Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53973
Two parts to this PR; I had to put them together because adding support for X causes more test code to be exercised, which in turn may require a fix for Y.
The first part is restoring the concept of storage to meta tensors. Previously, meta tensors had a nullptr storage (e.g., `meta_tensor.storage()` is an error.) As I was increasing the coverage of meta tensors, I started running into test cases (specifically memory overlap tests) that were failing because not having storage meant I couldn't check for memory overlap. After some discussion, we decided that it would make sense for meta tensors to model this as well (we already model strides, so getting accurate view information also seems useful). This PR does that by:
* Rewrite all of the factory functions in MetaTensor.cpp to use the generic versions (which are very carefully written to not actually poke at the data pointer, so everything works out). The key idea here is we give meta tensors a special allocator, MetaAllocator, which always returns a nullptr even if you ask for a nonzero number of bytes. resize_ is also made generic; the normal variant can be used directly rather than having to instruct it to avoid resizing storage
* Turn on memory overlap checking in TensorIterator even for meta tensors
* Although meta tensors now have storage, the concept of meta storage is NOT exposed to Python land (as it would imply I would have to codegen MetaFloatStorage, MetaDoubleStorage, etc. classes). So `x.storage()` still raises an error and I have a cludge in `__deepcopy__` to break storage sharing upon deep copy (this is wrong, but no tests exercise this at the moment).
The second part is adding more support for the most used functions in the test suite.
* Inplace operations have very simple meta functions. I added `fill_`, `zero_`, `random_`, `uniform_` and `normal_`. In the case of random, I take advantage of pbelevich's templates for defining random kernels, so that I can reuse the common scaffolding, and then just register a noop stub that actually does the RNG. (Look, another structured kernels tiny variant!)
* `copy_` is now implemented. Copying into a meta tensor is always OK, but copying out of a meta tensor raises an error (as we don't know what the "correct" data to copy out is in this case)
* `empty_strided` usage from structured kernels now is implemented (TBH, this could have been done as soon as `empty_strided` was added)
* Meta was missing in a few places in TensorOptions/DispatchKey utility functions, so I added them
* Autograd engine now correctly homes meta tensors with CPU tensors (they have -1 device index so CUDA queues wouldn't work anyway)
* `apply_`, `map_` and `map2_` are special cased to no-op on meta tensor self. These count as inplace operations too but they are implemented a little differently.
Getting more meta function support triggers a number of bugs in the test suite, which I then fix:
- Linear algebra functions sometimes don't report NotImplementedError because they get swallowed by catch all try blocks. This is tracked in https://github.com/pytorch/pytorch/issues/53739
- dlpack obviously doesn't work with meta tensors, I just disabled the test
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D27036572
Test Plan: Imported from OSS
Reviewed By: agolynski, bdhirsh
Pulled By: ezyang
fbshipit-source-id: 7005ecf4feb92a643c37389fdfbd852dbf00ac78
Summary:
```
index_add(Tensor self, int dim, Tensor index, Tensor source) -> Tensor
```
now becomes
```
index_add(Tensor self, int dim, Tensor index, Tensor source, Scalar alpha=1) -> Tensor
```
Generally, this sounds useful and harmless, and inside PyTorch, we are already needing this feature in `add_out_dense_sparse_cuda`, see the `SparseCUDATensorMath.cu` change in this PR.
**Test not added yet. Will add if after discussion we believe this is a good idea.**
- [ ] TODO: add test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54176
Reviewed By: ngimel
Differential Revision: D27319198
Pulled By: mruberry
fbshipit-source-id: fe43be082d1230c87c5313458213d5252be2ff23
Summary:
Added the support for half / bfloat / bool for `index_select`, as suggested by ngimel in
https://github.com/pytorch/pytorch/issues/49707#issuecomment-788140578
For the tests to pass, I also added the support for `index_add`.
I added `OpInfo` tests for `index_add` and more thorough forward tests for `index_select` to test these changes.
While doing so, I found that the support for scalar types in the derivative of `index_add` was not correct, so I corrected it.
Resolves https://github.com/pytorch/pytorch/issues/49707
It should also resolve similar issues that I encountered when porting `index_copy`, `take` and `put`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53898
Reviewed By: mruberry
Differential Revision: D27193294
Pulled By: ngimel
fbshipit-source-id: 5a0af2c62a0cf24f3cc9c74f230ab4f3712bbb7a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54079
Fixes https://github.com/pytorch/pytorch/issues/53815
Instead of testing if something is CUDA, we instead test if something
is not CPU. This in the general theming of "Don't be so darn CUDA
centric".
Intruigingly, we didn't have a is_cpu() method on Tensor. Which seems
like a big oversight and one of the reasons how we ended up in this
mess. So in it goes. Maybe we should also get this for Python bindings
as well (but in that case, should probably look into redoing all of the
is_X bindings so they aren't done manually).
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Reviewed By: ngimel
Differential Revision: D27109507
Pulled By: ezyang
fbshipit-source-id: abbe72c2e688c452ffe098d206cb79938b5824b1
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53759Fixes#53587, see issue for in-depth explanation of the bug.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Reviewed By: albanD
Differential Revision: D26971342
Pulled By: ezyang
fbshipit-source-id: 805983fed2658e27fb033f36a71fd30950a29328
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53665
ngimel pointed out to me where we already test the behavior of the `Upsample` ops in `test_nn.py`. This PR deleting my bespoke tests in `test_torch.py` and updates those in `test_nn.py` to test memory format properly.
There were two reasons the original test didn't pick up on a memory format regression:
- They didn't test the memory format of the output tensor explicitly, i.e. `output.is_contiguous(memory_format=...)`
- Even with that change, the test tensors were to simple to fail the tests. From some trial and error, it looks like one of the first two dimensions in the inputs needs to be > 1 in order for the `channels_last` memory format to actually re-order the strides.
Test Plan: Imported from OSS
Reviewed By: ngimel
Differential Revision: D26929683
Pulled By: bdhirsh
fbshipit-source-id: d17bc660ff031e9b3e2c93c60a9e9308e56ea612
Summary:
Fixes https://github.com/pytorch/pytorch/issues/48841 for half datatype (it was fixed for other datatypes before).
The reason for https://github.com/pytorch/pytorch/issues/48841 happening for half was that `exponential_` for half was producing 0s.
Exponential distribution implementation on cuda is here e08aae2613/aten/src/ATen/native/cuda/DistributionTemplates.h (L535-L545)
with `transformation::exponential` defined here
e08aae2613/aten/src/ATen/core/TransformationHelper.h (L113-L123)
It takes a uniformly distributed random number and takes `log` of it. If necessary, the result is then converted to low precision datatype (half). To avoid 0's, before applying `log`, ones are replaced with std::nextafter(1,0). This seems fine, because log(1-eps) is still representable in half precision (`torch.tensor([1.], device="cuda").nextafter(torch.tensor([0.], device="cuda")).log().half()` produces 5.96e-8) , so casting to `scalar_t` should work. However, since fast log approximation is used (`__logf`), the log result is ~3e-9 instead of more accurate 5.96e-8, and underflows when casting to half. Using `::log` instead of fast approximation fixes it, however, it comes with ~20% perf penalty on exponential kernel for fp32 datatype, probably more for half.
Edit: alternative approach used now is to filter all small values returned by transformation. The result is equivalent to squashing of 1's to 1-eps that was used before, and computing correct log of 1-eps (which is -eps, exactly equal even for doubles). This doesn't incur noticeable performance hit.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53480
Reviewed By: mruberry
Differential Revision: D26924622
Pulled By: ngimel
fbshipit-source-id: dc1329e4773bf91f26af23c8afa0ae845cfb0937
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53535
During the port to structured kernels for upsample kernels, I missed that a subset of them explicitly pass `memory_format` information from the input to the output tensors.
Note 1:
I added the logic into the `meta` function of each op, which feels morally correct since this logic affects the output shape/metadata. One consequence is that all backend implementations will get the logic. I synced with fmassa that this seems reasonable.
Note 2:
This logic used to happen in the following operators, which this PR fixes:
- upsample_nearest3d
- upsample_trilinear3d
- upsample_nearest2d
- upsample_bilinear2d
I explicitly didn't patch the other upsample kernels, which look like they never forwarded memory_format information:
- `upsample_bicubic2d` (maybe this should though? `UpSampleBicubic2d.cpp` isn't currently written to do anything different for `channels_last` tensors)
- All of the `upsample_{mode}1d` operators. Probably because, afaik, channels_last isn't supported for 3d tensors
- The corresponding backwards operator for every upsample op.
Note 3:
I'm also wondering why memory_format isn't just directly a part of the `tensor::options()` method, which would cause all ops to universally forward memory_format information from input to output tensors, rather than just the upsample ops. My guess is:
- BC-breakage. I'm not sure whether this would really *break* people, but it's an API change
- performance. `tensor::options()` is called everywhere, and adding a call to `suggest_memory_format()` would probably noticeably hit microbenchmarks. We could probably deal with that by making `memory_format` a precomputed field on the tensor?
Test Plan: Imported from OSS
Reviewed By: H-Huang
Differential Revision: D26891540
Pulled By: bdhirsh
fbshipit-source-id: b3845f4dd5646b88bf738b9e41fe829be6b0e5cf
Summary:
Helps make master green by removing this hefty memory allocating from CPU test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53561
Reviewed By: malfet, albanD
Differential Revision: D26897941
Pulled By: janeyx99
fbshipit-source-id: 9f6c2d55f4eea1ab48665f7819fc113f21991036
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53276
- One of the tests had a syntax error (but the test
wasn't fine grained enough to catch this; any error
was a pass)
- Doesn't work on ROCm
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D26820048
Test Plan: Imported from OSS
Reviewed By: mruberry
Pulled By: ezyang
fbshipit-source-id: b02c4252d10191c3b1b78f141d008084dc860c45
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53143
Meta is now an honest to goodness device type, like cpu, so you can use
device='meta' to trigger allocation of meta tensors. This way better
than empty_meta since we now have working API for most factory functions
(they don't necessarily work yet, though, because need to register Meta
versions of those functions.)
Some subtleties:
- I decided to drop the concept of CPU versus CUDA meta tensors; meta
tensors are device agnostic. It's hard to say exactly what the
correct level of abstraction here is, but in this particular case
implementation considerations trump semantic considerations: it
is way easier to have just a meta device, than to have a meta device
AND a cpu device AND a cuda device. This may limit the applicability
of meta tensors for tracing models that do explicit cpu()/cuda()
conversions (unless, perhaps, we make those operations no-ops on meta
tensors).
- I noticed that the DeviceType uppercase strings are kind of weird.
Are they really supposed to be all caps? That's weird.
- I moved the Meta dispatch key to live with the rest of the "device"
dispatch keys.
- I intentionally did NOT add a Backend for Meta. For now, I'm going to
hope meta tensors never exercise any of the Backend conversion code;
even if it does, better to fix the code to just stop converting to and
from Backend.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Reviewed By: samestep
Differential Revision: D26763552
Pulled By: ezyang
fbshipit-source-id: 14633b6ca738e60b921db66a763155d01795480d
Summary:
Fixes https://github.com/pytorch/pytorch/issues/52213
Nans were previously inconsistently propagated due to std::min always returning first argument if one of the args in nan
when reduction functor was called on 2 `-inf` arguments, `std::min(x,y) - std::max(x,y)` resulted in `-inf - (-inf)` = nan, even though logcumsumexp is well defined for `-inf, -inf` pair.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52947
Reviewed By: H-Huang
Differential Revision: D26718456
Pulled By: ngimel
fbshipit-source-id: a44433889da352cc959786dd15b6361a68fcfed7
Summary:
This PR adds functionality to skip a test based on CUDA version.
This way, we can be more specific when skipping a test, such as when the test only fails for a particular CUDA version.
This allows us to add back the skipped tests for CUDA 11.2 for other CUDA versions, such as 10.1 and 11.1.
I tested this locally (by using 11.0 instead of 11.2), but will run all the CI to make sure it works.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52359
Reviewed By: walterddr
Differential Revision: D26487951
Pulled By: janeyx99
fbshipit-source-id: 45c71cc6105ffd9985054880009cf68ea5ef3f6a
Summary:
Fixes https://github.com/pytorch/pytorch/issues/51719, https://github.com/pytorch/pytorch/issues/28142
**Change**
- Update `torch.Tensor.unflatten` to support users pass`-1` as the inferred size for both tensors and named tensors.
- Examples of using `-1` in the `unflatten` function are added to the docs.
- Fix the rendered issue of original `unflatten` docs by removing a blank line between its example section.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51955
Reviewed By: agolynski
Differential Revision: D26467198
Pulled By: zou3519
fbshipit-source-id: 6a3ede25561223187273796427ad0cb63f125364
Summary:
Reference: https://github.com/pytorch/pytorch/issues/50006
We should probably add aliases for these operators to be consistent with NumPy names i.e. `np.degrees` and `np.radians`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51283
Reviewed By: ngimel
Differential Revision: D26171163
Pulled By: mruberry
fbshipit-source-id: 1869604ed400820d95f6ff50a0e3cba1de1ffa84
Summary:
Adding CUDA 11.2 to Windows CI.
Disabled tests:
The following ran into `CUDA error: misaligned address` for CUDA 11.2: (issue linked below)
`test_where_scalar_valid_combination_cuda_complex128` in test_torch.py
`test_sgn_complex_cuda` in test_autograd.py
The following ran into `CUDA error: too many resources requested for launch` for CUDA 11.2: (https://github.com/pytorch/pytorch/issues/52002)
test_EmbeddingBag_per_sample_weights_and_new_offsets_cuda_int64_float64
test_EmbeddingBag_per_sample_weights_and_offsets_cuda_int64_float64
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51598
Reviewed By: mrshenli
Differential Revision: D26344965
Pulled By: janeyx99
fbshipit-source-id: 3c9a4ed16d748969e96593220ec0a9f33e1ffcef
Summary:
Toward fixing https://github.com/pytorch/pytorch/issues/47624
~Step 1: add `TORCH_WARN_MAYBE` which can either warn once or every time in c++, and add a c++ function to toggle the value.
Step 2 will be to expose this to python for tests. Should I continue in this PR or should we take a different approach: add the python level exposure without changing any c++ code and then over a series of PRs change each call site to use the new macro and change the tests to make sure it is being checked?~
Step 1: add a python and c++ toggle to convert TORCH_WARN_ONCE into TORCH_WARN so the warnings can be caught in tests
Step 2: add a python-level decorator to use this toggle in tests
Step 3: (in future PRs): use the decorator to catch the warnings instead of `maybeWarnsRegex`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48560
Reviewed By: ngimel
Differential Revision: D26171175
Pulled By: mruberry
fbshipit-source-id: d83c18f131d282474a24c50f70a6eee82687158f
Summary:
Implements `np.diff` for single order differences only:
- method and function variants for `diff` and function variant for `diff_out`
- supports out variant, but not in-place since shape changes
- adds OpInfo entry, and test in `test_torch`
- automatic autograd because we are using the `Math` dispatch
_Update: we only support Tensors for prepend and append in this PR. See discussion below and comments for more details._
Currently there is a quirk in the c++ API based on how this is implemented: it is not possible to specify scalar prepend and appends without also specifying all 4 arguments.
That is because the goal is to match NumPy's diff signature of `diff(int n=1, int dim=-1, Union[Scalar, Tensor] prepend=None, Union[Scalar, Tensor] append)=None` where all arguments are optional, positional and in the correct order.
There are a couple blockers. One is c++ ambiguity. This prevents us from simply doing `diff(int n=1, int dim=-1, Scalar? prepend=None, Tensor? append=None)` etc for all combinations of {Tensor, Scalar} x {Tensor, Scalar}.
Why not have append, prepend not have default args and then write out the whole power set of {Tensor, Scalar, omitted} x {Tensor, Scalar, omitted} you might ask. Aside from having to write 18 overloads, this is actually illegal because arguments with defaults must come after arguments without defaults. This would mean having to write `diff(prepend, append, n, dim)` which is not desired. Finally writing out the entire power set of all arguments n, dim, prepend, append is out of the question because that would actually involve 2 * 2 * 3 * 3 = 36 combinations. And if we include the out variant, that would be 72 overloads!
With this in mind, the current way this is implemented is actually to still do `diff(int n=1, int dim=-1, Scalar? prepend=None, Tensor? append=None)`. But also make use of `cpp_no_default_args`. The idea is to only have one of the 4 {Tensor, Scalar} x {Tensor, Scalar} provide default arguments for the c++ api, and add `cpp_no_default_args` for the remaining 3 overloads. With this, Python api works as expected, but some calls such as `diff(prepend=1)` won't work on c++ api.
We can optionally add 18 more overloads that cover the {dim, n, no-args} x {scalar-tensor, tensor-scalar, scalar-scalar} x {out, non-out} cases for c++ api. _[edit: counting is hard - just realized this number is still wrong. We should try to count the cases we do cover instead and subtract that from the total: (2 * 2 * 3 * 3) - (3 + 2^4) = 17. 3 comes from the 3 of 4 combinations of {tensor, scalar}^2 that we declare to be `cpp_no_default_args`, and the one remaining case that has default arguments has covers 2^4 cases. So actual count is 34 additional overloads to support all possible calls]_
_[edit: thanks to https://github.com/pytorch/pytorch/issues/50767 hacky_wrapper is no longer necessary; it is removed in the latest commit]_
hacky_wrapper was also necessary here because `Tensor?` will cause dispatch to look for the `const optional<Tensor>&` schema but also generate a `const Tensor&` declaration in Functions.h. hacky_wrapper allows us to define our function as `const Tensor&` but wraps it in optional for us, so this avoids both the errors while linking and loading.
_[edit: rewrote the above to improve clarity and correct the fact that we actually need 18 more overloads (26 total), not 18 in total to complete the c++ api]_
Pull Request resolved: https://github.com/pytorch/pytorch/pull/50569
Reviewed By: H-Huang
Differential Revision: D26176105
Pulled By: soulitzer
fbshipit-source-id: cd8e77cc2de1117c876cd71c29b312887daca33f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51578https://github.com/pytorch/pytorch/pull/49710 introduced an edge case in which
drawing a single sample resulted in ignoring the `dtype` arg to `draw`. This
fixes this and adds a unit test to cover this behavior.
Test Plan: Unit tests
Reviewed By: danielrjiang
Differential Revision: D26204393
fbshipit-source-id: 441a44dc035002e7bbe6b662bf6d1af0e2cd88f4
Summary:
Performs the update that was suggested in https://github.com/pytorch/pytorch/issues/41489
Adjust the functionality to largely match that pf the scipy companion PR https://github.com/scipy/scipy/pull/10844/, including
- a new `draw_base2` method
- include zero as the first point in the (unscrambled) Sobol sequence
The scipy PR is also quite opinionated if the `draw` method doesn't get called with a base 2 number (for which the resulting sequence has nice properties, see the scipy PR for a comprehensive discussion of this).
Note that this update is a **breaking change** in the sense that sequences generated with the same parameters after as before will not be identical! They will have the same (better, arguably) distributional properties, but calling the engine with the same seed will result in different numbers in the sequence.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49710
Test Plan:
```
from torch.quasirandom import SobolEngine
sobol = SobolEngine(3)
sobol.draw(4)
sobol = SobolEngine(4, scramble=True)
sobol.draw(5)
sobol = SobolEngine(4, scramble=True)
sobol.draw_base2(2)
```
Reviewed By: malfet
Differential Revision: D25657233
Pulled By: Balandat
fbshipit-source-id: 9df50a14631092b176cc692b6024aa62a639ef61
Summary:
Reference: https://github.com/pytorch/pytorch/issues/33152
Changes
* Enable complex support for masked_scatter
* Enable half support for masked_scatter CPU
* Enable complex autograd support for masked_scatter CPU and masked_select (both CPU and CUDA).
**Note**:
Complex Support for masked_scatter CUDA is disabled as it depends on `masked_fill` which is yet to be ported to ATen.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51281
Reviewed By: ailzhang
Differential Revision: D26127561
Pulled By: anjali411
fbshipit-source-id: 6284926b934942213c5dfc24b5bcc8538d0231af
Summary:
Fixes https://github.com/pytorch/pytorch/issues/3307
Previously, `self.grad` was not ~cloned~ deepcopied to the returned tensor in `deepcopy`. Added a test and an implementation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/50663
Reviewed By: heitorschueroff
Differential Revision: D26074811
Pulled By: albanD
fbshipit-source-id: 536dad36415f1d03714b4ce57453f406ad802b8c
