar
Min Xu committed
67

Min Xu's avatar
Min Xu committed
68
69
70
71
72
## [0.3.9] - 2021-07-26
### Fixed
- FSDP: fixed metadata saving and shard consolidation for MoE cases. When a model has
        shared parameters or mixture of expert layers, the handling of state dict
        metadata was broken. This release fixes that. [#746]
73
- OSS: fixed the buckets which would stay in fp16 if `broadcast fp16` was required [#751]
anj-s's avatar
anj-s committed
74
75

### Added
Min Xu's avatar
Min Xu committed
76
77
- FSDP: better performance; use `_allgather_base` and `_reduce_scatter_base` when they are
        available from pytorch nightly version (will be in 1.10 releases) [#729]
78
- FSDP: prepared FSDP internals for supporting multiple groups of flatten parameters (to support more general optimization) [#746]
anj-s's avatar
anj-s committed
79
80
81
82
83
84
85
86
87
88

## [0.3.8] - 2021-07-12
### Fixed
- checkpointing: Use dummy tensor to ensure backward pass is called. [#701]
- checkpointing: Ensure internal fwd counter is not incremented in eval mode. [#709]
- checkpointing: Use non-blocking CPU transfer to improve perf. [#719]
- FSDP: Fixed bug where buffers returned in `state_dict()` could still be half precision when `mixed_precision` is set to `True`. [#705]
- FSDP: Ensure requires_grad of FlatParameter is consistent with requires_grad of the original parameters. [#721]
- doc: Thoroughly improved the doc for FSDP. [#711]
- cleanup: Remove examples/ doc from the repo. [#712]
89
- cleanup: Future proof storage size test. [#735]
anj-s's avatar
anj-s committed
90
91
- cleanup: Migrate away from legacy torchtext iterators. [#713]
- chore: Updated torch 1.9 to release version. [#717]
Min Xu's avatar
Min Xu committed
92
93

### Added
94
- FSDP: supporting multiple flatten parameter groups [#708] [#711]
anj-s's avatar
anj-s committed
95
- chore: Add the latest numpy version to requirements-test.txt to prevent mypy errors on certain PR commits [#732]
Min Xu's avatar
Min Xu committed
96
97
98
99
100
101
102

## [0.3.7] - 2021-05-17
### Fixed
- setup.py: hide CUDA extensions behind `BUILD_CUDA_EXTENSIONS` envvar [#634]
- checkpointing: rename and move the `checkpoint_activations` wrapper [#654]
- FSDP: fix `local_state_dict` potentially called child class's `state_dict` [#574]
- FSDP: fix extra process groups being created by default. Old behavior can cause excessive GPU memory usage [#678] [#681]
103
104
105
- FSDP: fix forward pass not overlapping compute and allgather [#671]
- FSDP: improved frozen weight support [#657]
- FSDP: workaround AMP autocast cache issue with `clear_autocast_cache` flag [#650]
Min Xu's avatar
Min Xu committed
106
- FSDP: Rename API arg `cpu_offload` to `move_params_to_cpu` to better reflect functionality. We will deprecate `cpu_offload` in an upcoming release [#676]
107
108
- MoE: several fixes [#666] [#667] [#668]
- SDP: re-expose the module property [#647]
Min Xu's avatar
Min Xu committed
109
- wrap: support wrapping based on `wrapper_config` [#685]
110

111
### Added
112
113
- FSDP: added `force_input_to_fp32` flag for SyncBatchNorm [#659]
- FSDP: better memory usage for reduce bucket [#633]
114
115
- FSDP: added `local_metadata_dict` to save sharding relating information [#683]
- FSDP: added `consolidate_shard_weights` to reconstruct the consolidated (non-sharded) model weights from saved sharded weights and metadata on the disk [#683]
Min Xu's avatar
Min Xu committed
116
- Experimental SyncBatchNorm [#662] [#680]
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
117

Min Xu's avatar
Min Xu committed
118
119
## [0.3.6] - 2021-04-26
### Added
120
- FSDP: Consolidate cpu\_adam optimizer state dict ([#607](https://github.com/facebookresearch/fairscale/pull/607))
Min Xu's avatar
Min Xu committed
121
122
123
124
125
126
127
128
129

### Fixed
- FSDP: handle model with multiple forward pass and checkpoint ([#621](https://github.com/facebookresearch/fairscale/pull/621))
- FSDP & SDP: check before calling `_specify_ddp_gpu_num` ([#626](https://github.com/facebookresearch/fairscale/pull/626))
- FSDP: relax checking root condition ([#620](https://github.com/facebookresearch/fairscale/pull/620))
- SDP: removing an assert which does not seem always accurate ([#625](https://github.com/facebookresearch/fairscale/pull/625))
- FSDP: changing FSDP init to by pass pg validation ([#619](https://github.com/facebookresearch/fairscale/pull/619))
- OSS: to 100% coverage ([#618](https://github.com/facebookresearch/fairscale/pull/618))

Min Xu's avatar
Min Xu committed
130
131
132
133
134
135
136
137
138
139
## [0.3.5] - 2021-04-19
### Added
- [offload] Add API, tutorial and smaller doc string changes. ([#576](https://github.com/facebookresearch/fairscale/pull/576))

### Fixed
- FSDP: fixing training with freezing weights ([#614](https://github.com/facebookresearch/fairscale/pull/614))
- SDP: privatizing all the things ([#611](https://github.com/facebookresearch/fairscale/pull/611))
- FSDP: Make `_get_default_cuda_device` more robust to modules without params ([#606](https://github.com/facebookresearch/fairscale/pull/606))
- OffloadModel: Add prev codepath of using OffloadModel without activation checkpointing ([#608](https://github.com/facebookresearch/fairscale/pull/608))

Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
140
## [0.3.4] - 2021-04-13
Min Xu's avatar
Min Xu committed
141
### Added
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
142
143
- FSDP: Add no broadcast optim state option ([#560](https://github.com/facebookresearch/fairscale/pull/560))

Min Xu's avatar
Min Xu committed
144
### Fixed
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
145
146
147
148
149
- ShardedDDP: Properly handle .eval() mode ([#587](https://github.com/facebookresearch/fairscale/pull/587))
- ShardedDDP: Handle model being moved back to CPU prior to state consolidation ([#573](https://github.com/facebookresearch/fairscale/pull/573))
- FSDP: much faster state consolidation ([#595](https://github.com/facebookresearch/fairscale/pull/595))
- FSDP: Add gradient pre-dedivide to prevent overflow with large world sizes ([#565](https://github.com/facebookresearch/fairscale/pull/565))
- Offload: (experimental) Fix activation offloading to CPU ([#588]((https://github.com/facebookresearch/fairscale/pull/588) )
Min Xu's avatar
Min Xu committed
150

Min Xu's avatar
Min Xu committed
151
152
## [0.3.3] - 2021-04-1
### Added
153
- FSDP: changed `auto_wrap_bn` utility function so that single FSDP group is optional ([#556](https://github.com/facebookresearch/fairscale/pull/556))
Min Xu's avatar
Min Xu committed
154
155
156
157
158
159
160
161
- FSDP: optimizer state load/save ([#537](https://github.com/facebookresearch/fairscale/pull/537))
- FSDP: fix weight init when using apply() ([#543](https://github.com/facebookresearch/fairscale/pull/543))
- Multiprocess Pipe: retired old implementation
- Experimental: xpipe

### Fixed
- ShardedDDP deferred init ([#558](https://github.com/facebookresearch/fairscale/pull/558))

Min Xu's avatar
Min Xu committed
162
## [0.3.2] - 2021-03-18
Min Xu's avatar
Min Xu committed
163
### Added
164
- Experimental: Add spectrain support ([#372](https://github.com/facebookresearch/fairscale/issues/372))
Min Xu's avatar
Min Xu committed
165
- FSDP: enabled pytorch SyncBN (no asserting) ([#527](https://github.com/facebookresearch/fairscale/issues/527))
166
- FSDP: added `auto_wrap_bn` utility function ([#531](https://github.com/facebookresearch/fairscale/pull/531))
Min Xu's avatar
Min Xu committed
167
168

### Fixed
169
- OSS: fix a compatibily problem with lightning wrt optimizer state dict ([#510](https://github.com/facebookresearch/fairscale/issues/510))
170
- FSDP: fixed a bug when part of autograd graph is traversed multiple times in mixed precision mode ([#513](https://github.com/facebookresearch/fairscale/pull/513))
Min Xu's avatar
Min Xu committed
171
172
173
174

## [0.3.1] - 2021-03-09
### Added
- FSDP docs ([#455](https://github.com/facebookresearch/fairscale/issues/455))
175
- `enable_wrap` and `auto_wrap` APIs ([#446](https://github.com/facebookresearch/fairscale/issues/446))
Min Xu's avatar
Min Xu committed
176
177
- Added experimental.nn.OffloadModel API for training large models on a single GPU.([#432](https://github.com/facebookresearch/fairscale/issues/432))

178
### Fixed
Min Xu's avatar
Min Xu committed
179
180
181
182
- OSS: fix a broken state dict when using non contiguous param groups
- Several SDP fixes around performance and corner cases
- Many FSDP fixes
- AdaScale & SDP/FSDP test added but not officially supported
183

Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
184
## [0.3.0] - 2021-02-22
Min Xu's avatar
Min Xu committed
185
186
### Added
- FullyShardedDataParallel (FSDP) ([#413](https://github.com/facebookresearch/fairscale/issues/413))
187
- ShardedDDP fp16 grad reduction option ([#402](https://github.com/facebookresearch/fairscale/issues/402))
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
188
- Expose experimental algorithms within the pip package ([#410](https://github.com/facebookresearch/fairscale/pull/410))
Min Xu's avatar
Min Xu committed
189

190
### Fixed
Min Xu's avatar
Min Xu committed
191
- Catch corner case when the model is too small with respect to the world size, and shards are empty ([#406](https://github.com/facebookresearch/fairscale/pull/406))
192
- Memory leak in `checkpoint_wrapper` ([#412](https://github.com/facebookresearch/fairscale/pull/412))
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
193
194

## [0.1.7] - 2021-02-19
195
196
### Fixed
- ShardedDDP and OSS handle model trainability changes during training ([#369](https://github.com/facebookresearch/fairscale/issues/369))
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
197
198
199
- ShardedDDP state dict load/save bug ([#386](https://github.com/facebookresearch/fairscale/issues/386))
- ShardedDDP handle train/eval modes ([#393](https://github.com/facebookresearch/fairscale/issues/393))
- AdaScale handling custom scaling factors ([#401](https://github.com/facebookresearch/fairscale/issues/401))
200
201

### Added
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
202
- ShardedDDP manual reduce option for checkpointing ([#389](https://github.com/facebookresearch/fairscale/issues/389))
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
203

Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
204
205
206
207
208
209
210
211
212
213
## [0.1.6] - 2021-02-10
### Added
- Checkpointing model wrapper (#376)
- Faster OSS, flatbuffers (#371)
- Small speedup in OSS clipgradnorm (#363)

### Fixed
- Bug in ShardedDDP with 0.1.5 depending the init (KeyError / OSS)
- Much refactoring in Pipe (#357, #358, #360, #362, #370, #373)
- Better pip integration / resident pytorch (#375)
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
214
215

## [0.1.5] - 2021-02-03
216
### Added
217
218
219
- Pytorch compatibility for OSS checkpoints (#310)
- Elastic checkpoints for OSS, world size can vary in between save and loads (#310)
- Tensor views for OSS bucketing, reduced CPU use (#300)
220
- Bucket calls in ShardedDDP, for faster inter node communications (#327)
Benjamin Lefaudeux's avatar
Benjamin Lefaudeux committed
221
222
223
224
225
226
227
- FlattenParamWrapper, which flattens module parameters into a single tensor seamlessly (#317)
- AMPnet experimental support (#304)

### Fixed
- ShardedDDP properly handles device changes via `.to()` (#353)
- Add a new interface for AdaScale, AdaScaleWrapper, which makes it compatible with OSS (#347)

228

229
230
231
232
233
## [0.1.4] - 2021-01-07
### Fixed
- Missing cu files in the pip package


234
235
236
## [0.1.3] - 2021-01-04
### Fixed
- Release numbering within python and from pypi
237

238
## [0.1.2] - 2021-01-04
239
### Added
240
241
- AdaScale:
  . Added gradient accumulation feature (#202)
242
243
244
  . Added support of `torch.lr_scheduler` (#229)
  . Added support for `add_param_groups` (#266)
  . Added support for `scale != world_size` (#266)
245
246

### Fixed
Min Xu's avatar
Min Xu committed
247
248
- AdaScale: smoothing factor value fixed when using gradient accumulation (#235)
- Pipe: documentation on balancing functions (#243)
249
250
251
- ShardedDDP: handle typical NLP models
- ShardedDDP: better partitioning when finetuning

252

253
254
255
256
## [0.1.1] - 2020-12-01
### Fixed
- make sure pip package includes header files (#221)

msbaines's avatar
msbaines committed
257
258
259
260
261
262
## [0.1.0] - 2020-12-01
### Added
- ShardedDataParallel with autoreduce (#157)
- cpu support for Pipe (#188)
- ShardedOptim: Distributed Grad Scaler (for torch AMP)  (#182)
- OSS-aware clip grads, bridge sharded states (#167)
263
- oss: add `rank_local_state_dict` staticmethod (#174)
msbaines's avatar
msbaines committed
264
265
266
267
268
269
- support for PyTorch 1.7.0 (#171)
- Add implementation of AdaScale (#139)

### Fixed
- pip package install (#196, #200)

msbaines's avatar
msbaines committed
270
271
272
273
274
275
276
277
## [0.0.3] - 2020-10-14
### Added
- multi-process pipe

### Fixed
- multiple OSS fixes
- MegaTron+OSS DDP fix

msbaines's avatar
msbaines committed
278
279
## [0.0.2] - 2020-08-28
### Added
280
- add ddp that works with oss with `reduce()` not `all_reduce()` (#19)
msbaines's avatar
msbaines committed
281
282
283
284
285
286
287
288
289
- support for PyTorch v1.6
- add mixed precision Adam (#40)
- Adam optimizer state scaling (#44)

### Fixed
- properly restore a sharded optim state (#39)
- OSS restore state to proper device (#46)
- optim/oss: support optimizers with additional step kwargs (#53)
- optim/oss: fix state cast (#56)
290
- fix eval for `oss_ddp` (#55)
msbaines's avatar
msbaines committed
291
292
293
- optim/oss: work correctly with LRScheduler (#58)

## [0.0.1] - 2020-07-31
294
- Initial release.