Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/69508
Original Phabricator Diff: D32704467 (e032dae329)
Reland, fix is to not test traditional checkpoint when input does not require grad as that is unsupported as documented.
Original PR body:
Resubmission of https://github.com/pytorch/pytorch/pull/62964 with the
suggestions and tests discussed in
https://github.com/pytorch/pytorch/issues/65537.
Adds a `use_reentrant=False` flag to `checkpoint` function. When
`use_reentrant=True` is specified, a checkpointing implementation that uses
SavedVariableHooks instead of re-entrant autograd is used. This makes it more
composable with things such as `autograd.grad` as well as DDP (still need to
add thorough distributed testing).
As discussed in https://github.com/pytorch/pytorch/issues/65537, the tests that we need to add are:
- [x] Gradient hooks are called once
- [x] works when input does require grads but Tensor that require grads are captures (like first layer in a nn)
- [x] works for functions with arbitrary input/output objects
- [x] distributed tests (next PR)
Note that this is only for `torch.utils.checkpoint`, if this approach overall looks good, we will do something similar for `checkpoint_sequential`.
ghstack-source-id: 144948501
Test Plan: CI
Reviewed By: zhaojuanmao
Differential Revision: D32902634
fbshipit-source-id: 2ee87006e5045e5471ff80c36a07fbecc2bea3fe
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/69027
Resubmission of https://github.com/pytorch/pytorch/pull/62964 withe
suggestions and tests discussed in
https://github.com/pytorch/pytorch/issues/65537.
Adds a `use_reentrant=False` flag to `checkpoint` function. When
`use_reentrant=True` is specified, a checkpointing implementation that uses
SavedVariableHooks instead of re-entrant autograd is used. This makes it more
composable with things such as `autograd.grad` as well as DDP (still need to
add thorough distributed testing).
As discussed in https://github.com/pytorch/pytorch/issues/65537, we have added
the following tests:
-[ ] Gradient hooks are called once
ghstack-source-id: 144644859
Test Plan: CI
Reviewed By: pbelevich
Differential Revision: D32704467
fbshipit-source-id: 6eea1cce6b935ef5a0f90b769e395120900e4412
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52422
As mentioned in https://github.com/pytorch/pytorch/issues/52415,
`torch.utils.checkpoint` doesn't support checkpointing for functions which have
non-tensor inputs and outputs.
This PR resolves this issue by ensuring the autograd machinery ignores the
non-tensor inputs and outputs and processes the tensors accordingly.
ghstack-source-id: 124406867
Test Plan:
1) unit test
2) waitforbuildbot
Reviewed By: albanD
Differential Revision: D26507228
fbshipit-source-id: 0a5a1591570814176185362e83ad18dabd9c84b0
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45934https://pytorch.org/docs/stable/checkpoint.html pytorch checkpoint requires all input to the function being checkpointed to requires_grad, but this assumption is not necessarily try. consider the following two examples
```
output = MultiheadedMaskedAtten(input, mask)
output = LSTM(input, seq_length)
```
both length and mask are tensors that won't requires grad, currently if you try to checkpoint torch.autograd.backward will complain
```
File "/mnt/xarfuse/uid-124297/7d159c34-seed-nspid4026531836-ns-4026531840/torch/autograd/function.py
", line 87, in apply
return self._forward_cls.backward(self, *args)
File "/mnt/xarfuse/uid-124297/7d159c34-seed-nspid4026531836-ns-4026531840/torch/utils/checkpoint.py"
, line 99, in backward
torch.autograd.backward(outputs, args)
File "/mnt/xarfuse/uid-124297/7d159c34-seed-nspid4026531836-ns-4026531840/torch/autograd/__init__.py
", line 132, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError: element 1 of tensors does not require grad and does not have a grad_fn
```
this diff allows skipping the non-grad-requiring tensor when running autograd.backward.
added documentation for this feature as well.
Test Plan: added unit test to make sure partial tensor grads can be used in checkpoint().
Differential Revision: D24094764
fbshipit-source-id: 6557e8e74132d5a392526adc7b57b6998609ed12
Summary:
Fix typos in torch.utils/_benchmark/README.md
Add empty __init__.py to examples folder to make example invocations from README.md correct
Fixed uniform distribution logic generation when mixval and maxval are None
Fixes https://github.com/pytorch/pytorch/issues/42984
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42960
Reviewed By: seemethere
Differential Revision: D23095399
Pulled By: malfet
fbshipit-source-id: 0546ce7299b157d9a1f8634340024b10c4b7e7de
Summary:
See https://discuss.pytorch.org/t/training-with-gradient-checkpoints-torch-utils-checkpoint-appears-to-reduce-performance-of-model/78102/3?u=jwl for details.
Updated the docs to warn users about issues with checkpointing models that use `detach()` or `torch.no_grad()` to freeze their model layers/weights during training. When they do this, training with `checkpoint` will fail as it forces the outputs to require gradients when the model itself does not. Hence, during the backward pass it will output the error:
```
[4]<stderr>:RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
```
Maybe it is possible to fix this directly in the code, but I am not sure how in the current codebase.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37266
Differential Revision: D21262558
Pulled By: mrshenli
fbshipit-source-id: 529cf370534504baf8937ef17dac5d6916fbf5ae
Summary:
To support variadic inputs of `checkpoint_sequential` was deprecated at https://github.com/pytorch/pytorch/issues/21006. This case should be warned with `DeprecationWarning` for PyTorch 1.2, but it should be simply failed with `TypeError` since PyTorch 1.3. This patch removes the `DeprecationWarning` for PyTorch 1.2.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25985
Differential Revision: D18809875
Pulled By: albanD
fbshipit-source-id: e84dd8629c04979c4b2dc63e8ada94292e8cedd0
Summary:
I've reported inconsistency between `checkpoint_sequential` and `nn.Sequential` at https://github.com/pytorch/pytorch/issues/19260. Both should provide the same input signature but they don't. I think the consistency is important and I agree with apaszke that `nn.Sequential`'s semantics should be kept instead of `checkpoint_sequential`.
I hope `checkpoint_sequential` raises `TypeError` on variadic arguments since PyTorch 1.2.0. But for now, it's okay just to warn as `DeprecationWarning`. I've talked about this approach with soumith.
Please review this pull request. Any comment will be my pleasure.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21006
Differential Revision: D15530801
Pulled By: soumith
fbshipit-source-id: 0ceb2cc6a17dcc547d0d00ebaf9df8603be53183
Summary:
Currently, we cannot run a checkpointed function with None argument.
```python
out = torch.utils.checkpoint.checkpoint(run_fn, input_var, None)
```
```
File "/home/tunz/anaconda3/envs/torchdev/lib/python3.7/site-packages/torch/utils/checkpoint.py", line 14, in detach_variable
x = inp.detach()
AttributeError: 'NoneType' object has no attribute 'detach'
```
This PR makes checkpoint function to safely handle None argument.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17969
Differential Revision: D14475148
Pulled By: ezyang
fbshipit-source-id: 9afe9e9aac511a6df1e1620e9ac341536890d451
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14278
In this commit, we make checkpoint_sequential work for models with multiple tensor inputs. Previously, it only processed the first tensor and ignored the rest.
We introduce a new test in test/test_utils.py that replicates the issue referenced in this [GitHub issue](https://github.com/pytorch/pytorch/issues/11093), and we make sure that the test passes by changing the behavior of checkpoint_sequential to process all input tensors.
Reviewed By: ezyang
Differential Revision: D13144672
fbshipit-source-id: 24f58233a65a0f5b80b89c8d8cbced6f814004f7
Summary:
This issue was noticed, and fix proposed, by raulpuric.
Checkpointing is implemented by rerunning a forward-pass segment for each checkpointed segment during backward. This can result in the RNG state advancing more than it would without checkpointing, which can cause checkpoints that include dropout invocations to lose end-to-end bitwise accuracy as compared to non-checkpointed passes.
The present PR contains optional logic to juggle the RNG states such that checkpointed passes containing dropout achieve bitwise accuracy with non-checkpointed equivalents.** The user requests this behavior by supplying `preserve_rng_state=True` to `torch.utils.checkpoint` or `torch.utils.checkpoint_sequential`.
Currently, `preserve_rng_state=True` may incur a moderate performance hit because restoring MTGP states can be expensive. However, restoring Philox states is dirt cheap, so syed-ahmed's [RNG refactor](https://github.com/pytorch/pytorch/pull/13070#discussion_r235179882), once merged, will make this option more or less free.
I'm a little wary of the [def checkpoint(function, *args, preserve_rng_state=False):](https://github.com/pytorch/pytorch/pull/14253/files#diff-58da227fc9b1d56752b7dfad90428fe0R75) argument-passing method (specifically, putting a kwarg after a variable argument list). Python 3 seems happy with it.
Edit: It appears Python 2.7 is NOT happy with a [kwarg after *args](https://travis-ci.org/pytorch/pytorch/builds/457706518?utm_source=github_status&utm_medium=notification). `preserve_rng_state` also needs to be communicated in a way that doesn't break any existing usage. I'm open to suggestions (a global flag perhaps)?
**Batchnorm may still be an issue, but that's a battle for another day.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14253
Differential Revision: D13166665
Pulled By: soumith
fbshipit-source-id: 240cddab57ceaccba038b0276151342344eeecd7
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13141
This is an example diff to show what lint rules are being applied.
Reviewed By: mingzhe09088
Differential Revision: D10858478
fbshipit-source-id: cbeb013f10f755b0095478adf79366e7cf7836ff
* Autograd container for trading compute for memory
* add a unit test for checkpoint
* address comments
* address review comments
* adding some docs for the checkpoint api
* more comments
* more comments
* repro bug
* Fix a subtle bug/apply some review comments
* Update checkpoint.py
* Run everything in grad mode
* fix flake and chunk=1
* use imperative backward as per discussion
* remove Variable and also add models and test for models
* Add a simple thread local variable to check for autograd grad mode
* remove models and models test after debugging
* address review comments
* address more comments
* address more comments
