PyTorch 1.10 dev release notes

release/packaging
1

We have quite a few commits in the 1.10 release and some things that are interesting for people that develop within PyTorch.
You can find below a curated list of these changes:

Developers

Python API

  • Generic test parametrization functionality (#60753)
  • Ensure NativeFunctions.h codegen output is deterministic (#58889)
  • hide top-level test functions from pytest’s traceback (#58915)
  • remove pytest.UsageError (#58916)
  • introduce TestingErrorMeta for internal use (#58917)
  • make assert_equal an example how to partial torch.testing.assert_close (#58918)
  • split TestAsserts by functionality (#58919)
  • add support for constant (#60166)
  • ensure xml report path are relative to */pytorch/test (#60380)
  • Set TORCH_WARN_ONCE to always warn inside of assertNotWarn (#60020)
  • Print stdout and stderr to console on parallel runs (#60869)
  • Fix several test_ops cuda dtypes tests (#60922)
  • remove UsageError in favor of ValueError (#61031)
  • Remove test_out, test_variant_consistency_eager skips for addmv; fixed before (#61579)
  • [OpInfo] Added ReductionOpInfo subclass of OpInfo and ported sum test (#62737)
  • Added API tests to ReductionOpInfo and ported amax/amin/nansum tests (#62899)
  • Revert "Revert D30558877: Ported std/var to ReductionOpInfo (#65262)
  • Cleanup functional.py after lu_unpack was removed (#58669)
  • Fix torch.finfo.bits typo in stub (#58819)
  • masked_scatter thrust->cub (#58865)
  • improve broadcasts_input error message (#58295)
  • OpInfo
  • ATen migration: nonzero , renorm, torch.lstsq , crossKernel, glu, nll_loss_forward, (#59149, #59250, #59400, #60039, #61153, #60097)
  • Move sharding to after all tests have been excluded (#59583)
  • Fix Python 3.8 expecttest (#60044)
  • Improve error messages of torch.testing.assert_close in case of mismatching values (#60091)
  • update docstring examples of torch.testing.assert_close (#60163)
  • tests for diagnostics in callable msg in torch.testing.assert_close (#60254)
  • Increase tolerance for test_grad_scaling_clipping (#60458)
  • Add a test case for findDanglingImpls (#61104)
  • Don’t check stride by default in Torch testing (#60637)
  • Fix ignoring Tensor properties in torch.overrides (#60050)
  • mvlgamma: int → float type promotion (#59934)
  • Use python3.6 compatible APIs in clang_tidy.py (#60659)
  • correct filename issue for test_cpp_extensions_aot (#60604)
  • First step to rearrange files in tools folder (#60473)
  • Add exclusion list to _check_kernel_launches.py (#60562)
  • use explicitly non-returning GPU atomics (#60607)
  • Test parametrization for instantiated device-specific tests (#60233)
  • Use maximum of tolerances in case of mismatching dtypes (#60636)
  • remove randn? from torch.testing namespace (#61840)
  • Parity tests for functional optimizer step_param (#61756)
  • Functional Optim: Test kwargs parity for SGD (#62078)
  • Make SkipInfo with a unittest skip(#63481)
  • Modify compare scalars testing error message when atol=0 and rtol=0 (#60897)
  • Added reference tests to ReductionOpInfo (#64273)
  • Remove opinfo warning from floor_divide (#58682)
  • Support torch.concat alias, add cat OpInfo & remove OpInfo test_out skips {cat, stack, hstack, vtack, dstack} (#62560)
  • Remove run_functional_checks from test_autograd and create necessary OpInfos (#64993)
  • Fixing user inputs for low, high in make_tensor (#61108)
  • kill SkipInfo (#64878)
  • Add skipIfTBB decorator (#64942)

Complex Number

  • Skip conjugate and negate fallback for view ops and their in-place versions (#64392).
  • Remove component wise comparison of complex values in TestCase.assertEqual (#63572).

Autograd

  • Added autograd not implemented boxed fallback (#63458)
  • Added support for using gradients named for outputs in derivatives.yaml (#63947)
  • Forward AD now being tested using OpInfo (#58304, #60498)
  • Added change to extract TestAutogradComplex into its own test file (#63400)
  • No longer skip gradgrad checks because fast gradcheck made it faster (#60435)
  • Improve debug-only autograd checks to ensure ops are properly view and inplace as advertised on their schema (#60286)

torch.nn

  • Testing
    • Added OpInfo entries for nn.functional.{avg_pool2d, conv2d, conv_transpose2d, cosine_similarity, grid_sample, interpolate, layer_norm, linear, mse_loss, one_hot, relu, softmax, softplus, unfold} (#62455, #65233, #62882, #62959, #62311, #61956, #63276, #61971, #62254, #62253, #62076, #62077, #62317, #62705)
    • Added ModuleInfo-based testing for forward reference function comparison, module string representation, pickling, in-place vs. out-of-place comparisons, and factory kwarg usage (#61935, #63737, #63736, #63739, #62340, #62999)
    • nn.AvgPool2d: Empty caching allocator before large subtest in test_avg_pool2d (#63528)
    • nn.Conv2d: Disabled test test_Conv2d_groups_nobias for ROCm (#59158)
    • Tweaked TF32 tolerance thresholds for nn.{Conv2d, Conv3d} tests (#60209, #60451)
    • nn.LayerNorm: Restored deleted numerics test (#64385)
    • nn.LayerNorm: Set behavior to always use fast gradcheck for test variant 3d_no_affine_large_feature (#61848)
    • Remove using directive from header to fix Sandcastle tests (#59728)
    • Fixed logic to correctly avoid TF32 operations for tests that explicitly disable TF32 (#59624)
  • Migrated {multilabel_margin_loss, nll_loss2d, nll_loss2d_backward, rrelu_with_noise, thnn_conv2d, thnn_conv_depthwise2d} from TH / THC to ATen (#60708, #62826, #60299, #57864, #62006, #63428, #62281)
  • Removed dispatch from parallel regions for nn.{BatchNorm*d, Embedding, CTCLoss} (#60596, #60597, #60599)
  • Back out “[pytorch][PR] Adds dtype to nn.functional.one_hot” (#59080)
  • Removed unused variables (D29994983, D30000830)
  • nn.Module: Added option to pass module instance to _load_state_dict_pre_hooks (#62070)
  • nn.EmbeddingBag: Replace thrust with cub for sorting (#64498)

Dataloader

  • Fixed test_fd_limit_exceeded to use fork explicitly as multiprocessing method for different default methods across Python versions on MacOS (#60868)
  • Fixed test_ind_worker_queue to set max_num_worker based on system resource (#63779)
  • Fixed test_multiprocessing_contexts that jetson device doesn’t support sharding cuda Tensor (#64757)
  • Fixed TestSetAffinity to use CPU ID based on hardware (#65042)
  • Added test to validate the generator attached to Sampler is created lazily (#63646)

AMD

  • Updated ROCm PyTorch persons of interest (#55206)
  • Make Jeff and Jithun .circleci/docker code owners (#60958)
  • Skipped test_masked_scatter_large_tensor_cuda (#61313)
  • Fixed bug in #60313 (#61073)
  • Updated CI images for ROCm 4.3.1 (#64610)
  • Added ROCm as a platform for which tests can be disabled (#63813)
  • Added a skip for ROCm for a flaky test (#62664)
  • Added distributed/_sharded_tensor/test_sharded_tensor to ROCM_BLOCKLIST (#63508)
  • Disabled test test_Conv2d_groups_nobias_v2 for ROCm (#58701)
  • Disable CircleCI ROCm build (#64434)
  • [ROCm] allow user to override PYTORCH_ROCM_ARCH (#60602)
  • [ROCM] enable fft tests (#60313)

CUDA

  • Added change to wrap cudaStreamSynchronize calls (#61889)
  • Fix formatting repeat_interleave op files (#58313)
  • Added kernel launch checks after each kernel launch to silence the check (#58432)
  • Added change to move code to Blas.cpp, clean up THC magma (#58526)
  • Introduced step 0 of cuDNN v8 convolution API integration (#51390)
  • Added #pragma once to CUDA foreach headers (#58209)
  • Removed THCReduce.cuh (#59431)
  • Fixed const correctness and loop variable type in CUDACachingAllocator (#59819)
  • Fixed bad change in a CUDACachingAllocator loop (#59903)
  • Fixed compile failure on CUDA92 (#60017)
  • Added a follow-up fix for compilation error on CUDA92 (#60287)
  • Fixed version comparison for defining CUDA11OrLater (#60010)
  • Moved remaining Sort in THC to ATen (#58953)
  • Fixed kernel launch check in cross kernel (7bf195f360)
  • Removed some unnecessary functions from CUDAHooks (#59655)
  • Added change to use make_unique instead of std::unique_ptr in CudaCachingAllocator (#61272)
  • Added change to split zeta_kernel out of BinaryMiscOpsKernel.cu (#62261)
  • Added cast return of cudaGetLastError() to void when discarding (#62518)
  • Replaced hardcoded values in IndexKernel.cu (#63372)
  • Killed THCUNN (#63429)
  • Migrated legacy lstsq from THC to ATen (CUDA) (#63504)
  • Migrated THCTensor_copyIgnoringOverlaps to ATen (#63505)
  • Added change to convert mul to use opmath_gpu_kernel_with_scalars (#64019)
  • Added acc_gpu_kernel_with_scalars and port add to use it (#63884)
  • CUDA graphs: added hotfix for test_graph_ (#64339)
  • Migrated uses of THCReduceApplyUtils to cuda_utils::BlockReduce (#64713)
  • Removed 9.2 related macros for CONSTEXPR (#65066)
  • Removed CUDA 9.2 references conditionals and workarounds (#65070)
  • Makes a streaming backward test try gradient stealing more directly (#60065)

C++ API

  • Made it easy to grep out variant of unary/binary op kernels in codebase (#60128)
  • Enhanced comparison tests for c10::optional (#62887)

TorchScript

  • Removed metadata.pkl file from PyTorch model (#61760)

torch.package

  • Added a hack that allows typing.io/re use case (#60666)
  • Added change to wrap torch::deploy API functions in safe rethrow macros (#58412)
  • Added proper handling of re-packaging mocked modules (#61434)
  • Added support in Digraph to track predecessors (#61146)
  • Added register_module_source, a helper function to get some python source code loaded on each interpreter (#58290)

Distributed

DistributedDataParallel

  • Retired _module_copies field in DDP reducer. (#59094)
  • Merged work and future_work in reducer (#59574)
  • Removed legacy code from DDP internals (#58592, #58595, #58603)
  • Removed unneeded code around DDP buffer synchronization logic (#64777, #64473)

torch.distributed

  • Added a split for custom class bindings out of python binding code (#58992)
  • Added the use of irange in a few places (#55325)
  • Corrected launcher tests (#59635)
  • Added the change to use TORCH_CHECK instead of std::runtime_error (#59667, #59683,#59918, #63928)
  • Moved c10d to libtorch(_cuda) (#59561, #59562, #59696, #59697, #59563)
  • Enabled ncclAvg for reductions (#62303)
  • Made JIT operation call to accept Stack& instead of Stack* (#63414)
  • Cleaned up autograd/distributed autograd engine thread state (#63115)

torch.distributed.rpc

  • Introduced new rpc.barrier API (#53423)
  • Added some TORCH_API annotations to RPC (#60169)
  • Moved RPC agents to libtorch (#60170)
  • Moved torch/lib/c10d to torch/csrc/distributed/c10d (#60543)
  • Added RPC Parameter Server Benchmark (#58003, #60784, #60785)
  • Added helpers to manipulate futures (#57846)
  • Made RequestCallback collect Futures from methods, rather than providing them (#57847)
  • Added change to centralize setting messageId in RequestCallback (#57848)
  • Made some methods of RequestCallback return void instead of bool (#57849)
  • Unified invoking JIT operands (#57850)
  • Unified invoking JIT functions (#57851)
  • Unified async execution for JIT functions (#57852)
  • Added change to simplify OwnerRRef completion (#57853)
  • Removed getScriptRemoteCallType (#57854)
  • Unified assignment of OwnerRRef result (#57856)
  • Added change to simplify process(Script|Python)(Remote)?Call (#57857)
  • Deduplicated Python object serialization (#57858)
  • Unified fetching RRefs (#57859)
  • Made remaining RRef methods return futures (#57860)
  • Made remaining autograd methods return futures (#57861)
  • Added change to always use intrusive_ptr for Message (1 out of 2) (#59205)
  • Added change to always use intrusive_ptr for Message (2 out of 2) (#59206)
  • Provided pre-extracted DataPtrs when completing a Future with a Message (#59208)
  • Created CUDA-aware futures in RequestCallback (#59209)
  • Removed LazyStreamContext (#59298, #59299)
  • Allowed Future::then to return pre-extracted DataPtrs (#59207)
  • Merged TensorPipe’s CPU and CUDA channel registry (#59375)
  • Made CUDA serde support for TensorPipe agent pluggable (#59376)
  • Moved CUDA-related stuff of TensorPipe agent to separate file (#59377)
  • Prepared for TensorPipe separating its CUDA-specific headers (#59788)
  • Added change to reduce overhead when Future invokes callbacks inline (#57638)
  • Added change to prevent using anything other than intrusive_ptr for Future (#58421)
  • Allowed Future::then to return pre-extracted DataPtrs (#58424)

torch.distributed.optim.ZeroRedundancyOptimizer

  • Cleaned up ZeRO (#60285)
  • Fixed ZeRO sort to be by numel (#60556)
  • Refactored non-joined process computation (#61555)
  • Made broadcast_object_list accept a device parameter (#61305)
  • Made _Join, _Joinable, _JoinHook public (#62605)
  • Refactored commonalities between two approaches (#62624)
  • Add change to gate DistributedOptimizers on RPC availability (#62937)

torch.fx

  • Introduced prototype for guarding against mutable operations in tracing (#64295)

Benchmark

  • PyTorch core should not require NumPy: torch.utils.benchmark (#60564)
  • PyTorch core should not require NumPy: Operator Microbenchmarks (#64707)

Performance_as_a_product

  • OpenMP: Refactored parallel_reduce to share code with parallel_for (#60184)
  • Removed many unnecessary constructor calls of Vectorized (#58875)
  • Fixed arange functions for VSX specializations of Vec256 (#58553)

Composability

  • Added query() and synchronize() to c10::Stream, increasing parity with at::cuda::CUDAStream (#59560)
  • Added change to make PyObject_FastGetAttrString accept const char* (#59758)
  • Added change to tag PyObject on TensorImpl per torchdeploy interpreter (#57985)
  • Added ExclusivelyOwned and MaybeOwned, which can be used to help reduce the number of tensor refcount bumps in some parts of the codebase (#59419, #63450)
  • Improved CI testing parity across pytorch/pytorch and pytorch/xla (#59888, #59989)
  • Added a boxed CPU fallback kernel (#58065)
  • Added an optional Device parameter to pin_memory/is_pinned that does nothing (#60201)
  • Fixed Dispatching not considering List[Optional[Tensor]] for dispatch (#60787)
  • Added support for registering boxed functors to the dispatcher (#62658)
  • Structured kernels**:** added the ability to precompute values in the meta function that can be reused later in the impl function (#61746)
  • Improved build times by factoring out core parts of the tensor class from TensorBody.h (which is used as an input to codegen) to TensorBase.h (which is not) (#63612)
  • Added basic support for meta tensor serialization. (#62192)
  • Added a new per-operator C++ API that can be accessed through (with add.Tensor as an example) ATEN_FN2(add, Tensor). This API exposes a struct that contains some compile-time constant info about each operator, such as its name, overload name, and schema. (#60214)

Build_Frontend

  • CUDA-9.2 builds or older are no longer supported (#65065, #61462, #65024)
  • Update CMake minimum version to 3.10 (#63660)

Foreach_Frontend

  • Foreach Binary Test Refactor now supports complex dtypes via binary ops with one tensorlist and one scalarlist (#59907)
  • Foreach Test Refactor: Pointwise, Min/Maximum (#61327)

Sparse_Frontend

  • Setup CI to run test_sparse_csr tests (#58666)
  • Remove some uses of deprecated Tensor methods (#58990)
  • Improved test for unsupported backwards for sparse COO Tensor (#59971)
  • Increased dtype test coverage for sparse CSR tensors (#60656)
  • Improved the representative power of random sparse CSR tensors used by tests (#60283)
  • Removed dispatch in parallel regions for sparse COO CPU kernels (#60598)
  • Fixed variable initialization issues in C++ code (#60896)
  • Improved error messages of torch.testing.assert_close for sparse inputs (#61583)
  • Moved some .cu files that do not rely on CUDA to .cpp (#63894)
  • Add support for sparse Tensors to torch.testing.assert_close (#58844)
1 Like
2

And here are the updates to the documentation!

Documentation

Python API

  • Updated torch.quasirandom.SobolEngine docstring w/ correct behavior (#62548)
  • Added docstring for torch.is_inference_mode_enabled (#59047)
  • Updated document dim argument to tensor.size (#58777)
  • Fixed name of dims kwarg in torch.tile docs (#59471)
  • Added doc for torch.einsum sublist format (#57038)
  • Fixed torch.fft documentation examples and added runs of doctests in the test suite (#60304)
  • Improved documentation for torch.set_rng_state (#60422)
  • torch.index_{put, add, copy, fill}: Removed unused sample input argument from functions (#60486)
  • Fixed incorrect doc about the dtype for torch.randint described in issue #56347 (#60507)
  • Added permute in torch docs (#60821)
  • torch.chunk: Fixed a bug when chunk may return less than the requested number of chunks silently if some undocumented division constraints are not met (#61097)
  • Updated faq.rst so OOM section mentions checkpoint (#62709)
  • Documented that torch.triangular_solve has optional out= parameter (#63253)
  • Fixed incorrect type in docstring for torch.optim.Optimizer.state_dict (#63153)
  • Fixied trapezoid() comments for clarity (#64592)
  • Removed .data from some docs (#65358)
  • Added mention of vsplit, hsplit and tensor_split in Tensor views doc (#63191)
  • Added cross references to torch.transpose and torch.t (#63177)
  • Removed input parameter from Tensor.flatten docs (#63180)
  • Merged note block in torch.lu documentation (#63156)
  • Formated torch.permute docstring with common args (#61833)
  • Added math section for the following: torch.optim.{Adagrad, Adadelta, Adam, AdamW, Adamax, RAdam, Rprop, SGD} (#63251, #63252, #63254, #63255, #63903, #63772, #63866, #63805)
  • Added copy button to code snippets in docs (#63149)
  • Updates set_default_dtype documentation (#63233)
  • Extended warning on norm docs (#63310)
  • Added expecttest to as dependency for testing in CONTRIBUTING.md (#61163)
  • Clarified the documentation of torch.meshgrid (#62977)
  • torch.autograd.function.FunctionCtx: Improved custom function documentation (#60312)
  • Documented why wrappers exist in torch.functional (#62847)
  • Clarified that torch.finfo.tiny is the smallest normal number (#63241)
  • Modernized torch.{add,mul} documentation (#63309)
  • Added mathematical description for optimizers Nesterov Adam (torch.optim.NAdam , torch.optim.RMSProp) (#63793, #63721)
  • Improved torch.hub docs (#63783)
  • Fixed default value of last_epoch for torch.optim.swa_utils.SWALR doc (#62799)

Complex Number

  • Added a note to the documentation of polar (#63259)

Autograd

  • Updated intro doc for derivatives.yaml (#60614)
  • Clarified default value of requires_grad for tensors (#61038)
  • Added docs describing saved tensor hooks (#62362)
  • Clarified deprecation notice for only_inputs argument for .grad (#63631)

torch.nn

  • nn.AvgPool2d: Fixed description for divisor_override in docs (#62670)
  • nn.ConvTranspose*d: Clarified docs and added reference links (#60291)
  • nn.LayerNorm: Clarified normalization behavior with additional example and diagram and added parameter docs (#59178, #63144)
  • nn.LSTM: Fixed parameter shape docs for proj_size > 0 (#65102)
  • nn.LSTM: Added parameter shape docs for bidirectional case (#60120)
  • nn.MaxPool1d: Fixed docs to indicate implicit padding of negative infinity for non-zero padding amount (#59404)
  • nn.Module: Added description for behavior of registering None with register_buffer() / register_parameter() (#59015, #60125, #60549)
  • nn.Module: Fixed link in register_module_backward_hook() docs (#61999)
  • nn.Module: Added register_module_full_backward_hook() to docs (#58954)
  • nn.Module: Updated full backward hook docs with not-same-object note (#63245)
  • nn.MultiheadAttention: Expanded input / output docs + corrected shape info for query, key, and value embedding dimensions (#61977)
  • nn.MultiMarginLoss: Expanded docs to include example and shape information (#63760)
  • nn.MultiMarginLoss: Fixed equation format error (#59188)
  • Clarified whether probabilities or log-probabilities are expected for inputs and targets of nn.{BCELoss, BCELossWithLogits, CrossEntropyLoss, KLDivLoss, NLLLoss} (#60553)
  • Clarified input and output shape documentation for nn.{Flatten, Unflatten} (#60980)
  • Added code comments to clarify initialization behavior for nn.{Conv*d, Linear} (#58931, #58487)
  • Added previously undocumented nn.{GLU, FeatureAlphaDropout``} to docs (#60590)
  • Expanded description for unsafe flag in ParametrizationList docs (#60899)
  • F.interpolate: Clarified docs for recompute_scale_factor (#58989)
  • F.kl_div: Fixed formatting of reduction mode in docs (#59636)
  • Expanded modules note to include more detail on module state / initialization, hook examples, and introductions to distributed, profiling, quantization, pruning, FX, TorchScript, and parametrizations (#63963)

Profiler

  • Added a warning about potential overhead to the documentation of torch.profiler.profile with_shapes argument (#59469)
  • Updated kineto repo README files (#306, #434, #433)

Dataloader

  • Added a warning about “memory leak” when DataLoader has num_workers > 0 (#64337)

AMD

  • Added note on torch.distributed backends on ROCm (#58975)
  • Added note on ifdefing based on CUDA_VERSION for ROCm path (#62850)

CUDA

  • Fixed cppdoc example syntax (#62385)
  • Removed old references to 9.2 in documentation (#65059)
  • CUDA graphs: moved memory sharing intro paragraph (#64996)
  • CUDA graphs: designated CUDA graphs as Beta, not prototype feature (#65247)

TorchScript

  • Updated optimize_for_inference docs (#64428)

torch.package

  • Added torch.package documentation preamble (#59491)
  • Added tutorials section to torch.package docs (#59499)
  • Added “how do I” section to torch.package docs (#59503)
  • Added explanation section to torch.package docs (#59833)
  • Added table of contents to torch.package docs (#59842)
  • Added docstring for PackageExporter.intern (#59602)
  • Cleaned up doc strings (#59843)

Mobile

  • Updated iOS readme to use lite interpreter (#59841)

Quantization

  • Fixed docstring for fusing functions (#58638)
  • Improved documentation on natively supported backends (#58925)
  • Added description of qconfig and qengine to quantization page (#63582)
  • Clarified operator support for quantization (#63270)
  • Fixed quantization.rst doc (#64802)
  • Added overload information for torch.fake_quantize_per_tensor_affine (#63258)
  • Added missing documentation for torch.quantized_batch_norm (#63240)
  • Added missing docstrings for torch.quantized_max_pool{1,2}d (#63242)
  • Updated quantize_per_tensor to mention overloads (#63165)

Distributed

torch.distributed

  • Documented debugability features in torch.distributed (#59726, #58322)
  • Fixed broken hyperlinks (#59425)
  • Added compare_set in distributed docs (#61351)
  • Added tutorial link (#62785)
  • Improved distributed.get_rank() API docstring (#63296)
  • Added warning on isend about modifying after send (#61875)

DistributedDataParallel

  • Enhanced comm hook docs (#61677)
  • Updated DDP documentation to mention outputs not used in loss is supported (#60275)
  • Documented bf16 gradient compression hook (#64346)

torch.distributed.optim.ZeroRedundancyOptimizer

  • Introduced minor documentation fixes (#61785)

torch.futures

  • Fixed torch.futures docstring examples (#61029)

torch.fx

  • Improved error message when Proxy object is iterated (#58302)
  • Fixed link to fuser example (#59670)
  • Fixed subgraph API call_module warning about no owning module (#61463)
  • Added note about python set pitfall (#61597)
  • Added documentation for shape inference algorithm (#64312)

ONNX

  • Introduced a major rewrite of documentation (#60249)

Misc

Performance_as_a_product

  • Added warning about poor performance creating Tensor from list of numpy.array’s (#51680)

Build_Frontend

  • Fixed mismatch in README.md Docker Image section (#59199)
  • Fixed wrong command in README.md (#62472)
  • Updated the pytorch-gdb example so that it works on current master (#61175)
  • Advertise USE_PRECOMPILED_HEADERS in CONTRIBUTING.md (#62827)

Infra (Releng)

  • Added note about nightly versions bump (#59324)
  • Created SECURITY.md (#61356)
  • Fixed grammatical errors on the PyTorch Contribution Guide (#61818)
  • Added small fixes to the Contributing.txt (#63385)
  • Updated persons_of_interest.rst (#63907)