# Motivation
## for `torch.amp.GradScaler`,
- `torch.cpu.amp.GradScaler(args...)` is completely equivalent to `torch. amp.GradScaler("cpu", args...)`.
- `torch.cuda.amp.GradScaler(args...)` is completely equivalent to `torch.amp.GradScaler("cuda", args...)`.
So, we intend to depreate them and **strongly recommend** developer to use `torch.amp.GradScaler`.
## for `custom_fwd` and `custom_bwd`,
this is a good solution to make the custom function run with or without effect even in an autocast-enabled region and can be shared by other backends, like CPU and XPU.
So we generalize it to be device-agnostic and put them int `torch/amp/autocast_mode.py` and re-expose to `torch.amp.custom_fwd` and `torch.amp.custom_bwd`. Meanwhile, we deprecate `torch.cuda.amp.custom_fwd` and `torch.cuda.amp.custom_bwd`.
# Additional Context
Add UT to cover the deprecated warning.
No need for more UTs to cover the functionality of `torch.amp.custom_f/bwd`, the existing UTs that previously covered the functionality of `torch.cuda.amp.custom_f/bwd` can cover them.
To facilitate the review, we separate these code changes to two PRs. The first PR cover `torch.amp.GradScaler`. The follow-up covers `custom_fwd` and `custom_bwd`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126527
Approved by: https://github.com/jgong5, https://github.com/gujinghui, https://github.com/janeyx99, https://github.com/EikanWang
Add non-package python modules to the public API checks.
The original change is to remove the `ispkg` check in this line
https://github.com/pytorch/pytorch/blob/main/docs/source/conf.py#L518
Everything else is to add the appropriate modules to the rst files, make sure every module we provide can be imported (fixed by either making optional dependencies optional or just deleting files that have been un-importable for 3 years), make API that are both modules and functions (like torch.autograd.gradcheck) properly rendered on the docs website without confusion and add every non-documented API to the allow list (~3k of them).
Next steps will be to try and fix these missing docs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110568
Approved by: https://github.com/zou3519
Summary:
Working towards https://docs.google.com/document/d/10yx2-4gs0gTMOimVS403MnoAWkqitS8TUHX73PN8EjE/edit?pli=1#
This PR:
- Ensure that all the submodules are listed in a rst file (that ensure they are considered by the coverage tool)
- Remove some long deprecated code that just error out on import
- Remove the allow list altogether to ensure nothing gets added back there
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73983
Reviewed By: anjali411
Differential Revision: D34787908
Pulled By: albanD
fbshipit-source-id: 163ce61e133b12b2f2e1cbe374f979e3d6858db7
(cherry picked from commit c9edfead7a01dc45bfc24eaf7220d2a84ab1f62e)
Summary:
Should close https://github.com/pytorch/pytorch/issues/42218
Numerically, `grid_sampler` is fine in fp16 or fp32, but takes several inputs and expects their dtypes to match, so it belongs on the autocast promote list.
`grid_sampler` currently uses `gpuAtomicAdd`, notoriously slow in fp16 because it calls cuda's atomicAdd __half overload which uses a software compare-and-swap loop internally. To allow good performance if both inputs happen to be FP16, the PR also modifies `grid_sampler_[2,3]d_backward_kernel`s to use `fastAtomicAdd` instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/58618
Reviewed By: mruberry
Differential Revision: D29257199
Pulled By: ngimel
fbshipit-source-id: 3cc7505945b480427f2fc1beb36bee80bf3853b3
Summary:
Temporary fix for https://github.com/pytorch/pytorch/issues/42218.
Numerically, grid_sampler should be fine in fp32 or fp16. So grid_sampler really belongs on the promote list. But performancewise, native grid_sampler backward kernels use gpuAtomicAdd, which is notoriously slow in fp16. So the simplest functionality fix is to put grid_sampler on the fp32 list.
In https://github.com/pytorch/pytorch/pull/58618 I implement the right long-term fix (refactoring kernels to use fp16-friendly fastAtomicAdd and moving grid_sampler to the promote list). But that's more invasive, and for 1.9 ngimel says this simple temporary fix is preferred.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/58679
Reviewed By: soulitzer
Differential Revision: D28576559
Pulled By: ngimel
fbshipit-source-id: d653003f37eaedcbb3eaac8d7fec26c343acbc07
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52859
This reverts commit 92a4ee1cf6.
Added support for bfloat16 for CUDA 11 and removed fast-path for empty input tensors that was affecting autograd graph.
Test Plan: Imported from OSS
Reviewed By: H-Huang
Differential Revision: D27402390
Pulled By: heitorschueroff
fbshipit-source-id: 73c5ccf54f3da3d29eb63c9ed3601e2fe6951034
Summary:
[index_put](https://pytorch.org/docs/master/tensors.html#torch.Tensor.index_put) requires src and dst tensors to be the same dtype, so imo it belongs on the promote list when autocast is active (output should be widest dtype among input dtypes).
i also put some other registrations in alphabetical order.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41035
Differential Revision: D22418305
Pulled By: ngimel
fbshipit-source-id: b467cb16ac6c2ba1f9e43531f69a144b17f00b87
Summary:
Currently, a custom autograd function written with
```
torch.cuda.amp.custom_fwd(cast_inputs=dtype)
def forward(ctx, *args):
...
```
casts incoming floating-point CUDA tensors to `dtype` unconditionally, regardless of whether the function executes in an autocast-enabled region. I think I had the wrong idea there. Autocast-disabled regions should give the user control of input types. Also, `custom_fwd(cast_inputs=dtype)`-decorated functions' behavior should align with native fp32list/fp16list functions. C++-side casting wrappers have no effect when autocast is disabled, and `custom_fwd`'s casting should behave the same way.
The present PR changes `custom_fwd` so it only casts in autocast-enabled regions (also updates custom_fwd to ignore fp64 inputs, like the C++ wrappers).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36171
Differential Revision: D22179511
Pulled By: ngimel
fbshipit-source-id: 5a93d070179a43206066bce19da0a5a19ecaabbd
Summary: NVIDIA's Apex is updating to no longer rely on this behavior, but we're reverting this Python2->Python3 update to unblock internal apex users.
Test Plan: Sandcaslte + OSS CI.
Reviewed By: ngimel
Differential Revision: D22146782
fbshipit-source-id: f9483d2cbf9dc3a469ad48a6c863edea3ae51070
Summary:
BC-breaking note:
If a user is using one of these dunders directly they will not longer be available. Users should update to Python3 compatible dunders.
Original PR note:
`__div__` (and `__idiv__` and `__rdiv__`) are no longer special dunders in Python3. This PR replaces them with the `__truediv__` (`__itrudediv__`, `__rtruediv__`) dunders, since we no longer support Python2.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39151
Differential Revision: D22075713
Pulled By: mruberry
fbshipit-source-id: d318b47b51f7cc4c3728b1606a34d81e49ba0fa1
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:
Also, windows memory failures responsible for the earlier reversion have been fixed.
This PR (initially) contains 2 commits:
* a revert of the revert
* all changes to implement the original Apex scale update heuristic, squashed into a single commit for easier diff review
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33366
Differential Revision: D20099026
Pulled By: ngimel
fbshipit-source-id: 339b9b6bd5134bf055057492cd1eedb7e4461529
Summary:
This PR implements the gradient scaling API that mruberry, jjsjann123, ngimel, zdevito, gchanan and I have been discussing. Relevant issue/RFC: https://github.com/pytorch/pytorch/issues/25081.
Volume-wise, this PR is mostly documentation and tests. The Python API (found entirely in `torch/cuda/amp/amp_scaler.py`) is lightweight . The exposed functions are intended to make the implementation and control flow of gradient scaling convenient, intuitive, and performant.
The API is probably easiest to digest by looking at the documentation and examples. `docs/source/amp.rst` is the homepage for the Automatic Mixed Precision package. `docs/source/notes/amp_examples.rst` includes several examples demonstrating common but not-immediately-obvious use cases. Examples are backed by tests in `test_cuda.py` (and thankfully the tests pass :P).
Two small utility kernels have been added in `native/cuda/AmpKernels.cu` to improve performance and avoid host-device synchronizations wherever possible.
Existing optimizers, both in the wild and in Pytorch core, do not need to change to use the scaling API.
However, the API was also designed to establish a contract between user scripts and optimizers such that writers of _new_ custom optimizers have the control points they need to implement fast, optionally sync-free updates. User scripts that obey the scaling API can drop such custom optimizers in and reap performance benefits without having to change anything aside from the optimizer constructor itself. [I know what the contract with custom optimizers should be](35829f24ef/torch/cuda/amp/amp_scaler.py (L179-L184)), but I'm waiting for review on the rest of the API before I go about documenting it (it will be given a dedicated section in `docs/source/notes/amp_examples.rst`.
Currently, the gradient scaling examples do not include the auto-casting API as discussed in https://github.com/pytorch/pytorch/issues/25081. The gradient scaling API is intended to be orthogonal/modular relative to autocasting. Without auto-casting the gradient scaling API is fully use-_able_, but not terribly use-_ful_, so it's up to you guys whether you want to wait until auto-casting is ready before merging the scaling API as well.
### Todo
- [ ] How do I get c10 registered status for my two custom kernels? They're very simple.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26512
Differential Revision: D19859905
Pulled By: mruberry
fbshipit-source-id: bb8ae6966214718dfee11345db824389e4286923
