Summary:
Add a new device type 'XPU' ('xpu' for lower case) to PyTorch. Changes are needed for code related to device model and kernel dispatch, e.g. DeviceType, Backend and DispatchKey etc.
https://github.com/pytorch/pytorch/issues/48246
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49786
Reviewed By: mrshenli
Differential Revision: D25893962
Pulled By: ezyang
fbshipit-source-id: 7ff0a316ee34cf0ed6fc7ead08ecdeb7df4b0052
Summary:
This PR fixes unexpected `SystemError` when warnings are emitted and warning filters are set.
## Current behavior
```
$ python -Werror
>>> import torch
>>> torch.range(1, 3)
UserWarning: torch.range is deprecated in favor of torch.arange and will be removed in 0.5. Note that arange generates values in [start; end), not [start; end].
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
SystemError: <built-in method range of type object at 0x7f38c7703a60> returned a result with an error set
```
## Expected behavior
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
UserWarning: torch.range is deprecated and will be removed in a future release because its behavior is inconsistent with Python's range builtin. Instead, use torch.arange, which produces values in [start, end).
```
## Note
Python exception must be raised if `PyErr_WarnEx` returns `-1` ([python docs](https://docs.python.org/3/c-api/exceptions.html#issuing-warnings)). This PR fixes warnings raised in the following code:
```py
import torch
torch.range(1, 3)
torch.autograd.Variable().volatile
torch.autograd.Variable().volatile = True
torch.tensor(torch.tensor([]))
torch.tensor([]).new_tensor(torch.tensor([]))
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/44371
Reviewed By: mrshenli
Differential Revision: D23598410
Pulled By: albanD
fbshipit-source-id: 2fbcb13fe4025dbebaf1fd837d4c8e0944e05010
Summary:
This PR moves `DispatchKey::Autograd` to an alias dispatch key mapping to `AutogradCPU, AutogradCUDA, AutogradXLA, AutogradOther, AutogradPrivate*` keys.
A few things are handled in this PR:
- Update alias dispatch key mapping and precompute dispatchTable logic
- Move `Autograd` key from `always_included` set to TensorImpl constructor.
- Update `dummyTensor` constructor to take `requires_grad` as optional argument so that it's closer to the real application in op_registration_test.
- Use `BackendSelect` key for both backend select before and after autograd layer. (1 liner in backend_select codegen)
A few planned followups ordered by priority:
- [cleanup] Update `test_dispatch.py` to include testing `Autograd`.
- [cleanup] Add Math alias key and move catchAll to Math. (to remove 2.2 in `computeDispatchTableEntryWithDebug`)
- [new feature] Add support for Math in native_functions.yaml
- [cleanup] Add iterator like functionality to DispatchKeySet
- [cleanup/large] Only add Autograd backend keys when tensor requires grad. (cc: ljk53 ?)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/43070
Reviewed By: ezyang
Differential Revision: D23281535
Pulled By: ailzhang
fbshipit-source-id: 9ad00b17142e9b83304f63cf599f785500f28f71
Summary:
**BC-Breaking Note**
This PR changes the behavior of the torch.tensor, torch.as_tensor, and sparse constructors. When given a tensor as input and a device is not explicitly specified, these constructors now always infer their device from the tensor. Historically, if the optional dtype kwarg was provided then these constructors would not infer their device from tensor inputs. Additionally, for the sparse ctor a runtime error is now thrown if the indices and values tensors are on different devices and the device kwarg is not specified.
**PR Summary**
This PR's functional change is a single line:
```
auto device = device_opt.has_value() ? *device_opt : (type_inference ? var.device() : at::Device(computeDeviceType(dispatch_key)));
```
=>
```
auto device = device_opt.has_value() ? *device_opt : var.device();
```
in `internal_new_from_data`. This line entangled whether the function was performing type inference with whether it inferred its device from an input tensor, and in practice meant that
```
t = torch.tensor((1, 2, 3), device='cuda')
torch.tensor(t, dtype=torch.float64)
```
would return a tensor on the CPU, not the default CUDA device, while
```
t = torch.tensor((1, 2, 3), device='cuda')
torch.tensor(t)
```
would return a tensor on the device of `t`!
This behavior is niche and odd, but came up while aocsa was fixing https://github.com/pytorch/pytorch/issues/40648.
An additional side affect of this change is that the indices and values tensors given to a sparse constructor must be on the same device, or the sparse ctor must specify the dtype kwarg. The tests in test_sparse.py have been updated to reflect this behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41984
Reviewed By: ngimel
Differential Revision: D22721426
Pulled By: mruberry
fbshipit-source-id: 909645124837fcdf3d339d7db539367209eccd48
Summary:
AT_CHECK has been deprecated and provides no more features than
TORCH_CHECK
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34846
Differential Revision: D20481339
Pulled By: mrshenli
fbshipit-source-id: 1777e769a069a78e03118270294e5e273d516ca7
Summary:
Fixes https://github.com/pytorch/pytorch/issues/33899
In the issue, we have
```
TypeError("expected %s (got %s)", dispatch_key, toString(other.key_set()).c_str());
```
which results in `dispatch_key` being interpreted as a c-string by `sprintf`. Adding `__attrbute__((format))` to the `TypeError` constructor allows gcc or clang to detect this at compile time. Then `-Werror=format` makes it a hard error at compile time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34019
Differential Revision: D20194842
Pulled By: ezyang
fbshipit-source-id: fa4448916c309d91e3d949fa65bb3aa7cca5c6a8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31490
When this happens, a dense tensor is constructed from a sparse constructor.
Fixes: https://github.com/pytorch/pytorch/issues/16154
Test Plan: Imported from OSS
Reviewed By: cpuhrsch, mrshenli
Differential Revision: D19196498
Pulled By: gchanan
fbshipit-source-id: 57a6324833e35f3e62318587ac74267077675b93
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31117
After this diff, we will have completely removed the named tensor
feature flagging. This means that named tensors are always on and that
there is no mechanism to turn them off. There should be no more follow-up
diffs.
I performed the deletion of the header with
```
find . -type f -print0 | xargs -0 sed -i '/#include
<ATen\/core\/EnableNamedTensor.h>/d'
```
Test Plan: - wait for CI
Differential Revision: D18934952
Pulled By: zou3519
fbshipit-source-id: 253d059074b910fef15bdf885ebf71e0edf5bea5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30894
This PR begins the process of removing BUILD_NAMEDTENSOR macros. There
will be followups.
Reasons for removing the macros:
- BUILD_NAMEDTENSOR is always on and has been on since pytorch 1.3.0.
- Since we don't test building without it, it is useless to keep around.
- Code becomes nicer to read without the macros
Reasons for not removing the macros:
- potential for feature flagging
Now, I argue against needing to feature flag. The main reason why we
might want to feature flag is if we need to disable the feature.
We'd need a fast switch to disable the feature if someone discovers
in the future that named tensors caused some regression in some existing workflows.
In https://github.com/pytorch/pytorch/pull/25798, I did a variety of
macro- and micro- benchmarks to determine the performance impact of named
tensors on regular tensors.
[The
microbenchmarks](https://github.com/pytorch/pytorch/pull/25798#issuecomment-529014810)
were not very stable, and running the
microbenchmarks for more iterations doesn't actually help because the
noise is not distributed in a nice way. Instead of microbenchmarks I ran
a [profiler
(perf)](https://github.com/pytorch/pytorch/pull/25798#issuecomment-555707645)
to estimate how much overhead named tensors add to unnamed code. I
estimated the overhead to be less than 100ns for `add` and even smaller
for `mm`; there are ways to optimize even futher if we find this to be a
problem.
[Initial
macrobenchmarks](https://github.com/pytorch/pytorch/pull/25798#issuecomment-530539104)
were also not very stable. I ran imagenet for some number of epochs. To
make them more stable, I got rid of the data loading (which seemed to
vary between runs). [In some benchmarkers without data
loading](https://github.com/pytorch/pytorch/pull/25798#issuecomment-562214053),
we can see that the results are less noisy now. These results support
no noticeable regressions in speed.
Test Plan: - wait for CI
Differential Revision: D18858543
Pulled By: zou3519
fbshipit-source-id: 08bf3853a9f506c6b084808dc9ddd1e835f48c13
Summary:
Given that pybind11 implements these gil functions, I don't think it makes sense for Pytorch to have its own bespoke versions.
Fixes https://github.com/pytorch/pytorch/issues/29065
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29095
Differential Revision: D18301806
Pulled By: ezyang
fbshipit-source-id: 03da6a26c41ee65aaadf7b67b9f0b14d2def2a5a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29213
A trivial use of make_variable is one where requires_grad=False. This
transformation is not technically semantics preserving, as make_variable
will create a shallow copy of the tensor in question; however, I
am guessing that we have the invariant that we don't actually make
use of this shallow copy in a nontrivial way.
There were some cases where the surrounding code expected a Variable proper
to be returned; I retained those sites.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Differential Revision: D18353503
Pulled By: ezyang
fbshipit-source-id: 57fe34d82e009c0cc852266fb0b79d6d9c62bb03
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29239
There were a few main changes, summarized below.
Rename `propagate_names`
----------------------------------------------
There are two main APIs now, `propagate_names_if_nonempty(Tensor&,
ArrayRef<Dimname>)` and `propagate_names(Tensor&, ArrayRef<Dimname>)`
The former propagates names if they are not empty and the latter
unconditionally tries to propagate names.
`names` can be empty if name inference did not occur (see the next
section).
Removed usages of `optional` in name inference
----------------------------------------------
Previously, we used `optional<ArrayRef<Dimname>>` and
`optional<vector<Dimname>>`. `nullopt` represens that no name inference
happened.
The problem with this is that these types are not implicitly convertible
to each other and dealing with them is painful as a result (users have
to manually unwrap `optional<vector>` and convert to
`optional<arrayref>`.
To fix this, I rewrote most named inference functions to use an empty array as an
indicator value:
- If an array is empty, then no name inference occured
- If an array is not empty, then name inference occured.
Removed `vector<Dimname>&&` overloads
----------------------------------------------
These were originally meant for efficiency: instead of copying a vector
of names we could move it directly inside the tensor and replace the old
names. However, looking around the code base, we do copies for
`IntArrayRef` for sizes and strides instead of optimizing them, so the
perf gain is probably not critical. I removed `vector<Dimname>&&` overloads
to stop optimizing prematurely.
Furthermore, one potential design for a faster named inference api is
to construct names directly on a tensor's names object; in this design
there is also no `vector<Dimname>&&` overload.
Plans
----------------------------------------------
After this PR I'll keep attempting to cleaning up `propagate_names`
functions. There are a lot of `propagate_names_for_{blah}` functions
that exist that probably don't need to.
Test Plan: - `python test/test_namedtensor.py -v`
Differential Revision: D18350090
Pulled By: zou3519
fbshipit-source-id: eb5dd6cbd2d4f1838431db5edbdb207204c5791d
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/28620
All Tensors are Variables now, they just happen to have requires_grad=False. Tensors ALWAYS have `VariableTensorId` in their type set.
When constructing this patch, I had to make decisions about what I would fix in this patch, and what I would leave for follow up PRs. Here is the cleanup that happens in this patch:
- The `is_variable` property is removed from TensorOptions. I removed this immediately because unlike Tensor::is_variable, TensorOptions::is_variable doesn't respect our VariableTensorId thread-local state. This means that there were a bunch of places where TensorOptions::is_variable was false, which is obviously bogus in the world when tensor and variable are merged. Instead of keeping the method as a function that always returns true, I just opted to remove it entirely (it's not public API.) All places we set `is_variable` are deleted.
- Knock on effect: there is no longer a separate DeprecatedTypeProperties for the variable and non-variable versions of type.
- Knock on effect: instead of asserting on TensorOptions::is_variable, instead we just test `at::impl::variable_is_excluded()`
- There is now only one copy of the cuDNN RNN dropout cache, not two (I'm not sure why we had two to begin with)
Some cleanup that doesn't happen in this patch:
- Eliminating unnecessary uses of `make_variable`
- Eliminating `Tensor::is_variable`
The most subtle part of this patch is retaining tracing behavior: the fact that everything is a Variable means that more code gets routed to VariableType than before; this can change traces. I identified two places where we didn't appropriately turn off VariableType, mostly factory functions:
- `torch.tensor` must turn off VariableType before invoking `at::empty` to construct the tensor, as it subsequently does direct data access
- `tensor_slow` (invoked when you pass a Python scalar to a tensor argument) must turn off VariableType before calling `scalar_to_tensor` so the scalar gets traced as constant, rather than as a call to `scalar_to_tensor`.
Honestly, these are all giant hacks, and should be replaced with a more specialized guard that just toggles tracing.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Reviewed By: dreiss
Differential Revision: D18171156
Pulled By: ezyang
fbshipit-source-id: 5b6a045beba37492647e350190f495114e86504d
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26060
This PR enables BUILD_NAMEDTENSOR by default. This is done via including
a header, `c10/core/EnableNamedTensor`, that sets `BUILD_NAMEDTENSOR`.
In the future, the plan is to get rid of the flag entirely: we can
incrementally delete usages after this PR goes in.
This PR also maintains the namedtensor ci vs regular ci distinction.
`test/test_namedtensor.py` only runs if TEST_NAMEDTENSOR=1 is specified.
TEST_NAMEDTENSOR=1 is set on the namedtensor ci. I'll remove this
distinction later and send out an announcement about it; devs will be
responsible for named tensor failures after that.
The initial reason why we had the BUILD_NAMEDTENSOR flag was so that we
could quickly prototype named tensor features without worrying about
adding overhead to the framework. The overheads can be categorized as
memory overhead and performance overhead.
Memory overhead: named tensors adds 1 additional word per Tensor. This
is because TensorImpl stores a `unique_ptr<NamedTensorMetaInterface>`
field. This is not a lot of overhead.
Performance overhead: At all entry points to name inference, we check
if inputs to an op are named. If inputs are not named, we short-circuit
and don't do name inference. These calls should therefore be as
efficient as error-checking code and not take up a lot of time.
My plan is to benchmark a few functions and then post the results in a
comment to this PR.
Test Plan: - [namedtensor ci]
Differential Revision: D17331635
Pulled By: zou3519
fbshipit-source-id: deed901347448ae2c26066c1fa432e3dc0cadb92
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25424
Test Plan
- new tests [namedtensor ci]
Test Plan: Imported from OSS
Differential Revision: D17120399
Pulled By: zou3519
fbshipit-source-id: 93d7944f2ec4c5a7256f505323b879af706131df
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25308
Instead of storing a single TensorTypeId in a Tensor, we store a bitset of tensor type IDs in a Tensor, TensorTypeSet. This class comes with some unit tests. This is in preparation for making Variable a TensorTypeId. In order to help flush out places where this makes a semantic difference, we rename `Tensor::type_id()` to `Tensor::type_set()` and smoke out all of the locations where this was semantically meaningful.
Because the new tensor type set is 64-bits, this increases the size of Tensor by a word.
Listing of semantic changes:
* Many TensorImpl related constructors just propagate TensorTypeId to a parent constructor. These are pretty simple to adjust.
* Backend extensions are now in the business of explicitly constructing a TensorTypeSet and then passing it in. This is probably OK for now but when Variable drops, these dispatch IDs may get immediately overwritten to have Variable set.
* `sparseTensorSetToDeviceType` and similar functions previously did an equality test with TensorTypeId, to determine what an appropriate device type is. This equality is now replaced with a set inclusion test. This is valid, under the assumption that we don't ever have weird sets like "this tensor is simultaneously a sparse CPU tensor and a sparse CUDA tensor", which will be true in the short term plan of adding Variable to the dispatch ID.
* `impl::dispatchTypeId` was generally introduced for cases where we legitimately need to convert from `TensorTypeSet -> TensorTypeId` in a dispatch related manner. At the moment, the implementation is trivial, but they will soon be adjusted to handle TLS. I've tried to make these call sites as forwards compatible as possible:
* `checked_tensor_unwrap` and co now use `dispatchTypeId`. When Variable is added to the type set, these will always be called in a context where the Variable type ID is disabled, so we will get the correct underlying tensor type ID.
* Uses of `Backend` in dispatch are now replaced with `TensorTypeSet`. The general heuristic here for whether or not to accept a `TensorTypeId` or `TensorTypeSet` is that we want to make the generated code as simple as possible. It is easier to retrieve a `TensorTypeSet`, so that's a more appropriate API in these cases.
* In some cases, I could not conveniently switch an implementation to the new semantics, because it was blocked on some other refactor. In this case, I introduced `legacyExtractTypeId`, which gives what would be a BC-compatible `TensorTypeSet` to `TensorTypeId` implementation that will continue to report the same values it would have prior to this change. This is **different** from `dispatchTypeId`, because this function does NOT respect TLS; it always ignores Variable type IDs.
* c10 dispatcher tests, which are oblivious to Variable dispatch, use this BC function (actually, they use `extractTypeId`, an overload for Tensor.
* The implementation of `new_*` methods heavily relies on tensor type ID, I chose not to unwind this. PR to refactor this at https://github.com/pytorch/pytorch/pull/25475
* Slicing also relies on tensor type ID, see `torch/csrc/autograd/python_variable_indexing.cpp` (though in some cases in this file, I was able to replace use of tensor type ID with TensorOptions)
* In some cases, there is an equality test on tensor type ID which would be better done by testing "tensor axes". In those cases, I replaced those equality tests with more equality tests.
* Example: `torch/csrc/nn/type_checks.h`
* There is a total punt in `torch/csrc/tensor/python_tensor.cpp` where "instance of" checking is done via dispatch ids. In general, the Variable-ness of a tensor doesn't participate in instanceof testing. It's not entirely clear what to do here.
* Instead of storing `Backend` in `VariableInfo`, we now just store Layout.
c10 dispatcher test updates were done with:
```
:%s/\([^ ]\+\)\.type_id()/extractTypeId(\1)/g
:%s/\([^( ]\+\)->type_id()/extractTypeId(*\1)/g
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25308
Differential Revision: D17092791
Test Plan: sandcastle and ossci
Reviewed By: bwasti
Pulled By: ezyang
fbshipit-source-id: 22207d14fe62dd31ee19cc5011af22e3d9aabb5b
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25475
I got sucked into this rabbit hole when I was trying to understand
what I should do with TensorTypeId occurrences in
torch/csrc/utils/tensor_new.cpp. I eventually concluded that all of my problems
were because Tensor.new_empty was hand implemented and not actually a native
function. So I made it a native function.
There are a bunch of other new_* functions which should get this
treatment, but I'm sending out this PR just to show how it can
be done.
The general recipe:
1. Implement a concept of TensorOptions merging (TensorOptions::merge_in).
This represents the notion of taking a tensor, but "overriding" some
of its values with specific overrides. One subtlety here is how
devices get merged; see the comments for what our existing behavior is,
and how I preserve it.
2. Implement new_empty as a native function, using options merging.
3. Add another special case to Python binding generation to treat new_*
similar to *_like (i.e., handle TensorOptions correctly). The logic
here is probably wrong, actually; we should codegen TensorOptions
correctly no matter what happens, but new_empty follows the same
pattern as empty_like so I opted not to touch this code too much.
4. Delete the now defunct manual binding code.
5. Delete manual type annotations that are no longer necessary since
we're going through native.
I didn't handle memory format correctly here. I don't know if this function
should accept memory format; prior memory format patches didn't add support
for memory format to new_like. If we had put memory format in TensorOptions
this wouldn't have been a question.
ghstack-source-id: 89294185
Test Plan: sandcastle & ossci
Differential Revision: D17133000
fbshipit-source-id: 00f4e98bd5174f6fd54e8aba2910ea91824771d9
Summary:
This PR implements auto-conversion of GPU arrays that support the `__cuda_array_interface__` protocol (fixes#15601).
If an object exposes the `__cuda_array_interface__` attribute, `touch.as_tensor()` and `touch.tensor()` will use the exposed device memory.
#### Zero-copy
When using `touch.as_tensor(...,device=D)` where `D` is the same device as the one used in `__cuda_array_interface__`.
#### Implicit copy
When using `touch.as_tensor(...,device=D)` where `D` is the CPU or another non-CUDA device.
#### Explicit copy
When using `torch.tensor()`.
#### Exception
When using `touch.as_tensor(...,device=D)` where `D` is a CUDA device not used in `__cuda_array_interface__`.
#### Lifetime
`torch.as_tensor(obj)` tensor grabs a reference to `obj` so that the lifetime of `obj` exceeds the tensor
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20584
Differential Revision: D15435610
Pulled By: ezyang
fbshipit-source-id: c423776ba2f2c073b902e0a0ce272d54e9005286
Summary:
Make it possible to construct a pinned memory tensor without creating a storage first and without calling pin_memory() function. It is also faster, as copy operation is unnecessary.
Supported functions:
```python
torch.rand_like(t, pin_memory=True)
torch.randn_like(t, pin_memory=True)
torch.empty_like(t, pin_memory=True)
torch.full_like(t, 4, pin_memory=True)
torch.zeros_like(t, pin_memory=True)
torch.ones_like(t, pin_memory=True)
torch.tensor([10,11], pin_memory=True)
torch.randn(3, 5, pin_memory=True)
torch.rand(3, pin_memory=True)
torch.zeros(3, pin_memory=True)
torch.randperm(3, pin_memory=True)
torch.empty(6, pin_memory=True)
torch.ones(6, pin_memory=True)
torch.eye(6, pin_memory=True)
torch.arange(3, 5, pin_memory=True)
```
Part of the bigger: `Remove Storage` plan.
Now compatible with both torch scripts:
` _1 = torch.zeros([10], dtype=6, layout=0, device=torch.device("cpu"), pin_memory=False)`
and
` _1 = torch.zeros([10], dtype=6, layout=0, device=torch.device("cpu"))`
Same checked for all similar functions `rand_like`, `empty_like` and others
It is fixed version of #18455
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18952
Differential Revision: D14801792
Pulled By: VitalyFedyunin
fbshipit-source-id: 8dbc61078ff7a637d0ecdb95d4e98f704d5450ba
