[Feature] Support torch ZeroRedundancyOptimizer #551

nijkah · 2022-09-27T03:02:12Z

Device: V100 32GB (32510MiB) x4
Input Size: 512
Total Batch Size: 16

upernet_r50_4xb4-80k_ade20k-512x512

Method	Time	Memory
w/o zero	0.611	7053
w/ zero	0.552	6868

Time may be dependent to the environment.

Co-authored-by: Junhwa Song ethan9867@gmail.com @KKIEEK
Signed-off-by: Junhwa Song ethan9867@gmail.com
Signed-off-by: Hakjin Lee nijkah@gmail.com

Check reduced memory
Reproduce Performance
Check saving optimizer's state
Check loading optimizer's state

Co-authored-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Hakjin Lee <nijkah@gmail.com>

codecov · 2022-09-27T04:27:09Z

Codecov Report

Base: 78.07% // Head: 78.04% // Decreases project coverage by -0.03% ⚠️

Coverage data is based on head (7e91b88) compared to base (36af1f0).
Patch coverage: 61.11% of modified lines in pull request are covered.

❗ Current head 7e91b88 differs from pull request most recent head 00ed3b3. Consider uploading reports for the commit 00ed3b3 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #551      +/-   ##
==========================================
- Coverage   78.07%   78.04%   -0.04%     
==========================================
  Files         125      126       +1     
  Lines        8991     9009      +18     
  Branches     1845     1846       +1     
==========================================
+ Hits         7020     7031      +11     
- Misses       1659     1666       +7     
  Partials      312      312

Flag	Coverage Δ
unittests	`78.04% <61.11%> (-0.04%)`	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmengine/optim/optimizer/zero_optimizer.py	`58.82% <58.82%> (ø)`
mmengine/optim/optimizer/__init__.py	`100.00% <100.00%> (ø)`

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

nijkah · 2022-09-27T04:33:59Z

  File "/workspace/mmengine/mmengine/runner/runner.py", line 1718, in call_hook                                                                                                                                          
    getattr(hook, fn_name)(self, **kwargs)                                                                                                                                                                                        
  File "/workspace/mmengine/mmengine/hooks/checkpoint_hook.py", line 488, in after_train_iter
    self._save_checkpoint(runner)                                                                                                                                                                                                 
  File "/workspace/mmengine/mmengine/dist/utils.py", line 346, in wrapper                                                                                                                                                
    return func(*args, **kwargs)                                                                                                                                                                                                  
  File "/workspace/mmengine/mmengine/hooks/checkpoint_hook.py", line 294, in _save_checkpoint                                                                                                                            
    **self.args)                                                                                                                                                                                                                  
  File "/workspace/mmengine/mmengine/runner/runner.py", line 2109, in save_checkpoint                                                                                                                                    
    ) else self.optim_wrapper.state_dict()                                                                                                                                                                                        
  File "/workspace/mmengine/mmengine/optim/optimizer/optimizer_wrapper.py", line 199, in state_dict                                                                                                                      
    return self.optimizer.state_dict()                                                                                                                                                                                            
  File "/opt/conda/lib/python3.7/site-packages/torch/distributed/optim/zero_redundancy_optimizer.py", line 1152, in state_dict                                                                                                    
    "Optimizer state has not been consolidated on this rank. "                                                                                                                                                                    
RuntimeError: Optimizer state has not been consolidated on this rank. Please call `consolidate_state_dict(to=0)` on all ranks beforehand if you meant to save the global state.

I found that it has a problem when saving the optimizer's state_dict.

nijkah · 2022-09-27T04:57:43Z

Solved the problem to call self.state_dict.

HAOCHENYE · 2022-09-27T07:24:42Z

Thank you for your contributions! It could be better to add a unit test to use the ZeroRedundancyOptimizer 😊.

This reverts commit dd64538.

HAOCHENYE · 2022-09-29T06:26:46Z

tests/test_optim/test_optimizer/test_optimizer.py

+        if ZeroRedundancyOptimizer is None:
+            self.skipTest('ZeroRedundancyOptimizer is not available.')


Suggested change

if ZeroRedundancyOptimizer is None:

self.skipTest('ZeroRedundancyOptimizer is not available.')

if ZeroRedundancyOptimizer is None:

self.skipTest('ZeroRedundancyOptimizer is not available.')

Is this line duplicated with

@unittest.skipIf( digit_version(TORCH_VERSION) < digit_version('1.8.0'), reason='ZeRO needs Pytorch 1.8 or higher')

https://github.com/open-mmlab/mmengine/actions/runs/3134972777/jobs/5090129146#step:8:132
I found that importing ZeroRedundancyOptimizer failed in the Windows CPU CI with & torch1.8.1.
(The importing failure made _ZeroRedundancyOptimizer as object.)
So I added duplicated skip code.

I'll check it again.

I found that it has another condition.
torch.distributed.rpc should be available. I removed duplicated lines, and clarified this condition.

tests/test_optim/test_optimizer/test_optimizer.py

nijkah · 2022-09-30T08:34:59Z

I found another bug.

Currently, saving the ZeroRedundancyOptimizer's state_dict gives the error.
Each ZeroRedundancyOptimizer should call consolidate_state_dict for all processes, and only the main process's optimizer should return the state_dict.

This constraint makes refactor the CheckpointHook.
I'll have to think about it a little more.

nijkah · 2022-09-30T08:53:50Z

Hi, @HAOCHENYE @C1rN09

I changed the CheckpointHook's logic to remove master_only decorators when saving the checkpoint.
I think it doesn't affect the logic since save_checkpoint function in mmengine/runner/checkpoint.py is wrapped by master_only already.

Please notify me if the potential problem is expected.

HAOCHENYE · 2022-10-01T05:59:31Z

Hi, @HAOCHENYE @C1rN09

I changed the CheckpointHook's logic to remove master_only decorators when saving the checkpoint. I think it doesn't affect the logic since save_checkpoint function in mmengine/runner/checkpoint.py is wrapped by master_only already.

Please notify me if the potential problem is expected.

#553 also remove the master_only decorator to collect all weights from different rank. Does @C1rN09 have any idea about this?

mmengine/optim/optimizer/zero_optimizer.py

zhouzaida · 2022-10-21T12:32:20Z

mmengine/optim/optimizer/zero_optimizer.py

+            '`torch.distributed.optim.ZeroReundancyOptimizer` is only '
+            'available when pytorch version >= 1.8.')
+        assert is_available(), 'torch.distributed.rpc is not available.'
+        optimizer_class = getattr(torch.optim, optimizer_type)


Can it support custom Optimizer classes?

I'm still figuring it out now. Until now, it does not seem to have a specific dependency on torch's optimizers. It may be possible to custom Optimizer classes.

Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com>

zhouzaida · 2022-10-24T08:24:50Z

Hi @nijkah , the lint failed.

nijkah · 2022-10-24T09:21:01Z

Hi @C1rN09 @zhouzaida , I'm concerning about the option overlap_with_ddp.

The public API docs demonstrate:

overlap_with_ddp (bool, optional) – if True, step() is overlapped with DistributedDataParallel ‘s gradient synchronization; this requires (1) either a functional optimizer for the optimizer_class argument or one with a functional equivalent and (2) registering a DDP communication hook constructed from one of the functions in ddp_zero_hook.py; parameters are packed into buckets matching those in DistributedDataParallel, meaning that the parameters_as_bucket_view argument is ignored. If False, step() runs disjointly after the backward pass (per normal). (default: False)

So if someone wants to use overlap_with_ddp option, there should be some modifications like registering a DDP communication hook, removing consolidate_state_dict, and so on.

Will it be better to fix it as overlap_with_ddp=False?

nijkah · 2022-10-24T10:22:38Z

mmengine/optim/optimizer/zero_optimizer.py

+    def state_dict(self):
+        """Consolidate `state_dict`s from ranks to save the `state_dict`."""
+        self.consolidate_state_dict()
+        state_dict = super().state_dict() if is_main_process() else dict()


mmengine/mmengine/optim/optimizer/amp_optimizer_wrapper.py

Line 107 in bda92b4

state_dict['loss_scaler'] = self.loss_scaler.state_dict()

Due to this line, using ZeroRedundancyOptimizer with AmpOptimWrapper gave the error like

TypeError: 'NoneType' object does not support item assignment in <mmengine.hooks.checkpoint_hook.CheckpointHook object at XXXXXX>

So I modified it to return dict() instead of None when it is not the main process.

C1rN09 · 2022-10-25T03:55:47Z

Hi @C1rN09 @zhouzaida , I'm concerning about the option overlap_with_ddp.

The public API docs demonstrate:

overlap_with_ddp (bool, optional) – if True, step() is overlapped with DistributedDataParallel ‘s gradient synchronization; this requires (1) either a functional optimizer for the optimizer_class argument or one with a functional equivalent and (2) registering a DDP communication hook constructed from one of the functions in ddp_zero_hook.py; parameters are packed into buckets matching those in DistributedDataParallel, meaning that the parameters_as_bucket_view argument is ignored. If False, step() runs disjointly after the backward pass (per normal). (default: False)

So if someone wants to use overlap_with_ddp option, there should be some modifications like registering a DDP communication hook, removing consolidate_state_dict, and so on.

Will it be better to fix it as overlap_with_ddp=False?

Hi! If there is no easy solution, I think it's acceptable to fix overlap_with_ddp=False in this PR. You may add a few lines of comments/TODO to explain why. Probably we could introduce this feature in the future.

nijkah · 2022-10-25T08:16:19Z

@C1rN09 In some versions (e.g. 1.8.0), ZeroRedundancyOptimizer does not have a keyword argument for overlap_with_ddp.
So I just added a TODO comment.

C1rN09 · 2022-10-27T08:44:31Z

Hi, @nijkah I tested your branch on my cluster and got some different results, show as below:

Method	Time	Memory
w/o zero	~0.24	7053
w/ zero	~0.24	6868

What I'm confused about is the memory consumption. Since this model is only ~250MB, it seems that the maximum memory reduction that ZeroOptimizer can do with momentum SGD is 250MB * (1 + 1) * 0.75 ~ 375MB, which is close to my results. Could you provide your detailed modifications on models, configs, etc (if any), to help me find out what I've missed?

My cluster & configurations in case I miss something important:

Device: 4 x A100
Configs: same as upernet_r50_4xb4-80k_ade20k-512x512.py, except that I changed the dataset inheritence to cityscapes because my cluster has no ade20k dataset
Zero optimizer configs:

optimizer = dict(type='ZeroRedundancyOptimizer', optimizer_type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
optim_wrapper = dict(type='OptimWrapper', optimizer=optimizer, clip_grad=None)

nijkah · 2022-10-27T09:10:29Z

Hi, @C1rN09. I apologize for confusing this. 😞
You are totally correct. I found that I was confused about this today.

I was confused because I just tried to compare my result with the provided log in github.

After running the experiments again, I also could get the same result as you. I'll fix the description.

nijkah · 2022-10-28T04:50:24Z

nit: Hi @C1rN09. Since the SGD optimizer with momentum only stores the model params from the previous step,
should the expected memory reduction be 250MB * (1) * 0.75 ~ 175MB, which is more close to the result?

C1rN09 · 2022-10-28T07:11:44Z

nit: Hi @C1rN09. Since the SGD optimizer with momentum only stores the model params from the previous step, should the expected memory reduction be 250MB * (1) * 0.75 ~ 175MB, which is more close to the result?

Yes, I think you are right! From the experiment result I guess it only shards optimizer states, instead of optimizer states + grads, which I used to think it might do.

* [Feature] Support torch ZeRORedundancyOptimizer Co-authored-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Hakjin Lee <nijkah@gmail.com> * lint * Fix saving optimizer state_dict * Fix handling import error * Add test case * fix UT * Revert "fix UT" This reverts commit dd64538. * fix handling import in UT * Fix saving zero checkpoint and delete redundant master_only * lint * test unittest * Fix handling impor error * Fix UT condition * Edit docstrings * Fix typo * Skip redundant procudure in checkpoint hook * fix typo again * Update mmengine/optim/optimizer/zero_optimizer.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * Add api info * lint * Fix lint * Handling AmpOptimWrapper case * handling overlap_with_ddp * Fix error Signed-off-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Hakjin Lee <nijkah@gmail.com> Co-authored-by: Junhwa Song <ethan9867@gmail.com> Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com>

twmht · 2023-03-28T02:51:11Z

@nijkah

how can i use ZeroRedundancyOptimizer with AmpOptimizer in the training config?

nijkah · 2023-03-28T02:56:38Z

@nijkah

how can i use ZeroRedundancyOptimizer with AmpOptimizer in the training config?

I didn't test it yet. Is there any specific reason to use AmpOptimizer for 'mixed precision training'?
Unless it is necessary, I recommend you use AmpOptimWrapper. You can find a related configuration file and just change its optimizer part.

twmht · 2023-03-28T07:09:10Z

@nijkah

This is what i want

optimizer = dict(type='ZeroRedundancyOptimizer', optimizer_type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0005)
optim_wrapper = dict(type='AmpOptimWrapper', optimizer=optimizer, clip_grad=None)

Did you test that?

nijkah · 2023-03-28T07:31:04Z

@twmht Yes, it should work! 😄
I tested it and this is a related line I edited. #551 (comment)

[Feature] Support torch ZeRORedundancyOptimizer

d466fb8

Co-authored-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Junhwa Song <ethan9867@gmail.com> Signed-off-by: Hakjin Lee <nijkah@gmail.com>

zhouzaida requested a review from HAOCHENYE September 27, 2022 03:35

lint

d957bdc

nijkah marked this pull request as draft September 27, 2022 04:29

Fix saving optimizer state_dict

67bcec8

nijkah marked this pull request as ready for review September 27, 2022 04:57

Fix handling import error

da39cd2

Add test case

a408f69

nijkah changed the title ~~[Feature] Support torch ZeRORedundancyOptimizer~~ [Feature] Support torch ZeroRedundancyOptimizer Sep 27, 2022

nijkah added 3 commits September 27, 2022 14:24

fix UT

dd64538

Revert "fix UT"

48806af

This reverts commit dd64538.

fix handling import in UT

00ed3b3

HAOCHENYE reviewed Sep 29, 2022

View reviewed changes

tests/test_optim/test_optimizer/test_optimizer.py Outdated Show resolved Hide resolved

HAOCHENYE reviewed Sep 29, 2022

View reviewed changes

tests/test_optim/test_optimizer/test_optimizer.py Outdated Show resolved Hide resolved

zhouzaida requested a review from C1rN09 September 30, 2022 06:22

Fix saving zero checkpoint and delete redundant master_only

dd5986c

nijkah added 4 commits September 30, 2022 09:03

lint

764133d

test unittest

f9a6dad

Fix handling impor error

1c1da2e

Fix UT condition

0275d08

nijkah commented Oct 4, 2022

View reviewed changes

mmengine/optim/optimizer/zero_optimizer.py Outdated Show resolved Hide resolved

mmengine/optim/optimizer/zero_optimizer.py Outdated Show resolved Hide resolved

mmengine/optim/optimizer/zero_optimizer.py Outdated Show resolved Hide resolved

Edit docstrings

831e5c7

zhouzaida reviewed Oct 21, 2022

View reviewed changes

mmengine/optim/optimizer/zero_optimizer.py Outdated Show resolved Hide resolved

zhouzaida reviewed Oct 21, 2022

View reviewed changes

Update mmengine/optim/optimizer/zero_optimizer.py

9d63ea6

Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com>

nijkah dismissed C1rN09’s stale review via 9d63ea6 October 24, 2022 04:17

nijkah added 2 commits October 24, 2022 18:13

Add api info

e244e30

lint

41dff5a

nijkah added 2 commits October 24, 2022 18:22

Fix lint

929cf06

Handling AmpOptimWrapper case

d6d0f45

nijkah commented Oct 24, 2022

View reviewed changes

nijkah added 2 commits October 25, 2022 17:02

handling overlap_with_ddp

f0ac4cb

Fix error

adee57c

HAOCHENYE added the enhancement New feature or request label Oct 27, 2022

C1rN09 approved these changes Oct 27, 2022

View reviewed changes

zhouzaida approved these changes Oct 27, 2022

View reviewed changes

zhouzaida merged commit 0857f9f into open-mmlab:main Oct 27, 2022

nijkah mentioned this pull request Oct 31, 2022

[GPU Memory Issue] why does the memory go peak on the first training log? #651

Closed

nijkah deleted the zero_1_optimizer branch November 3, 2022 04:04

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feature] Support torch ZeroRedundancyOptimizer #551

[Feature] Support torch ZeroRedundancyOptimizer #551

nijkah commented Sep 27, 2022 •

edited

Loading

codecov bot commented Sep 27, 2022 •

edited

Loading

nijkah commented Sep 27, 2022

nijkah commented Sep 27, 2022

HAOCHENYE commented Sep 27, 2022

HAOCHENYE Sep 29, 2022

nijkah Sep 29, 2022 •

edited

Loading

nijkah Sep 30, 2022

nijkah commented Sep 30, 2022 •

edited

Loading

nijkah commented Sep 30, 2022 •

edited

Loading

HAOCHENYE commented Oct 1, 2022

zhouzaida Oct 21, 2022

nijkah Oct 24, 2022 •

edited

Loading

zhouzaida commented Oct 24, 2022

nijkah commented Oct 24, 2022 •

edited

Loading

nijkah Oct 24, 2022 •

edited

Loading

C1rN09 commented Oct 25, 2022

nijkah commented Oct 25, 2022

C1rN09 commented Oct 27, 2022 •

edited

Loading

nijkah commented Oct 27, 2022 •

edited

Loading

nijkah commented Oct 28, 2022 •

edited

Loading

C1rN09 commented Oct 28, 2022

twmht commented Mar 28, 2023

nijkah commented Mar 28, 2023 •

edited

Loading

twmht commented Mar 28, 2023

nijkah commented Mar 28, 2023 •

edited

Loading

		if ZeroRedundancyOptimizer is None:
		self.skipTest('ZeroRedundancyOptimizer is not available.')

[Feature] Support torch ZeroRedundancyOptimizer #551

[Feature] Support torch ZeroRedundancyOptimizer #551

Conversation

nijkah commented Sep 27, 2022 • edited Loading

upernet_r50_4xb4-80k_ade20k-512x512

codecov bot commented Sep 27, 2022 • edited Loading

Codecov Report

nijkah commented Sep 27, 2022

nijkah commented Sep 27, 2022

HAOCHENYE commented Sep 27, 2022

HAOCHENYE Sep 29, 2022

Choose a reason for hiding this comment

nijkah Sep 29, 2022 • edited Loading

Choose a reason for hiding this comment

nijkah Sep 30, 2022

Choose a reason for hiding this comment

nijkah commented Sep 30, 2022 • edited Loading

nijkah commented Sep 30, 2022 • edited Loading

HAOCHENYE commented Oct 1, 2022

zhouzaida Oct 21, 2022

Choose a reason for hiding this comment

nijkah Oct 24, 2022 • edited Loading

Choose a reason for hiding this comment

zhouzaida commented Oct 24, 2022

nijkah commented Oct 24, 2022 • edited Loading

nijkah Oct 24, 2022 • edited Loading

Choose a reason for hiding this comment

C1rN09 commented Oct 25, 2022

nijkah commented Oct 25, 2022

C1rN09 commented Oct 27, 2022 • edited Loading

nijkah commented Oct 27, 2022 • edited Loading

nijkah commented Oct 28, 2022 • edited Loading

C1rN09 commented Oct 28, 2022

twmht commented Mar 28, 2023

nijkah commented Mar 28, 2023 • edited Loading

twmht commented Mar 28, 2023

nijkah commented Mar 28, 2023 • edited Loading

nijkah commented Sep 27, 2022 •

edited

Loading

codecov bot commented Sep 27, 2022 •

edited

Loading

nijkah Sep 29, 2022 •

edited

Loading

nijkah commented Sep 30, 2022 •

edited

Loading

nijkah commented Sep 30, 2022 •

edited

Loading

nijkah Oct 24, 2022 •

edited

Loading

nijkah commented Oct 24, 2022 •

edited

Loading

nijkah Oct 24, 2022 •

edited

Loading

C1rN09 commented Oct 27, 2022 •

edited

Loading

nijkah commented Oct 27, 2022 •

edited

Loading

nijkah commented Oct 28, 2022 •

edited

Loading

nijkah commented Mar 28, 2023 •

edited

Loading

nijkah commented Mar 28, 2023 •

edited

Loading