Summary:
Currently, whether `AccumulateGrad` [steals](67cb018462/torch/csrc/autograd/functions/accumulate_grad.h (L42)) or [clones](67cb018462/torch/csrc/autograd/functions/accumulate_grad.h (L80)) an incoming gradient, the gradient ends up rowmajor contiguous, regardless of its param's layout. If the param's layout is channels last, or otherwise not rowmajor contigous, later kernels that apply gradients to params are forced into an uncoalesced memory access pattern for either the param or the gradient. This may not sound like a big deal but for any binary op on large tensors it's a >3X increase in gmem traffic => 3X slowdown.
The present PR changes `AccumulateGrad` to prefer, where possible, stashing gradients that match their params' layouts (["Gradient Layout Contract"](https://github.com/pytorch/pytorch/pull/34904/files#diff-ef1a56d24f66b280dcdb401502d6a796R29-R38)).
Allowing `AccumulateGrad` to stash non-rowmajor-contiguous grads means DDP allreduces and DP reduces must allow non-rowmajor-contiguous grads. This PR extends DDP and DP to allow gradients with non-rowmajor-contiguous strides as long as their layout is nonoverlapping and dense.
For good measure, I include changes that allow all five nccl primitives (allreduce, reduce, broadcast, allgather, reducescatter) to act on non-rowmajor-contiguous tensors (again as long as each input's layout is nonoverlapping and dense, and as long as all tensors participating in a given collective have the same layout). The primitive comm changes aren't necessary to enable the DDP changes, but I wasn't sure this would end up true until I had written both sets of changes. I think primitive comm enablement is reasonable to keep in the PR, especially since the code for it is simple.
Channels last params will be a major beneficiary of this PR, but I don't see it as channels-last-specific fix. The spirit is layout matching in general:
- Grads should be stashed with memory layouts matching their params.
- Src and dst tensors on opposite ends of collectives should have matching dense layouts.
This PR also updates autograd docs to describe potential BC-breaking changes below.
## BC notes
ngimel albanD gchanan
#### BC-breaking
In the common case where the user lets AccumulateGrad decide grad layouts, strides for grads of dense but non-rowmajor-contiguous params will change. Any user code that was accustomed to `view(-1)`ing these grads will break.
Also, the circumstances under which a grad can be stolen directly from the backward function that created it, as opposed to deep-copied by AccumulateGrad, have changed. In most cases we expect silent performance improvement, because we expect channels-last-aware backward kernels will create channels last gradients for channels last params. Now those can be stolen, whereas before this PR they were cloned and made rowmajor contiguous. IMO this is a mild BC breakage. Param backward hooks still see grads come in with whatever format the backward kernel gave them. The only BC breakage potential I see is if user code relies somehow on a grad in a hook having or not having the same deep memory as the eventual `param.grad`. Any such users hopefully know they're off the edge of the map and understand how to update their expectations.
#### BC escape hatches
At alband's recommendation, this PR's changes to AccumulateGrad do not alter the pre-PR code's decisions about whether grad is accumulated in or out of place. Accumulations of new grads onto an existing `.grad` attribute were (usually) in-place before this PR and remain in-place after this PR, keeping the existing `.grad`'s layout. After this PR, if the user wants to force accumulation into a grad with a particular layout, they can preset `param.grad` to a zeroed tensor with the desired strides or call `grad.contiguous(desired format)`. This likely won't be as performant as letting AccumulateGrad establish grad layouts by cloning or stealing grads with contract-compliant strides, but at least users have a control point.
One limitation (present before this PR and unchanged by this PR): Presetting `param.grad` does not ensure in-place accumulation all the time. For example, if `create_graph=True`, or if incoming `new_grad` is dense and existing `variable_grad` is sparse, accumulation occurs out of place, and the out-of-place result may not match the existing grad's strides.
----------------------------
I also noticed some potential DDP improvements that I considered out of scope but want to mention for visibility:
1. make sure Reducer's ops sync with AccumulateGrad streams
2. ~to reduce CPU overhead and incur fewer kernel launches, lazily create flat `contents` tensors by a single `cat` kernel only when a bucket is full, instead of `copy_`ing grads into `contents` individually as soon as they are received.~ PR includes a [minor change](https://github.com/pytorch/pytorch/pull/34904/files#diff-c269190a925a4b0df49eda8a8f6c5bd3R312-R315) to divide grads while copying them into flat buffers, instead of copying them in, then dividing separately. Without cat+div fusion, div-while-copying is the best we can do.
3. https://github.com/pytorch/pytorch/issues/38942
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34904
Differential Revision: D20496044
Pulled By: albanD
fbshipit-source-id: 248d680f4b1bf77b0a986451844ec6e254469217
Summary:
This updates assertEqual and assertEqual-like functions to either require both or neither of atol and rtol be specified. This should improve clarity around handling precision in the test suite, and it allows us to remove the legacy positional atol argument from assertEqual. In addition, the "message" kwarg is replace with a kwarg-only "msg" argument whose name is consistent with unittest's assertEqual argument.
In the future we could make "msg" an optional third positional argument to be more consistent with unittest's assertEqual, but requiring it be specified should be clear, and we can easily update the signature to make "msg" an optional positional argument in the future, too.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38872
Differential Revision: D21740237
Pulled By: mruberry
fbshipit-source-id: acbc027aa1d7877a49664d94db9a5fff91a07042
Summary:
This updates assertEqual and assertEqual-like functions to either require both or neither of atol and rtol be specified. This should improve clarity around handling precision in the test suite, and it allows us to remove the legacy positional atol argument from assertEqual. In addition, the "message" kwarg is replace with a kwarg-only "msg" argument whose name is consistent with unittest's assertEqual argument.
In the future we could make "msg" an optional third positional argument to be more consistent with unittest's assertEqual, but requiring it be specified should be clear, and we can easily update the signature to make "msg" an optional positional argument in the future, too.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38872
Differential Revision: D21717199
Pulled By: mruberry
fbshipit-source-id: 9feb856f94eee911b44f6c7140a1d07c1b026d3a
Summary:
Fixes https://github.com/pytorch/pytorch/issues/37182
The `zero_grad` wrapper from `_replicate_for_data_parallel` can't be pickled. So instead, I set an attribute `_is_replica = True` and check for this in `Module.zero_grad`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37307
Differential Revision: D21246119
Pulled By: mrshenli
fbshipit-source-id: 4755786d48a20bc247570ba672de9dd526914ce1
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35615
Python 2 has reached end-of-life and is no longer supported by PyTorch.
Now we can clean up a lot of cruft that we put in place to support it.
These changes were all done manually, and I skipped anything that seemed
like it would take more than a few seconds, so I think it makes sense to
review it manually as well (though using side-by-side view and ignoring
whitespace change might be helpful).
Test Plan: CI
Differential Revision: D20842886
Pulled By: dreiss
fbshipit-source-id: 8cad4e87c45895e7ce3938a88e61157a79504aed
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36411
This PR remove pytorch specific defined assertwarns and use the unit
test one, also format some tests
Test Plan: Imported from OSS
Differential Revision: D20998159
Pulled By: wanchaol
fbshipit-source-id: 1280ecff2dd293b95a639d13cc7417fc819c2201
Summary:
Initial integration of eager autocasting, supporting out-of-place ops only for easier review.
Relevant issue/RFC: https://github.com/pytorch/pytorch/issues/25081
In-place ops and ops with user-supplied `out=...` can certainly be supported as well (my initial WIP https://github.com/pytorch/pytorch/pull/29552 handled many) but require substantially more complex special casing in the autocasting backend and tests. Support for these ops (much of which has already been written) will be broken into later PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32140
Differential Revision: D20346700
Pulled By: ezyang
fbshipit-source-id: 12d77b3917310186fbddf11c59b2794dc859131f
Summary:
In DataParallel, replica parameters are not leaves (because they are computed via broadcast from master parameters), and should be treated as such. Fixes https://github.com/pytorch/pytorch/issues/33552
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33907
Differential Revision: D20150199
Pulled By: ngimel
fbshipit-source-id: 5965d4115b6b3a8433063126ff6269567872fbeb
Summary:
`.data` calls are unsafe and should not be used.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33874
Differential Revision: D20141059
Pulled By: izdeby
fbshipit-source-id: 8e11afc74f0cb04f5b18b458068fb813a6d51708
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32962
As per gchanan's comments on
https://github.com/pytorch/pytorch/pull/30445, I've used
`torch.set_default_dtype` in test_data_parallel instead of specifying
dtype=torch.double everywhere. Also, renamed dtype2prec to dtype2prec_DONTUSE
ghstack-source-id: 98388429
Test Plan: waitforbuildbot
Differential Revision: D19714374
fbshipit-source-id: eb55bbca33881625636ba9ea6dd4cb692f25668e
Summary:
Fixes https://github.com/pytorch/pytorch/issues/31768, second attempt of https://github.com/pytorch/pytorch/issues/32870
DataParallel creates replicas of the original `nn.Module` with the parameters duplicated onto the destination devices. Calling `backwards` will propagate gradients onto the original module parameters but calling `zero_grad` on the replica module doesn't clear the gradients from the parent module. However, any replica using backwards was broken anyway since the replica's parameters are not leaf nodes in autograd. So, we should issue a warning.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33064
Differential Revision: D19790178
Pulled By: albanD
fbshipit-source-id: 886f36640acef4834a6fa57a26ce16b42ff0e9ad
Summary:
Fixes https://github.com/pytorch/pytorch/issues/31768
`DataParallel` creates replicas of the original `nn.Module` with the parameters duplicated onto the destination devices. Calling `backwards` will propagate gradients onto the original module parameters but calling `zero_grad` on the replica module doesn't clear the gradients from the parent module,
~breaking any model that uses `backward`-`zero_grad` in its `forward`. I fix this by patching the replica module so that `zero_grad` clears grads on the parent as well.~
However, any replica using backwards was broken anyway since the replica's parameters are not leaf nodes in autograd. So, we should raise a warning.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32870
Differential Revision: D19730209
Pulled By: ezyang
fbshipit-source-id: cb9b2cb0c2e0aca688ce0ff3e56b40fbd2aa3c66
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30445
Create distributed and rpc directories under caffe/test for better management
of unit tests.
Differential Revision: D18702786
fbshipit-source-id: e9daeed0cfb846ef68806f6decfcb57c0e0e3606
