Fix lint errors in test_autograd

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17812

Reviewed By: eellison

Differential Revision: D14388897

fbshipit-source-id: 6e2671805dc8d57af68eb0a0cd6ccb24d9db45e2

Added a few extra python bindings to help with walking the IR graph from Python (#17822)

Summary:
These changes add the following new Python bindings:

- Values have a 'type' property now that allows getting to the 'type' object
- Blocks have now inputs and outputs as well as returnNode and paramNode properties
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17822

Differential Revision: D14410123

Pulled By: ezyang

fbshipit-source-id: 64ef79f85a7a43b83e4b127b1d39efcaa64b74dc

kthvalue consistency with sort in the presence of NaN (#17824)

Summary:
This PR causes kthvalue to be consistent with sort
(i.e. treat NaN as larger than any number), so that
`a.kthvalue(n) == a.sort()[n - 1]`.

One drawback is that median with a NaN argument does not return NaN,
which is a deviation from NumPy.

Thank you, ngimel, for raising this.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17824

Differential Revision: D14410092

Pulled By: ezyang

fbshipit-source-id: bdec2d8272dc4c65bcf2f9b8995e237774c44c02

Fix minor grammatical mistakes in torch/nn/modules/loss.py (#17892)

Summary:
Fixes some minor grammatical mistakes in the doc of `loss.py`.

I think in the doc:
>  Note that for some losses, there multiple elements per sample.

the "are" is lost between "there" and "multiple".

This mistake takes place in all the descriptions of parameter `size_average` and there are 17 of them.
It's minor but perfects the doc I think. 😁
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17892

Differential Revision: D14418177

Pulled By: ezyang

fbshipit-source-id: 412759f2f9b215819463bf8452ab0e0513218cd6

Remove (almost all) TensorOptions from native_functions.yaml (#17385)

Summary:
Stacked on top of https://github.com/pytorch/pytorch/pull/17386

Brings us to 1014/1106 of writing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17385

Differential Revision: D14248008

Pulled By: cpuhrsch

fbshipit-source-id: 033e00de91e3edf7ae01ca03ebe436c0446b3b5c

Restore full Windows tests (#17102)

Summary:
closes #17101
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17102

Differential Revision: D14420716

Pulled By: ezyang

fbshipit-source-id: 0134736e2d919afa683afa84cb2140f659042643

Prevent VS2017 from emitting ambiguous symbol errors (second time)

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17863

Differential Revision: D14404818

Pulled By: soumith

fbshipit-source-id: 9dac6b926e270e2a29ec2e4dba2e93984da0e5f5

Fix windows test hang (#17778)

Summary:
This PR resolves two concurrent issues discovered when running the test in windows. Details about the windows test can be found here: https://github.com/pytorch/pytorch/issues/17609

The change covers two fixes:
1. update running_preloaders_ upfront before creating worker thread to prevent underflow.
2. add a lock when updating stop_ to prevent dead lock in condition variable cv_write_.

The fix has been tested on both Windows and Linux. With --gtest_repeat=1000, the tests runs smoothly without issues.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17778

Differential Revision: D14404910

Pulled By: soumith

fbshipit-source-id: 2fbb8007e4b0bce4613e9a9fd31b8aace1bbfa8d

torch.btrifact for tensors with greater than 3 dimensions (#14964)

Summary:
Motivation:
- Earlier, `torch.btrifact` could not handle tensors with greater than 3 dimensions. This is because of the check:
>   AT_CHECK(THTensor_(nDimension)(a) == 3, "expected 3D tensor, got size: ", a->sizes());

What is in this PR?:
- Move `btrifact` to ATen
- Remove relation to TH/THC.
- Handle tensors with more than three dimensions
- Tests
- Docs modifications: added a note about the non-pivoting variant.

[blocked due to old magma-cuda binaries]
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14964

Differential Revision: D14405106

Pulled By: soumith

fbshipit-source-id: f051f5d6aaa45f85836a2867176c065733563184

Small clean up of aten_op

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17530

Reviewed By: ezyang

Differential Revision: D14237931

fbshipit-source-id: fb73d63d89fab0622097a49be6ed0b75ddb02a7c

add support for parsing class defs to the string frontend (#17628)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17628

This is not hooked up anywhere yet, just adding support.
This shares the same restrictions as the python frontend—namely, that the only exprs allowed right now are method defs.

Reviewed By: shannonzhu

Differential Revision: D14291654

fbshipit-source-id: 7798e5ff412a52ef8803c7bae8f439e50968a73a

add test for out of order methods (#17624)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17624

Just to make sure this path works

Reviewed By: shannonzhu

Differential Revision: D14288056

fbshipit-source-id: b719c0e90252b6821b1f9b22d3d98982985a6cb3

initializing class value (#17585)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17585

Create a sugared value that represents a class during initialization. This is
so that assignments to attributes correctly define attributes in __init__ but
raise an error elsewhere.

Reviewed By: shannonzhu

Differential Revision: D14263403

fbshipit-source-id: 09b2feeb272302f00a79c2a0302fbdf5483aed6a

Remove unused parameter in ProcessGroupGloo (#17718)

Summary:
This is not used anywhere and wasn't cleaned up prior to 1.0.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17718

Reviewed By: janewangfb

Differential Revision: D14355154

Pulled By: pietern

fbshipit-source-id: f8ff3c8f50cd6365b369a5c5b85d72d8940df048

Revert D14414435: [pytorch][PR] Remove remaining IR Expect files

Differential Revision:
D14414435

Original commit changeset: 0bfd7ce66ac2

fbshipit-source-id: 02de1814f3c4e581d3798059cee752517b176ed9

Remove remaining IR Expect files (#17886)

Summary:
Last batch of IR expect files removed. Includes some removal of expect files that are no longer used.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17886

Differential Revision: D14414435

Pulled By: eellison

fbshipit-source-id: 0bfd7ce66ac2f72a57f15f45ebd60b95e80b6c16

Bool tensor creation (cpu) (#17376)

Summary:
This PR enables bool tensor creation and some basic operations for the CPU backend. This is a part of Bool Tensor feature implementation work. The whole plan looks like this:
    1. Storage Implementation [Done]
    2. Tensor Creation.
        a) CPU (this PR)
        b) CUDA
    3. Tensor Conversions.
    4. Tensor Indexing.
    5. Tensor Operations.
    6. Back compatibility related changes.

**Change**:
Enable CPU tensors and these operations:
- torch.zeros
- torch.tensor
- torch.ones
- torch.randint
- torch.full
- torch.full_like
- torch.empty
- torch.empty_like

**Tested via**:
1) unit tests

2)
torch.zeros(2,2, dtype=torch.bool)
torch.tensor([True, False], dtype=torch.bool)
torch.tensor([-1, -1.1, 0, 1, 1.1, 2], dtype=torch.bool)
torch.ones([1,2], dtype=torch.bool)
torch.randint(10, (2, 2), dtype=torch.bool)
torch.full((2, 3), True, dtype=torch.bool)
torch.empty(4, dtype=torch.bool)

a = torch.tensor([0,0,1])
b = torch.full_like(a, True)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17376

Reviewed By: ezyang

Differential Revision: D14375995

Pulled By: izdeby

fbshipit-source-id: a65490b5360ee0e6e3accc54ce7e32e49ad2d2a8

Remove device guard from TypeDefault::copy()

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17833

Reviewed By: ezyang

Differential Revision: D14400901

Pulled By: li-roy

fbshipit-source-id: ababc95dadfc94a996a80c5332f45f76a300963d

re-enable torch.split tests (#17859)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17859

this has been fixed due to improvements in shape analysis

Reviewed By: driazati

Differential Revision: D14402781

fbshipit-source-id: 4ef2722ffedd9c8ac1eff55c244b421d7d3715ed

Fix lint in test_dataloader.py

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17878

Reviewed By: eellison

Differential Revision: D14409933

fbshipit-source-id: 20ee8953a21e29b4557aff62b5e48dddd630eef6

Optimize fused_dropout_kernel launch bounds for AMD hardware

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17870

Differential Revision: D14409990

Pulled By: ezyang

fbshipit-source-id: 0452282f459770823641b2527f47b1186ab14666

Deprecate torch.pstrf (#17866)

Summary:
Changelog:
- Add deprecation warning to torch.pstrf
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17866

Differential Revision: D14405527

Pulled By: soumith

fbshipit-source-id: 73f3b7d61c60eb57e4bffd08112e552ae3e6dfdc

Allow structseq to be input of operators where tuple is expected (#17208)

Summary:
Currently the following code gives an error on python 2 because `ret` is a structseq which is not a tuple
```python
ret = a.max(dim=0)
ret1 = torch.max(a, dim=0, out=ret)
```

This PR modify tuple check in python arg parser to allow structseq to be input of operators where tuple is expected, which would make the above code work.

Depend on: https://github.com/pytorch/pytorch/pull/17136
Partially fixes: https://github.com/pytorch/pytorch/issues/16813
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17208

Differential Revision: D14280198

Pulled By: VitalyFedyunin

fbshipit-source-id: beffebfd3951c4f5c7c8fe99a5847616a89491f3

Add PyTorch Governance, Contributor Guide, and List of Persons of Interest

Summary: Adding new documents to the PyTorch website to describe how PyTorch is governed, how to contribute to the project, and lists persons of interest.

Reviewed By: orionr

Differential Revision: D14394573

fbshipit-source-id: ad98b807850c51de0b741e3acbbc3c699e97b27f

Fix compilation error (#17860)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17860

att

Reviewed By: bddppq

Differential Revision: D14402751

fbshipit-source-id: 2d53b230dfd775372addeab1d3eaf0b9552fae9f

Revert D14392864: Fix lint in test_dataloader.py

Differential Revision:
D14392864

Original commit changeset: 12477b9cfe29

fbshipit-source-id: 1864a80d5cfaceeae55d0145340a578f978ab4a7

Removed dead code from THTensorMath.h (#17769)

Summary:
This PR removes dead code from THTensorMath.h
I found these unused methods while working on a PR where i plan to move **fill** and **zero** methods from TH/THC to Aten.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17769

Differential Revision: D14372732

Pulled By: izdeby

fbshipit-source-id: 94fd3b52c691ebc89d2bdc8905452e7498038bf5

Introducing array-like sequence methods __contains__ (#17733)

Summary:
for tensor

Fixes: #17000
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17733

Differential Revision: D14401952

Pulled By: soumith

fbshipit-source-id: c841b128c5a1fceda1094323ed4ef1d0cf494909

Revert "Add check for x64 Python before setup (#17707)" (#17864)

Summary:
This reverts commit 08fb9021da32e73bd7dec73104eea6a76dd44439.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17864

Differential Revision: D14404920

Pulled By: soumith

fbshipit-source-id: d41fc06e249f3437d4f80d1d6a5fdbd44c90462b

Registering of kl-divergence for independent distribution (#17681)

Summary:
This address issue https://github.com/pytorch/pytorch/issues/13545 and implements the proposed fix together with a single test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17681

Differential Revision: D14360161

Pulled By: ezyang

fbshipit-source-id: 427afc88e9054b5b0dc39ebbab1087b990695ea5

Fix lint in test_dataloader.py

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17820

Reviewed By: eellison

Differential Revision: D14392864

fbshipit-source-id: 12477b9cfe290428d51cc28e024c8cbe8bb7bf51

Further improvements of nn.container docs

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17731

Differential Revision: D14401894

Pulled By: soumith

fbshipit-source-id: cebb25859f78589cc4f4f8afb1e84c97f82b6962

fix faq typo

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17851

Differential Revision: D14401791

Pulled By: soumith

fbshipit-source-id: ed6d64d6f5985e7ce76dca1e9e376782736b90f9

Fix log_softmax and softmax if any dimension is 0-d (#17651)

Summary:
- Test added
- test_dim_function_empty: softmax and log_softmax on last dimension

fixes: #17262
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17651

Differential Revision: D14349009

Pulled By: gchanan

fbshipit-source-id: b6f728f5c6be8ae7615749e3f0c201886632923e

Correct loss docstrings (#17300)

Summary:
In the loss doc description, replace the deprecated 'reduct' and 'size_average' parameters with the 'reduction' parameter.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17300

Differential Revision: D14195789

Pulled By: soumith

fbshipit-source-id: 625e650ec20f13b2d22153a4a535656cf9c8f0eb

When openblas exists, "OpenBLAS_FOUND" is defined, rather than "OPENBLAS_FOUND". (#17841)

Summary:
See https://github.com/pytorch/pytorch/blob/master/cmake/Modules/FindOpenBLAS.cmake#L36

This typo lead to cmake fails to detect openblas on ubuntu.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17841

Differential Revision: D14400261

Pulled By: soumith

fbshipit-source-id: 287e019e122230cf6b70ab1ea94e5c514f429c88

Passing indices as a list to Subset instead of Tensor (#17649)

Summary:
Indices in Subset were stored as tensors earlier
passing as list in random_split to ensure integer indexing

fixes: #17466
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17649

Differential Revision: D14400250

Pulled By: soumith

fbshipit-source-id: cd20a959f33773c4babf8e861ea37ec61c2713a0

Clarify JIT docs

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17846

Differential Revision: D14400363

Pulled By: jamesr66a

fbshipit-source-id: 862316b5fd95526b6edebeca19d2cc522779df11

Add metadata for torch jit TracedModules. (#17640)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17640

Pull Request resolved: https://github.com/pytorch/pytorch/pull/17311

I've extended our model metadata framework in this diff to support
traced modules as well. Re-used a lot of components from the previous
implementation of ScriptModule metadata.

Tracing is a little different from Scripting since you can't just create a
subclass of TopLevelTraceModule (type returned by torch.jit.trace) and attach
metadata the way we did for ScriptModule. As a result, I've introduced a
separate API torch.fb.jit_trace which returns an instance of
TracedModuleWithMetadata which is a subclass of TopLevelTracedModule. As a
result, we can now attach metadata to this instance.

Reviewed By: dzhulgakov

Differential Revision: D14117966

fbshipit-source-id: 3eee5eef733cb8d6a219c02e2f41d08698eca326

Fix PySlice_Unpack not available on PyPy 3.6 yet (#17836)

Summary:
This is one of the fixes needed to support compilation on PyPy 3.6, see https://github.com/pytorch/pytorch/issues/17835
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17836

Differential Revision: D14399404

Pulled By: soumith

fbshipit-source-id: ca650a6e2066aed86ddd3314a95d0cb3c515c633

PyPy compatibility: let unmodified slots be inherited in the standard way (#17837)

Summary:
This is needed to fix a segfault on PyPy 3.6, see https://bitbucket.org/pypy/pypy/issues/2968/segfault-calling-cpyext_tp_new_tuple and https://github.com/pytorch/pytorch/issues/17835
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17837

Differential Revision: D14399408

Pulled By: soumith

fbshipit-source-id: 75328a30018313d3223dd3e3eef9240a416c049b

Run fp16 resnet50 training in bench script (#17831)

Summary:
cc xw285cornell
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17831

Differential Revision: D14398532

Pulled By: bddppq

fbshipit-source-id: 37c03cc2eebe3a6083e05631cb6ff03474e4a8a2

Int8 FC performance debugging (#17700)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17700

Add performance debugging utilities in DNNLOWP FC operator and the python script

Reviewed By: amylittleyang

Differential Revision: D14321299

fbshipit-source-id: 50dbd7b352a1da5d2ecb659d8003e71e70750063

Optimize LayerNormOp (#17604)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17604

Optimize LayerNormOp

i-am-not-moving-c2-to-c10

Reviewed By: houseroad

Differential Revision: D14274175

fbshipit-source-id: a7aa263a1b0eb109682d2be99306e7b2cdcc0faf

Remove some simple use cases of Type::ScalarType()

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17529

Reviewed By: ezyang

Differential Revision: D14237932

fbshipit-source-id: be633a1fc19215d53cfe083fdd7196acf2b7dd2f

Change Dispatch.h to use ScalarType over Type

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17527

Reviewed By: zou3519

Differential Revision: D14235395

fbshipit-source-id: 3f53e33f6794f1f14c2edf79014b8ef8397822c5

Revert D14361993: [pytorch][PR] [Onnx] - refactoring serialization of ONNX initializers to be name-based

Differential Revision:
D14361993

Original commit changeset: da93e945d557

fbshipit-source-id: 15eea001fbcd059ac13903405aeb9ea182c6ee8b

Open registration for c10 thread pool (#17788)

Summary:
1. Move ATen threadpool & open registration mechanism to C10
2. Move the `global_work_queue` to use this open registration mechanism, to allow users to substitute in their own
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17788

Reviewed By: zdevito

Differential Revision: D14379707

Pulled By: jamesr66a

fbshipit-source-id: 949662d0024875abf09907d97db927f160c54d45

Cast nn.Upsample.scale_factor to a float (#17732)

Summary:
Fixes #17106
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17732

Differential Revision: D14388192

Pulled By: driazati

fbshipit-source-id: d9c9e87a7c6db63c1de3ddebbb8dcf619f0dc34d

Fix lint in run_test.py

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17815

Reviewed By: eellison

Differential Revision: D14390308

fbshipit-source-id: 22efd62a1bbd1fc8155a942d7160d5b7d3158e6b

Fix lint in test/common_utils.py

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17814

Reviewed By: eellison

Differential Revision: D14390194

fbshipit-source-id: b4b3bbe20a15d0b9ed127b255e01c0d6d0832c1b

Replace tensor.type().scalarType() calls with tensor.scalar_type()

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17515

Reviewed By: ezyang

Differential Revision: D14233250

fbshipit-source-id: 6c7af8d2291c0c2b148001b30cf03834f34366c0

Catch exceptions in bound_shape_inference (#17775)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17775

Handles use input shape hint properly.

Reviewed By: zrphercule

Differential Revision: D14368735

fbshipit-source-id: 504cd96589e47aa432617e56362aa6b01a25ba9b

refactor caffe2 operator constructors - 11/9 (#17722)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17722

clangr codemod

Reviewed By: ezyang

Differential Revision: D14350584

fbshipit-source-id: adef54cedc9409b4fb365f6644e2621a9e47b2ff

Suppress C408 lint (don't use dict constructor) (#17813)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17813

We have a lot of manually written out dict() constructors,
and (1) I don't think use of curly brace syntax is much
of an improvement and (2) it seems like a waste of time to
fix them all.

Reviewed By: eellison

Differential Revision: D14390136

fbshipit-source-id: 6199bef4dea75b6079bcb9d9e8acf20a2e1a86e1

Add matches_jit_signature to recent native functions

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17805

Differential Revision: D14388004

Pulled By: cpuhrsch

fbshipit-source-id: c50580b6fe1e9cfefed91aaa526376325d9f9c0d

Add /MD to prevent linking errors on Windows

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17799

Differential Revision: D14385777

Pulled By: ezyang

fbshipit-source-id: 8c1d9f80c48399087f5fae4474690e6d80d740e6

Change message on unknown db type to be friendly (#17795)

Summary:
CreateDB actually returns nullptr when db type is unknown and throws when the file is missing
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17795

Reviewed By: ezyang

Differential Revision: D14383226

Pulled By: dzhulgakov

fbshipit-source-id: 1dcf75a6b4ba8b64a24d4e5daf02db3189d56b7b

Trace rnn max_batch_size (#17727)

Summary:
This causes the tracer to record the select / cast to int operation instead of just an int constant

Fixes #15319 but relies on a fix for #17583 first
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17727

Differential Revision: D14377886

Pulled By: driazati

fbshipit-source-id: 59453def54ba72756303f723993844dbeb5d2f8b

Remove legacy way of exposing caffe2 operators to PyTorch (#17742)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17742

This path isn't used anymore, and is incompatible with the changes stacked on top of this diff.
Removing it.
cc bwasti to check and confirm these can really be deleted

Reviewed By: ezyang

Differential Revision: D14362426

fbshipit-source-id: 32cdc19f28c2a981ae1e204901420998367ee588

Remove 'Tensor' key from ATen codegen. (#17782)

Summary:
We used to have different ATen Tensor types, but we don't anymore.  This was just being maintained by a codegen'ed comment.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17782

Reviewed By: ezyang

Differential Revision: D14378004

Pulled By: gchanan

fbshipit-source-id: 1bbf276393a391252d372cc385230c784bd78588

Remove ProcessorSpecificPlugin. (#17789)

Summary:
It doesn't seem to be used.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17789

Reviewed By: ezyang

Differential Revision: D14382423

Pulled By: gchanan

fbshipit-source-id: 0ac3236c48979a1b2bcd615e307e55f10fd8eb77

Remove THPPlugin. (#17790)

Summary:
It doesn't seem to be used.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17790

Reviewed By: ezyang

Differential Revision: D14380897

Pulled By: gchanan

fbshipit-source-id: 3c3884a08c3b6c1489347d439509b19e079c5861

Replace tens with hundreds.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17752

Differential Revision: D14366743

fbshipit-source-id: 39f6ac08180d780866e284024918d9abd197d239

Support failback for more operators in ideep (#17747)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17747

RMACRegions, Normalize and RoIPooling

Reviewed By: dskhudia

Differential Revision: D14365096

fbshipit-source-id: dafcb7077515e03c2880832a442015b70fc7140d

Cleanup include files in jit/passes/common_subexpression_elimination.h.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17784

Differential Revision: D14381529

Pulled By: ZolotukhinM

fbshipit-source-id: e32e17ee644ef888a6d56a8ee3648e7ac21758bf

Use return names in JIT operators

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17638

Differential Revision: D14295606

Pulled By: cpuhrsch

fbshipit-source-id: 62040ac65434411357808735f0fe6cd33cc1c30f

Change ConvPoolOp<Context>::SetOutputSize to ConvPoolOp<Context>::GetOutputSize (#17764)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17764

Original commit changeset: f1923fdca4a1

reverted int8 ops fixes the original runtime regression.
We'll ignore the memory regression since it is flaky, see D14228484

Reviewed By: dzhulgakov

Differential Revision: D13885233

fbshipit-source-id: ccbe4b94acb44b7b4cb3ae4d73e3f6091e1e1195

Clean up some old ScalarType stuff

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17755

Differential Revision: D14377135

Pulled By: li-roy

fbshipit-source-id: 35305760a1621340ba66c61a193ff61cfedfa7e8

add reference to flake8-mypy in contributing.md

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17759

Differential Revision: D14376813

Pulled By: eellison

fbshipit-source-id: cca1128e967ef7368633b94a3fa3c8e76a4a16f4

Move lerp to ATen, add functionality for tensor weights (#17348)

Summary:
Changelog:
- Remove TH/THC bindings
- Add tensor weights for `lerp`
- Modify derivatives appropriately
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17348

Differential Revision: D14355845

Pulled By: soumith

fbshipit-source-id: eaede4c09ee589d77ba6cf52583510ea8e3a2fcf

Refactor dispatcher (#17753)

Summary:
This is a side PR for a bool tensor feature. The idea of this change came from a feedback received in this [PR](https://github.com/pytorch/pytorch/pull/17376).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17753

Differential Revision: D14367989

Pulled By: izdeby

fbshipit-source-id: 4fa380e56e20f18e480be68920170dbc3a4eb91c

add layernorm to AD

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17702

Differential Revision: D14368472

Pulled By: wanchaol

fbshipit-source-id: 8db390e39444078258ad1d34ba74d6ddafa5d02b

move half<->float conversions to oss operators (#17548)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17548

expose half float operators to OSS

common/math/Float16.h is the original implementation
this is substituted by caffe2/c10/util/Half.h

from the comments seems like the both implementations don't handle denormals

Reviewed By: jspark1105

Differential Revision: D14244200

fbshipit-source-id: f90ba28c5bf6a2b451b429cc4925b8cc376ac651

Fix the update ONNX expect files (#17767)

Summary:
Fix the CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17767

Reviewed By: zrphercule

Differential Revision: D14370483

Pulled By: houseroad

fbshipit-source-id: e7b0bbde0797c41f5a010fa206fab80fe2792eb7

Cleanup testFusion/testOne: there are unused arguments.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17737

Differential Revision: D14366584

Pulled By: ZolotukhinM

fbshipit-source-id: 3c2dd2aabfecca475909e4eec4a077d900795da9

update of fbcode/onnx to 96c58ceeacf0f2b73d752e413e4fd78787a12da3 (#17676)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17676

Previous import was e18bb41d255a23daf368ffd62a2645db55db4c72

Included changes:
- **[96c58ce](https://github.com/onnx/onnx/commit/96c58ce)**: Fix shape inference when auto_pad is notset again (#1830) <Li-Wen Chang>
- **[873ddbb](https://github.com/onnx/onnx/commit/873ddbb)**: More extendable Runner (#1809) <Michał Karzyński>

Reviewed By: zrphercule

Differential Revision: D14321241

fbshipit-source-id: 12de9021afc61f5435f1b719cccf7b0f4ad73a84

Set the default ONNX opset to the latest stable opset (i.e., 9) (#17736)

Summary:
1) The changes in the new opset won't affect internal pipeline.
2) The CI won't be affected by the ONNX changes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17736

Reviewed By: zrphercule

Differential Revision: D14358710

Pulled By: houseroad

fbshipit-source-id: 4ef15d2246b50f6875ee215ce37ecf92d555ca6a

Add module attributes (#17309)

Summary:
Similar to `nn.Parameter`s, this PR lets you store any `IValue` on a module as an attribute on a `ScriptModule` (only from the Python front-end currently). To mark something as an attribute, it should wrapped in `jit.Attribute(value, type)` (ex. `self.table = torch.jit.Attribute(table, Dict[str, torch.Tensor])`)

Followup Work:
* (de)serializing for use in C++
* change `self.training` to be a `bool` attribute instead of a buffer
* mutable attributes
* string frontend support
* documentation
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17309

Differential Revision: D14354316

Pulled By: driazati

fbshipit-source-id: 67e08ab5229366b67fbc837e67b58831a4fb3318

- refactoring serialization of ONNX initializers to be name-based (#17420)

Summary:
Currently, serialization of model parameters in ONNX export depends on the order in which they are stored in a container (`list` on Python side and `std::vector` on C++ side). This has worked fine till now, but if we need to do any pass on that graph that mutates the parameter list, then strictly order-based serialization may not work.

This PR is the first in a set to bring in more passes (such as constant folding) related to ONNX export. This PR lays the groundwork by moving the serialization in ONNX export from order-based to name based approach, which is more amenable to some of the passes.

houseroad - As discussed this change uses a map for export, and removes the code from `export.cpp` that relies on the order to compute initializer names.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17420

Differential Revision: D14361993

Pulled By: houseroad

fbshipit-source-id: da93e945d55755c126de06641f35df87d1648cc4

ONNX Export for Max and Average Pooling in CEIL_MODE

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16769

Differential Revision: D14362175

Pulled By: houseroad

fbshipit-source-id: 65cfb1dfba6a43d39cc85374add368fe8e4e5645

use flake8-mypy (#17721)

Summary:
Use flake8 installed with mypy checks so that our linter matches fbcode. Mypy type errors also provide valuable signal
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17721

Differential Revision: D14357778

Pulled By: eellison

fbshipit-source-id: d8c9ea3fe3b5f550c3b70fe259e0eabf95e4c92d

use fp16<->fp32 intrinsic (#17496)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17496

As title.

Reviewed By: hyuen

Differential Revision: D14222907

fbshipit-source-id: d5d6c032e725ca8b52aca2be7401ec3c59f6a242

Implement a Caffe2 standalone LSTM operator (#17726)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17726

Pull Request resolved: https://github.com/pytorch/pytorch/pull/17725

Pull Request resolved: https://github.com/pytorch/pytorch/pull/17461

Implementing a standalone LSTM Operator in Caffe2 adopted from this Aten implementation: diffusion/FBS/browse/master/fbcode/caffe2/aten/src/ATen/native/RNN.cpp. The most tricky thing in this exercise was that caffe2::Tensor has no copy constructor that made it necessary to implement a custom templated copy constructor for the different Tensor containers used in the code. Also there was no way to use off-the-shelf C2 operators in my code easily so I had to copy some code that is doing basic matmul, cat, split, transpose and linear as utility functions.

Two things missing:

- Profiling this implementation against the current ONNXified LSTM op
- Make this operator available to use in PyTorch

Reviewed By: dzhulgakov

Differential Revision: D14351575

fbshipit-source-id: 3b99b53212cf593c7a49e45580b5a07b90809e64

caffe2:libtorch_cuda depends on caffe2:caffe2_gpu (#17729)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17729

When doing "import torch" in fbcode, previously the caffe2 cuda kernels weren't loaded because libcaffe2_gpu.so wasn't loaded.
Once you also did "from caffe2.python import workspace", then the cuda kernels were loaded because that triggered a runtime mechanism for loading libcaffe2_gpu.so.

We want the cuda kernels to always be available, so this diff adds a dependency from caffe2:libtorch_cuda to caffe2:caffe2_gpu.

Reviewed By: ezyang

Differential Revision: D14353498

fbshipit-source-id: 76a9fe69f231b308ab40eac393bb216c6fad3658

add tensor and cost inference functions (#17684)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17684

Adding tensor and cost inference functions to more int8 operators.

Reviewed By: yinghai

Differential Revision: D14174746

fbshipit-source-id: dfad975fa75899565c8fb61f1b7747a9206ebd22

ONNX Export Narrow op

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17550

Differential Revision: D14350401

Pulled By: houseroad

fbshipit-source-id: 4d88079bb7a8bbd270b0272009826eb3b202cc33

Keep the dim_type of hinted shape as BATCH if possible (#17734)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17734

If input is not BATCH, we will skip adjust its batch size during onnxifi transformation. So when we take hints, we take it as CONSTANT but later need to change it to BATCH if possible.

Reviewed By: jackm321

Differential Revision: D14355983

fbshipit-source-id: 63eb54a44afb1565c71486fdd73db07ca0ac4fd4

fix different round behavior on CPU and GPU #16498 (#17443)

Summary:
xxtemp, colesbury, bhushan23, zou3519,  convert gpu round behavior to half-to-even, consistent with torch cpu version and numpy. You feedback are welcomed.
See #16498
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17443

Differential Revision: D14261786

Pulled By: VitalyFedyunin

fbshipit-source-id: 98156436b545d72769831a89e2775d43ad913ebc

Warn about memory overlaps on expanded tensors (#17576)

Summary:
Eventually we should remove these when we're certain that all our ops
handle memory overlaps correctly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17576

Differential Revision: D14349990

Pulled By: zou3519

fbshipit-source-id: c3a09f6113b9b1bf93e7f13c0b426c45b2cdf21f

fix exp fam. formula

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17719

Differential Revision: D14349029

Pulled By: soumith

fbshipit-source-id: cf016756a9319436f7379e8377f8bd1e1b672b40

refactor caffe2 operator constructors - 10/9 (#17659)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17659

clangr codemod

Reviewed By: ezyang

Differential Revision: D14304675

fbshipit-source-id: 45fbd84c50651a70ae29bf46df3322715e99d225

Improve ONNX symbolic for logsoftmax and softmax (#17672)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17672

support dtype in the onnx symbolic

Reviewed By: zrphercule

Differential Revision: D14313987

fbshipit-source-id: e9364621b3f795191d880599711dfbcb220d0e31

Enable using CMD when building cpp extensions on Windows

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17706

Differential Revision: D14346482

Pulled By: ezyang

fbshipit-source-id: 7c85e51c701f6c0947ad324ef19fafda40ae1cb9

Do not rename net boundary inputs/outputs during ssaRewrite. (#17545)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17545

This diff avoids renaming boundary inputs of net during onnxifi transform.
It also removes adding mappings for the initializer during onnxifi op creation.
Thus gets read of the mapped ws creation during onnxifi op creation.

Reviewed By: zrphercule

Differential Revision: D14243161

fbshipit-source-id: 6eafa920c45f6a6bfacbbb443e8e84cf9778644c

Reapply D14078519 (#17596)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17596

Was reverted before, now fixed version.

Reviewed By: ezyang

Differential Revision: D14270288

fbshipit-source-id: c72490b5d02cc6098cb60145fa9a842b3c9a24c5

Batch of expect file removals (#17581)

Summary:
Another batch of removing expect files.

One note - I removed the Batched expect files without adding equivalent tests since they are already being tested in another ways, and we are no longer actively maintaining that project.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17581

Differential Revision: D14343578

Pulled By: eellison

fbshipit-source-id: ce0b1fd2b5b4ec80ad9003bab1b58f41645d3da6

(#14267)

Summary:
- Summary:

Added synchronized batch normalization, allows synchronization of stats across mini-batches between processes within a process group.
Current implementation uses a mixture of extended ATen native functions (cpp cuda extension) + torch.nn.modules (c10d python API)

- User-facing api:

1. torch.nn.utils.convert_sync_batchnorm(modules, process_group=None)

2. torch.nn.SyncBatchNorm(num_features, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True, ***process_group=None***)

- supported use case:
DistributedDataParallel with ***single-gpu multi-process***

a. User creates model containing `torch.nn.SyncBatchNorm` layers through one of the ways listed below:

  1. use layers directly:

     torch.nn.SyncBatchNorm(...)

     similar API as with torch.nn.BatchNormXd(...)
     with added argument `process_group` which is used to limit the scope of
     synchronization within each process group. Default value is None, which
     implies synchronization across all GPUs

  2. use torch.nn.utils.convert_sync_batchnorm(modules, process_group)

     recursively convert all `torch.nn.BatchNormXd` into `torch.nn.SyncBatchNorm`
     preserving values of parameters/buffers.
     the utility function also allows user to specify process_group value to all
     converted layers.

b. user wraps their model with
   `torch.distributed.parallel.DataParallelDistributed`, from this point, user
   should follow the general guidelines for DDP use guide

- Error checking

For use cases not supported, we error out:

1. Application launched without ddp:
   > import torch
   > sbn = torch.nn.SyncBatchNorm(10).cuda()
   > inp = torch.randn(5, 10, 3, 3).cuda()
   > sbn(inp) --> Error!
   > AttributeError: SyncBatchNorm is only supported within torch.nn.parallel.DistributedDataParallel

2. Application launched using DDP with multi-GPU per-process:
   > ddp_module = nn.parallel.DistributedDataParallel(module, device_ids=device_ids, output_device=args.local_rank)
   > ValueError: SyncBatchNorm is only supported for DDP with single GPU per process
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14267

Differential Revision: D14270035

Pulled By: ezyang

fbshipit-source-id: 4956d8fa565c32e9df5408d53719ff9f945f4d6d

Update ModuleDict doc about order

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17717

Differential Revision: D14346557

Pulled By: ezyang

fbshipit-source-id: 2484c7d8105f9aa8bce5567d1fa2d4f587cc9cc2

Update CODEOWNERS (#17720)

Summary:
teng-li is passing the baton to mrshenli. Thanks for all your work on distributed teng-li!! :tada:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17720

Differential Revision: D14350120

Pulled By: pietern

fbshipit-source-id: edfe784520c54630203cc8fbb296455d3dbf341b