2 Pull requests merged by 2 people

• Fix staging for CPU tensors in OSS DCP async_save

  #145408 merged Jan 23, 2025

• Prevent legacy_load when weights_only=True (correctly)

  #145111 merged Jan 17, 2025

193 Pull requests opened by 122 people

• Revert "Fix for MSVC problem on Windows Arm64 (#136765)"

  #145076 opened Jan 17, 2025

• Remove FFT from stride incorrect ops

  #145080 opened Jan 17, 2025

• partitioner: avoid inserting duplicates into heap

  #145082 opened Jan 17, 2025

• cpp_wrapper: Move #includes to per-device header files

  #145083 opened Jan 17, 2025

• `torch.distributions`: replace `numbers.Number` with `torch.types.Number`.

  #145086 opened Jan 17, 2025

• [POC] Extend torch function support to ALL arguments, not just scalar type (but not insides of list)

  #145089 opened Jan 17, 2025

• Test

  #145090 opened Jan 17, 2025

• cpp_wrapper/aot_inductor: handle conjugation and negation dispatch keys

  #145095 opened Jan 17, 2025

• Use STL string_view header

  #145098 opened Jan 17, 2025

• Maintain multiple configs

  #145103 opened Jan 17, 2025

• futher scheduler changes for invoke_quant: prologue low prec, (slightly) more aggressive fusion

  #145104 opened Jan 17, 2025

• WIP remove -E workaround for nvcc

  #145116 opened Jan 17, 2025

• [EXPERIMENTAL][dynamo] optimize `DictGetItemGuardAccessor`

  #145117 opened Jan 17, 2025

• WIP sccache simplified

  #145119 opened Jan 17, 2025

• [triton] Update triton pin to include warp specialization support

  #145120 opened Jan 17, 2025

• inductor: Don't throw an internal error when a nn.module is missing a attribute

  #145122 opened Jan 17, 2025

• Repro collective timeout and FR dump

  #145125 opened Jan 18, 2025

• test trigger dispatch

  #145126 opened Jan 18, 2025

• [executorch hash update] update the pinned executorch hash

  #145128 opened Jan 18, 2025

• [cuBLAS][cuBLASLt] Unify `cuBLASLt` workspaces with `cuBLAS` workspaces

  #145130 opened Jan 18, 2025

• [dynamo] Log guard latency

  #145132 opened Jan 18, 2025

• [inductor] [bug fix] Fix `conv` on processing uint

  #145136 opened Jan 18, 2025

• improve perf for layer_norm

  #145146 opened Jan 18, 2025

• [BE][Easy] increase pip timeout for nightly tool: 15s -> 60s

  #145147 opened Jan 18, 2025

• [BE][PYFMT] bump `ruff format` target version to py39: add parentheses around long `with`-statements

  #145148 opened Jan 18, 2025

• [inductor] Simplify _inductor/utils.py slightly

  #145150 opened Jan 18, 2025

• [BE]: Apply ruff PERF401 to torch

  #145153 opened Jan 18, 2025

• Make `inductor_utils.requires_gpu` accept MPS

  #145156 opened Jan 18, 2025

• [BE]: Update NCCL submodule to 2.24.3

  #145167 opened Jan 19, 2025

• Added weight to MSELoss Criterion

  #145169 opened Jan 19, 2025

• [BE]: Update CUTLASS submodule to 3.7.0

  #145172 opened Jan 19, 2025

• [scan] scan dim handling in user-facing scan()

  #145179 opened Jan 19, 2025

• Added torch check to ensure indices are not empty

  #145180 opened Jan 19, 2025

• Add transpose support for CppMicroGemmFP32Vec

  #145194 opened Jan 20, 2025

• Guard size oblivious within empty_tensor_restride_symint

  #145196 opened Jan 20, 2025

• Use std::string_view in get_fully_qualified_type_name

  #145197 opened Jan 20, 2025

• CI test: TestAutograd.test_gradcheck_nondeterministic

  #145205 opened Jan 20, 2025

• Update slow tests

  #145206 opened Jan 20, 2025

• Fix incorrect citation of authors in documentation

  #145209 opened Jan 20, 2025

• solve apl dependency issue

  #145215 opened Jan 20, 2025

• Refactoring Distributed test cases to be device agnostic [1/n]

  #145222 opened Jan 20, 2025

• Raise MutationError if there are side effects when returning generator

  #145223 opened Jan 20, 2025

• update sympy version 1.13.3 in setup.py (previously update only in requirement.txt)

  #145224 opened Jan 20, 2025

• fix test_cublas_workspace_explicit_allocation for gfx12

  #145227 opened Jan 20, 2025

• Expose the rendezvous keepalive arguments

  #145228 opened Jan 20, 2025

• Improve typing in torch/types.py

  #145237 opened Jan 21, 2025

• Improve typing in torch/__init__.py

  #145238 opened Jan 21, 2025

• Improve typing in torch/_C/__init__.pyi.in

  #145239 opened Jan 21, 2025

• add grad_output shape check for adaptive_avg_pool2d_backward

  #145241 opened Jan 21, 2025

• [inductor] Make serialized inductor patterns path configurable instead of using …

  #145243 opened Jan 21, 2025

• Improve typing by using bool and int

  #145244 opened Jan 21, 2025

• [Quant][CPU] add a wrapper op in quantized_decomposed for _weight_int4pack_mm_for_cpu

  #145245 opened Jan 21, 2025

• [Inductor UT] Set input tensors to corresponding device for test case in test_aot_indutor.py

  #145248 opened Jan 21, 2025

• [Inductor][CPU] Add a lowering pass for _weight_int4pack_mm_for_cpu

  #145250 opened Jan 21, 2025

• change the test wheel to release wheel when release wheel available

  #145252 opened Jan 21, 2025

• [TEST] tmp storage with CONSTANTHANDLE

  #145254 opened Jan 21, 2025

• [ARM] Add test_ops and test_memory_profiler to aarch64 tests

  #145260 opened Jan 21, 2025

• Fix SEGFAULT when None arg was passed in GraphContext.op(..)

  #145265 opened Jan 21, 2025

• Improve the caching allocator test for raw alloc

  #145269 opened Jan 21, 2025

• Remove unnecessary HPUHooksInterface method

  #145272 opened Jan 21, 2025

• Updates NCCL user buffer registration test for NCCL 2.24.3

  #145285 opened Jan 21, 2025

• update get start xpu

  #145286 opened Jan 21, 2025

• [ROCm] miopen benchmark behavior now better aligns with cudnn

  #145294 opened Jan 21, 2025

• Add unique identifer to bmm thread_mm functions

  #145303 opened Jan 21, 2025

• [BE] Remove test_ops from FIXME_inductor_dont_reset_dynamo

  #145307 opened Jan 21, 2025

• [Utilization] post-test-process workflow

  #145310 opened Jan 21, 2025

• windows builds with VS2022

  #145319 opened Jan 21, 2025

• inductor: Explicitly test that torch.compile(option=...) does something

  #145321 opened Jan 21, 2025

• fix a small typo in comments

  #145323 opened Jan 21, 2025

• Add stft option to align window for center = false

  #145324 opened Jan 21, 2025

• [utilization] pipeline to create clean db records

  #145327 opened Jan 22, 2025

• [WIP] [AOTInductor] Use AtenTensorHandle as the constant map's holder.

  #145331 opened Jan 22, 2025

• PEP585: Missed conversions

  #145342 opened Jan 22, 2025

• [dtensor][cp] experiment: call flex_attention on DTensor

  #145353 opened Jan 22, 2025

• ehnace logging statically known by adding size_oblivious(..)

  #145354 opened Jan 22, 2025

• [WIP] Fix avg_pool crash with negative numbers

  #145358 opened Jan 22, 2025

• removed check for ConvTranspose3D on MPS

  #145366 opened Jan 22, 2025

• [ARM] Fix broken tests in test_tensor_creation_ops on AArch64

  #145367 opened Jan 22, 2025

• Enable C++ API parity tests on AArch64

  #145370 opened Jan 22, 2025

• [inductor][BE] Enable test_cpu_cpp_wrapper in fbcode

  #145373 opened Jan 22, 2025

• [torchbench] Increase tolerance for amp only poolformer_m36

  #145375 opened Jan 22, 2025

• Use AOTI as inductor backend when fullgraph_package is enabled.

  #145381 opened Jan 22, 2025

• [DO NOT MERGE] pre-merge runs only on MI200 and post-merge runs on both MI300

  #145389 opened Jan 22, 2025

• Update NJT linear_backward to return non-aliased tensor bias grad

  #145399 opened Jan 22, 2025

• [auto_functionalized] Support `Tensor(a!)[]?`

  #145400 opened Jan 22, 2025

• Update OSS nested tensor docs to focus on NJT

  #145402 opened Jan 22, 2025

• Use guard_size_oblivious in debug tensor writer

  #145403 opened Jan 22, 2025

• [distributions] Catch inf gradient in beta distribution

  #145404 opened Jan 22, 2025

• [export][be] Clean up local imports from export [2/n]

  #145406 opened Jan 22, 2025

• [dynamo][fbcode] Turn on inline_inbuilt_nn_modules

  #145407 opened Jan 22, 2025

• Add Torchao docs link to Pytorch libraries

  #145412 opened Jan 22, 2025

• [dynamo][not ready - just for CI] Remove all builtin skiplist

  #145415 opened Jan 22, 2025

• TopK ROCm Tuning

  #145416 opened Jan 22, 2025

• [dynamo][hop] test torch.compiling all user-facing HOPs

  #145422 opened Jan 22, 2025

• Tag storages with offset in file when with FakeTensorMode

  #145424 opened Jan 22, 2025

• Fix aot inductor intermediate debug printing

  #145426 opened Jan 22, 2025

• add pt2 callbacks for backward pass

  #145427 opened Jan 23, 2025

• [ca][hop] test CA on all HOPs

  #145429 opened Jan 23, 2025

• [dynamo] Re-enable `test_torch_name_rule_map_updated`

  #145431 opened Jan 23, 2025

• [dynamo] save/restore system random state more carefully

  #145435 opened Jan 23, 2025

• Advance past fc window for stft center

  #145437 opened Jan 23, 2025

• [draft_export] add LOC for data-dep error logging

  #145443 opened Jan 23, 2025

• [Docs] Add clarification for target types in CrossEntropyLoss doc

  #145444 opened Jan 23, 2025

• Record inputs at time of tracing, constrain to them for triton fn

  #145448 opened Jan 23, 2025

• Fix incorrect type comparison

  #145449 opened Jan 23, 2025

• Replace is_same with is_same_v for concise syntax

  #145450 opened Jan 23, 2025

• Add check that envvar configs are boolean

  #145454 opened Jan 23, 2025

• Update TorchBench commit to main

  #145455 opened Jan 23, 2025

• [c10d] Flush file in file recorder

  #145458 opened Jan 23, 2025

• [AOTInductor] Align behavior between CPU and GPU

  #145459 opened Jan 23, 2025

• [TEST ONLY] Conv with `oc = 0`

  #145462 opened Jan 23, 2025

• OpenReg: Refactor impl_registry

  #145465 opened Jan 23, 2025

• [Intel GPU] Add TORCH_API macro to export symbol NestedTensor_to_mask for libtorch_xpu

  #145467 opened Jan 23, 2025

• fix test_convolution error when use cudnn.flags

  #145474 opened Jan 23, 2025

• [dynamo] added support to trace torch.cuda.is_current_stream_capturing

  #145475 opened Jan 23, 2025

• Adapt Dynamo Tests to HPUs

  #145476 opened Jan 23, 2025

• Modify enable logic of COLLECTIVE_COMM profiler activity type

  #145478 opened Jan 23, 2025

• [Dynamo] Fix names collisions with foreach decomps

  #145479 opened Jan 23, 2025

• simplify torch.utils.cpp_extension.include_paths; use it in cpp_builder

  #145480 opened Jan 23, 2025

• feat: add SVE dispatch for non-FBGEMM qembeddingbag

  #145486 opened Jan 23, 2025

• Remove unnecessary "special linking" for `BLAS_LIBRARIES`

  #145487 opened Jan 23, 2025

• [torchbench] Add meta function for _cudnn_rnn_flatten_weight

  #145488 opened Jan 23, 2025

• cpp_wrapper: Move #includes to per-device header files

  #145490 opened Jan 23, 2025

• [BE]: Fix OrderedSet equality oversight

  #145492 opened Jan 23, 2025

• [inductor] Make triton kernel autotune config defaults backward-compatible

  #145494 opened Jan 23, 2025

• [compiled_autograd] Rename interface to pyinterface

  #145495 opened Jan 23, 2025

• Remove truncated normal initialization for 16-bit (and lower) tensors

  #145499 opened Jan 23, 2025

• [ROCm] Update workflow to use root user instead of jenkins user

  #145504 opened Jan 23, 2025

• Bump AOTriton to 0.8.2b

  #145508 opened Jan 23, 2025

• [dynamo][guards] Log guard latency to tlparse

  #145509 opened Jan 23, 2025

• Add istft option to align window for center = false

  #145510 opened Jan 23, 2025

• Add missing autoreleasepool around runUniqueGraph to prevent leaks

  #145512 opened Jan 23, 2025

• [WIP][inductor][5/N] triton support post-#5512, fix 1 and None handling

  #145515 opened Jan 23, 2025

• [ROCM MI300 skips for flaky unit tests

  #145518 opened Jan 23, 2025

• Fix allow_mutation_on_saved_tensors for inplace foreach

  #145520 opened Jan 23, 2025

• General Changes for multi accelerators

  #145521 opened Jan 23, 2025

• inductor.config.descriptive_names = False is not actually supported

  #145523 opened Jan 23, 2025

• Fix IdentationError of code example

  #145525 opened Jan 23, 2025

• [MPS] Add bilineard2d_aa implementation

  #145526 opened Jan 23, 2025

• fix intermediate debug information with cpp_wrapper

  #145527 opened Jan 23, 2025

• [utils] add try_import method for importing optional modules

  #145528 opened Jan 23, 2025

• Work around buggy use_const_ref_for_mutable_tensors

  #145530 opened Jan 23, 2025

• Disable slow gradcheck for nn.Transformer ModuleInfo

  #145531 opened Jan 23, 2025

• Make sure that benchmark_harness is set before running

  #145532 opened Jan 23, 2025

• Remove det_singular OpInfo

  #145533 opened Jan 23, 2025

• Increase the number of perf benchmark shards

  #145534 opened Jan 23, 2025

• Removes threadfence from topk kernel to improve AMD performance

  #145536 opened Jan 23, 2025

• [dynamo] Properly model torch profiler context objects

  #145537 opened Jan 23, 2025

• Make sure not using cpp wrapper when setting nvtx training annotation

  #145538 opened Jan 23, 2025

• Add accuracy issue support in AOTI Minifier

  #145539 opened Jan 23, 2025

• [BE][hop] make it easier to use speculate_subgraph

  #145540 opened Jan 23, 2025

• [BE] Type annotate wrapper_benchmark.py and cuda_combined_scheduling.py

  #145542 opened Jan 23, 2025

• Testing #144594

  #145546 opened Jan 23, 2025

• [dynamo][refactor] Move collections.namedtuple out of SkipFunctionVariable

  #145547 opened Jan 23, 2025

• fix unbacked + view incorrectness

  #145548 opened Jan 23, 2025

• [ca] add test_reset for 2.6 release validation

  #145549 opened Jan 23, 2025

• If mypy fails it should report the error back to lintrunner

  #145550 opened Jan 23, 2025

• Remove incorrect BuiltinVariable.call_hasattr()

  #145551 opened Jan 23, 2025

• Turn on mypy for _dynamo/variables/builtin.py

  #145552 opened Jan 23, 2025

• Fix call to create_load_global

  #145553 opened Jan 23, 2025

• Fix dynamo use of `list[int]` in graph break

  #145554 opened Jan 23, 2025

• Add CUDA 12.8 installation and Linux CD Docker images

  #145557 opened Jan 23, 2025

• [dynamo][builtin-skipfile-cleanup] Support tuple.__new__

  #145558 opened Jan 23, 2025

• [dynamo][builtin-skipfile-cleanup] Remove collections

  #145559 opened Jan 23, 2025

• [Not for land] hacking up mx

  #145562 opened Jan 24, 2025

• Refactor fuzzer and add support for Dynamo

  #145565 opened Jan 24, 2025

• Advance docker release latest verison to cuda 12.4

  #145566 opened Jan 24, 2025

• Add CUDA 12.8 installation and manylinux-cuda12.8

  #145567 opened Jan 24, 2025

• [mps] Hoist erfinv logic out of the kernel in preparation for moving.

  #145568 opened Jan 24, 2025

• [aotinductor] update unbacked symint runtime assertion msg

  #145569 opened Jan 24, 2025

• [BE] mv test/inductor_skips/* to test/inductor_expected_failures/

  #145572 opened Jan 24, 2025

• [inductor/profiler] add kernel kwargs instrumentation

  #145573 opened Jan 24, 2025

• [inductor][3/N] triton support post-#5512, tt.divisibility format

  #145575 opened Jan 24, 2025

• [Inductor][CPP] fix torch logit decomposition

  #145576 opened Jan 24, 2025

• [inductor] Fix duplicate detection in _dynamic_scale_rblock

  #145577 opened Jan 24, 2025

• Spruce up docs for emulate_precision_casts

  #145579 opened Jan 24, 2025

• [MPS][BE] Implement bilineard2d as shader

  #145581 opened Jan 24, 2025

• [inductor][4/N] triton support post-#5512, fix constexpr signatures

  #145583 opened Jan 24, 2025

• [ROCm] Eliminate the need for divisions in layernorm for default vector size

  #145584 opened Jan 24, 2025

• [Custom Ops] Add a new API to allow users to register an autocast for the custom op

  #145588 opened Jan 24, 2025

• Replace decorators in UTs to cover additional devices

  #145589 opened Jan 24, 2025

• [dynamo][benchmarks] Stop benchmarking compile time of dead code

  #145590 opened Jan 24, 2025

• [CCA] remove TODO for hardware_destructive_interference_size

  #145591 opened Jan 24, 2025

• Fix constants with non-functional operators

  #145593 opened Jan 24, 2025

• [micro_pipeline_tp] add logging for all-gather-matmul fusion

  #145594 opened Jan 24, 2025

• [micro_pipeline_tp] support pattern matching row-wise scaled_mm with sharded scale

  #145595 opened Jan 24, 2025

• Add `torch._foreach_copy_` doc

  #145597 opened Jan 24, 2025

• [NFC] Fix some minor typos.

  #145599 opened Jan 24, 2025

• Add device support for chunk_cat, all_gather_copy_in, and split_with_…

  #145600 opened Jan 24, 2025

• [ATen][CUDA][Transformers] Add Blackwell support to SDPA

  #145602 opened Jan 24, 2025

• [dynamo] refactor dynamo__custom_eval_frame to C++, refactor SKIP_CODE[_RECURSIVE]

  #145603 opened Jan 24, 2025

• WIP error_prop sc

  #145605 opened Jan 24, 2025

• [BE][CI] bump ruff to 0.9.3

  #145606 opened Jan 24, 2025

148 Issues closed by 54 people

• DISABLED test_arange_dynamic_cuda (__main__.TestInductorDynamicCUDA)

  #127067 closed Jan 24, 2025

• torch.sin/cos/tan+torch.floor/round may bring wrong results with torch.compile

  #145466 closed Jan 24, 2025

• torch crashes on ubuntu:24.04 during SDPA-CuDNN test

  #145580 closed Jan 24, 2025

• Error: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory

  #145369 closed Jan 24, 2025

• Inference super slow with torchvision model fasterrcnn_mobilenet_v3_large_fpn

  #145032 closed Jan 24, 2025

• DISABLED test_bw_decoding_fails_float16 (__main__.TestFlexDecoding)

  #141761 closed Jan 24, 2025

• Int8 Inference Slowdown Comparing to FP32 when using PyTorch 2 Export Quantization with X86 Backend through Inductor

  #144434 closed Jan 24, 2025

• lerp_ doesn't correctly type promote

  #140601 closed Jan 24, 2025

• HOP input mutation analysis is not comprehensive

  #137639 closed Jan 23, 2025

• CUDA memory leak in model_container::run_const_fold

  #126059 closed Jan 23, 2025

• Flexattention: ValueError: Shape element 1 must be a power of 2

  #133321 closed Jan 23, 2025

• DISABLED test_fn_grad_linalg_det_singular_cuda_complex128 (__main__.TestBwdGradientsCUDA)

  #93044 closed Jan 23, 2025

• DISABLED test_forward_mode_AD_linalg_det_singular_cuda_complex128 (__main__.TestFwdGradientsCUDA)

  #93045 closed Jan 23, 2025

• qmul.cpp:34:10: error: redefinition of 'xnn_binary_params' 34 | struct xnn_binary_params { | ^

  #145497 closed Jan 23, 2025

• Async distributed checkpointing works incorrectly with tensors on CPU

  #144657 closed Jan 23, 2025

• cannot pickle 'torch._C._aoti.AOTIModelPackageLoader' object

  #145411 closed Jan 23, 2025

• DISABLED test_trigger_bisect_on_error (__main__.ExcTests)

  #131303 closed Jan 23, 2025

• PR #89436 looks like it causes or enables a memory leak

  #90464 closed Jan 23, 2025

• "Unknown builtin op" error during jit.load() of TorchScript module with @custom_op

  #143773 closed Jan 23, 2025

• Memory Leak in MPS Backend During LSTM Iterations (Out of Memory Error)

  #145374 closed Jan 23, 2025

• cloning third_party/kleidiai fails

  #145273 closed Jan 23, 2025

• DISABLED test_unbacked_bindings_for_divisible_u_symint_training_ir_to_decomp_non_strict (__main__.TrainingIRToRunDecompExportNonStrictTestExport)

  #138582 closed Jan 23, 2025

• DISABLED test_unbacked_bindings_for_divisible_u_symint (__main__.TestExport)

  #138586 closed Jan 23, 2025

• DISABLED test_unbacked_bindings_for_divisible_u_symint_training_ir_to_decomp (__main__.TrainingIRToRunDecompExportTestExport)

  #138583 closed Jan 23, 2025

• DISABLED test_unbacked_bindings_for_divisible_u_symint_non_strict (__main__.NonStrictExportTestExport)

  #138585 closed Jan 23, 2025

• DISABLED test_unbacked_bindings_for_divisible_u_symint_retraceability (__main__.RetraceExportTestExport)

  #138584 closed Jan 23, 2025

• DISABLED test_unbacked_bindings_for_divisible_u_symint_retraceability_non_strict (__main__.RetraceExportNonStrictTestExport)

  #138676 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_retraceability_non_strict (__main__.RetraceExportNonStrictTestExport)

  #138675 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_training_ir_to_decomp_non_strict (__main__.TrainingIRToRunDecompExportNonStrictTestExport)

  #131136 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_non_strict (__main__.NonStrictExportTestExport)

  #131088 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_retraceability (__main__.RetraceExportTestExport)

  #131083 closed Jan 23, 2025

• [inductor] [cuda] [silence] `F.gumbel_softmax` return inconsistent resutls compared with eager

  #145470 closed Jan 23, 2025

• Failed to export the model to ONNX

  #144750 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_training_ir_to_decomp (__main__.TrainingIRToRunDecompExportTestExport)

  #131082 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_serdes_non_strict (__main__.SerDesExportNonStrictTestExport)

  #138884 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv_serdes (__main__.SerDesExportTestExport)

  #131119 closed Jan 23, 2025

• DISABLED test_slice_with_floordiv (__main__.TestExport)

  #131101 closed Jan 23, 2025

• Link to `third_party/eigen` git submodule is broken

  #145496 closed Jan 23, 2025

• [Tracking Issue] Mixed precision does not work with ignored modules

  #90318 closed Jan 23, 2025

• torch.jit.trace wrong function mapping: > maps to aten::lt

  #145485 closed Jan 23, 2025

• libtorch_python.dylib not getting symlinked correctly in OSX 13 with pytorch-cpu

  #145469 closed Jan 23, 2025

• torch.backends.cudnn.flags use error when test

  #145472 closed Jan 23, 2025

• Non_blocking copy behavior on non-cuda/non-privateuse1 accelerator might be unexpected

  #143641 closed Jan 23, 2025

• DISABLED test_device_mode_ops_sparse_sampled_addmm_cpu_float32 (__main__.TestDeviceUtilsCPU)

  #132720 closed Jan 23, 2025

• torch._neg_view correctness

  #145428 closed Jan 23, 2025

• [RFC] Improve performance for softmax op for cuda in some specific size

  #144645 closed Jan 23, 2025

• DISABLED test_autograd_function_backed_op (__main__.TestCustomOpWithCompiledAutograd)

  #121342 closed Jan 22, 2025

• DISABLED test_aot_sequence_nr (__main__.DynamicShapesAotAutogradFallbackTests)

  #106440 closed Jan 22, 2025

• DISABLED test_no_grad_copy (__main__.TestAutograd)

  #139734 closed Jan 22, 2025

• Footgun: tracer.root.register_module( in HOPs

  #140760 closed Jan 22, 2025

• DISABLED test_second_order_accurate (__main__.TestGradient)

  #116746 closed Jan 22, 2025

• No period in docstring of torch.compiler.disable

  #145365 closed Jan 22, 2025

• DISABLED test_large_weight_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #127068 closed Jan 22, 2025

• DISABLED test_large_mmaped_weights_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #127202 closed Jan 22, 2025

• DISABLED test_cpp_frontend_module_has_same_output_as_python (__main__.TestCppExtensionJIT)

  #116105 closed Jan 22, 2025

• [aarch64] multiple inductor test failures related to vec128_bfloat16

  #144818 closed Jan 22, 2025

• [Inductor][GPU] Input is padded with incorrect value when executing `torch.nn.functional.pad` on gpu

  #144462 closed Jan 22, 2025

• [inductor][gpu] torch.fft.fft outputs incorrect results when `n>1`

  #143719 closed Jan 22, 2025

• [inductor][gpu] torch.nn.functional.avg_pool1d outputs incorrect result when input.numel() is 1

  #143720 closed Jan 22, 2025

• Release Pyotrch version 2.6.0 in pypi

  #145142 closed Jan 22, 2025

• [Device] `ConvTranspose` bahaves differently on CPU and CUDA when `out_channels=0`

  #142466 closed Jan 22, 2025

• DISABLED test_fs_preserve_sharing (__main__.TestMultiprocessing)

  #91467 closed Jan 22, 2025

• DISABLED test_min_cut_partitioner_recomputable_ops (__main__.TestPartitioning)

  #104327 closed Jan 22, 2025

• make latexpdf

  #145221 closed Jan 22, 2025

• DISABLED test_cpp_extension_recommends_custom_ops_dynamic_shapes (__main__.DynamicShapesMiscTests)

  #127813 closed Jan 22, 2025

• DISABLED test_fake_crossref_backward_no_amp_index_fill_cuda_float32 (__main__.TestFakeTensorCUDA)

  #99126 closed Jan 22, 2025

• Dtype available for `torch.optim.Adam` and `torch.optim.AdamW` when `fused=True` is different from described

  #145282 closed Jan 22, 2025

• Optimizer state cannot get offloaded to CPU

  #144397 closed Jan 22, 2025

• When calling a custom function of a LlamaForCausalLM using FSDP causes RuntimeError

  #145281 closed Jan 22, 2025

• XPU builds validations

  #145290 closed Jan 22, 2025

• DISABLED test_open_device_registration (__main__.TestCppExtensionOpenRgistration)

  #100152 closed Jan 22, 2025

• DISABLED test_basic (__main__.TestPythonDispatch)

  #145096 closed Jan 22, 2025

• DISABLED test_max_autotune_remote_caching_dynamic_False (__main__.TestMaxAutotuneRemoteCache)

  #145360 closed Jan 22, 2025

• loss.backward() breaking somewhere when modulating a nested tensor using scale and shift (RuntimeError: Function AddBackward0 returned an invalid gradient at index 0 - got [1, 4, 64] but expected shape compatible with [4, 1, 64])

  #145256 closed Jan 22, 2025

• DISABLED test_comprehensive_fft_ifft_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #127344 closed Jan 22, 2025

• nn.Embedding backwards pass for nested tensors

  #145257 closed Jan 22, 2025

• DISABLED test_aot_module_simplified_fake_tensor_gm_raises (__main__.TestAOTModuleSimplified)

  #124590 closed Jan 22, 2025

• Dynamo graph break on PEP585 generic types

  #145226 closed Jan 22, 2025

• trace.save_real_tensors segfaults on resnet

  #143524 closed Jan 21, 2025

• Investigate potential cost savings for inductor workflows

  #138476 closed Jan 21, 2025

• DISABLED test_identity_float32 (__main__.TestTemplatedSDPA)

  #124659 closed Jan 21, 2025

• Accessing secrets variables in CI

  #144853 closed Jan 21, 2025

• [CD] Nightly Release Linux Manywheel builds add size check

  #137362 closed Jan 21, 2025

• Flakybot fails to fetch test ownership information

  #144964 closed Jan 21, 2025

• Binaries Python 3.13t failing linux-aarch64-binary-manywheel and linux-binary-manywheel

  #145234 closed Jan 21, 2025

• DISABLED test_alibi_causal_float32 (__main__.TestTemplatedSDPA)

  #124588 closed Jan 21, 2025

• DISABLED test_alibi_bias_float32 (__main__.TestTemplatedSDPA)

  #124526 closed Jan 21, 2025

• torch/_prims/executor.py #TODO : caching

  #145171 closed Jan 21, 2025

• isin prevents dynamic shapes in modules

  #142507 closed Jan 21, 2025

• DISABLED test_autograd_in_attr (__main__.TestPythonDispatch)

  #145068 closed Jan 21, 2025

• DISABLED test_returning_symint (__main__.TestPythonRegistration)

  #144920 closed Jan 21, 2025

• DISABLED test_register_functional_op_multiple_returns (__main__.TestPythonRegistration)

  #142807 closed Jan 21, 2025

• DISABLED test_override_aten_ops_with_multiple_libraries (__main__.TestPythonRegistration)

  #142460 closed Jan 21, 2025

• DISABLED test_register_fallthrough (__main__.TestPythonRegistration)

  #142494 closed Jan 21, 2025

• DISABLED test_override_cuda_with_jiterator (__main__.TestPythonRegistration)

  #142495 closed Jan 21, 2025

• DISABLED test_register_functional_op_with_optional (__main__.TestPythonRegistration)

  #117871 closed Jan 21, 2025

• DISABLED test_register_functional_op_no_returns (__main__.TestPythonRegistration)

  #117834 closed Jan 21, 2025

• Performance regression when using @torch.compile compared to no compilation

  #144822 closed Jan 21, 2025

• F.scaled_dot_product_attention get query @ key

  #145276 closed Jan 21, 2025

• `torch.compile` may produce wrong result with `BicubicInterp+Neg+Linear+Tan`.

  #145264 closed Jan 21, 2025

• `torch.compile` may produce wrong result with `torch.nn.functional.interpolate`.

  #145268 closed Jan 21, 2025

• Calculation Results Become NaN After Using `torch.compile` with `Matmul+Concat4+Mul+Linear+Tan`.

  #145266 closed Jan 21, 2025

• A confusion about Bidirectional GRU

  #145073 closed Jan 21, 2025

• CPU-only PyTorch on M1 MacBook always gets "RuntimeError: Placeholder storage has not been allocated on MPS device!"

  #145229 closed Jan 21, 2025

• Exporting a model with dynamic axes and dynamo fails with `TypeError: unhashable type: 'list'`

  #144860 closed Jan 21, 2025

• Cannot build static windows libraries

  #111905 closed Jan 21, 2025

• Investigate CUDA enabled build-time difference between MSVC and GCC+WSL

  #91623 closed Jan 21, 2025

• DISABLED test_out_of_order_index_ds (__main__.TestOutOfOrderDataLoader)

  #142343 closed Jan 21, 2025

• BUG: torch.exp for complex types on Linux chokes in some cases

  #136063 closed Jan 21, 2025

• torch.asin returns incorrect value with complex input on cpu

  #138327 closed Jan 21, 2025

• Floating point exception (core dumped) in `thnn_conv2d`

  #143489 closed Jan 21, 2025

• [XPU] Nightly binary builds for XPU Linux and Windows are failing since 01.11.2025

  #144967 closed Jan 20, 2025

• [Perf] Flash-Attn Bwd slow down w/ cutlass 3.6.0 in General

  #144729 closed Jan 20, 2025

• When using `torch.jit.trace` with `Linear+MaxPool2d+BatchNorm2d`, different results are observed.

  #145207 closed Jan 20, 2025

• Inconsistent results between CPU and GPU for many operators with complex inputs containing `Inf`

  #141487 closed Jan 20, 2025

• torch.nn.functional.normalize producing nan values with a large p value and tensor of complex numbers

  #135428 closed Jan 20, 2025

• `1/torch.inf` produce inconsistent results

  #106845 closed Jan 20, 2025

• torch.sigmoid producing nan for tensor of negative complex numbers on cpu

  #135777 closed Jan 20, 2025

• [complex] torch.{exp}: does not match numpy

  #48010 closed Jan 20, 2025

• DISABLED test_mm_concat_cuda (__main__.FreezingGpuTests)

  #145185 closed Jan 20, 2025

• torch.onnx.export failed with Process finished with exit code 136 (interrupted by signal 8:SIGFPE)

  #144144 closed Jan 20, 2025

• [XPU] Keep going jobs of `ciflow/xpu` when case fist failed.

  #145048 closed Jan 20, 2025

• DISABLED test_comprehensive_fft_fft_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #122715 closed Jan 20, 2025

• unexpected behaviour of `torch.chunk`

  #145026 closed Jan 19, 2025

• [DCP] BUG: FsspecWriter calls os.fsync on .finish(), therefore program crashes on checkpoint save

  #144752 closed Jan 19, 2025

• massive number of runtime asserts can hamper compile times

  #144792 closed Jan 18, 2025

• DISABLED test_integers_t1_uint8_np_longlong (__main__.TestArrayFromScalar)

  #145135 closed Jan 18, 2025

• DISABLED test_dtype_passthrough_dtype_complex128 (__main__.TestDLPack)

  #145134 closed Jan 18, 2025

• Noisy warning - torch.fx.experimental.symbolic_shapes: [WARNING] Ignored guard (...), this could result in accuracy problems

  #101265 closed Jan 18, 2025

• DISABLED test_flex_attention (__main__.TestCompiledAutograd)

  #144912 closed Jan 18, 2025

• DISABLED test_register_functional_op_one_return (__main__.TestPythonRegistration)

  #117816 closed Jan 18, 2025

• [Inductor] Test failure in test_comprehensive_nn_functional_max_pool2d_cuda

  #131072 closed Jan 17, 2025

• [AOTI] AOTI doesn't work well with torch.select

  #132360 closed Jan 17, 2025

• "index_cuda" not implemented for 'Float8_e4m3fn'

  #133605 closed Jan 17, 2025

• eps in layernorm.cpp causes a numerical transformation

  #140092 closed Jan 17, 2025

• [Compiled_autograd] running deepspeed Zero3 failed for torch.compile with compiled_autograd

  #141646 closed Jan 17, 2025

• DISABLED test_autograd_cpp_node_data_dependent (__main__.TestCompiledAutograd)

  #125579 closed Jan 17, 2025

• DISABLED test_autograd_cpp_node_saved (__main__.TestCompiledAutograd)

  #131103 closed Jan 17, 2025

• DISABLED test_autograd_cpp_node_saved_float (__main__.TestCompiledAutograd)

  #133197 closed Jan 17, 2025

• DISABLED test_autograd_cpp_node_saved_int (__main__.TestCompiledAutograd)

  #133283 closed Jan 17, 2025

• DISABLED test_non_traceable_autograd_cpp_node (__main__.TestCompiledAutograd)

  #134738 closed Jan 17, 2025

• DISABLED test_autograd_cpp_node_saved_dynamic (__main__.TestCompiledAutograd)

  #135685 closed Jan 17, 2025

• `unbind_copy` gives unexpected results on 1-dimensional inputs, or 0-dimensional outputs

  #130829 closed Jan 17, 2025

• [inductor][cpu]pyhpc_isoneutral_mixing performance regression in 2024-07-30 nightly release

  #132281 closed Jan 17, 2025

• [XPU] unrecognized device for new_qtensor: xpu:0

  #144848 closed Jan 17, 2025

• torch.select could not guard on data-dependent expression error

  #143249 closed Jan 17, 2025

• > if graph capture is thread local

  #137844 closed Jan 17, 2025

• non-strict export doesn't work with nn.Sequential slicing

  #137455 closed Jan 17, 2025

145 Issues opened by 85 people

• `torch.ops.aten.copy` causes SIGSEGV when handling sparse CSR tensors with invalid metadata

  #145604 opened Jan 24, 2025

• add scalar inputs with out causes error in torch.compile

  #145598 opened Jan 24, 2025

• [Dynamo] compile torch.logit with different data types

  #145596 opened Jan 24, 2025

• _pickle.UnpicklingError: invalid load key, ''.

  #145592 opened Jan 24, 2025

• [dynamo] mark_dynamic not working as intended with input shapes

  #145587 opened Jan 24, 2025

• [compile / strict export] torch._dynamo.exc.Unsupported: CollectiveFunctionRewriteVariable can't support async_op=True for <function all_reduce at 0x7f40be5724d0>

  #145574 opened Jan 24, 2025

• [BE] Automate update stable_cuda version so that we can set it when introducing new cuda version

  #145571 opened Jan 24, 2025

• Enable CUDA 12.8.0

  #145570 opened Jan 24, 2025

• [dynamo] Dynamo doesn't prune dead input cell object

  #145564 opened Jan 24, 2025

• Unable to build pytorch after #143806

  #145563 opened Jan 24, 2025

• Confusing as_storage_and_layout(x, want_contiguous=True) behavior

  #145561 opened Jan 24, 2025

• Error `RuntimeError: CUDA error: no kernel image is available for execution on the device` when doing `!=` operation on Jetson orin agx.

  #145560 opened Jan 23, 2025

• Docs fonts are bold on Mac, in 2.7

  #145556 opened Jan 23, 2025

• need to document `FlopCounterMode`

  #145555 opened Jan 23, 2025

• [RFC] Cuda support matrix for Release 2.7

  #145544 opened Jan 23, 2025

• Module.to() fail in dynamo when swap_module_params_on_conversion is true

  #145529 opened Jan 23, 2025

• use_const_ref_for_mutable_tensors doesn't work with out= overloads

  #145522 opened Jan 23, 2025

• Make debugging flaky tests easier by having relevant logs in one place

  #145516 opened Jan 23, 2025

• Can't properly implement backward method for custom op in C++ when the op takes List of tensors as argument

  #145514 opened Jan 23, 2025

• Activation Checkpointing composability with split backward computation

  #145511 opened Jan 23, 2025

• pip failure when trying to download nightly whl from pytorch.download.org : ERROR: THESE PACKAGES DO NOT MATCH THE HASHES FROM THE REQUIREMENTS FILE

  #145501 opened Jan 23, 2025

• `torch._inductor.aoti_compile_and_package` fails when using dynamic shapes (PyTorch 2.6.0 RC)

  #145500 opened Jan 23, 2025

• Unexpected behavior of `torch.nn.init.trunc_normal` with bf16 tensors

  #145498 opened Jan 23, 2025

• Add a lint rule to avoid the word `interface` in C++

  #145493 opened Jan 23, 2025

• Cannot print symbolic tensors from C++

  #145491 opened Jan 23, 2025

• OrderedSet is backed by normal Dict, does not check ordering in equality

  #145489 opened Jan 23, 2025

• [custom ops] [2.7 nightly] custom ops with typing.List breaks when importing annotations from future

  #145481 opened Jan 23, 2025

• [XPU] torch 2.7.0.dev20250121+xpu Import Error

  #145477 opened Jan 23, 2025

• torch.backends.cudnn.flags use error when test

  #145473 opened Jan 23, 2025

• Is there a PyTorch version that can work properly on the Thor platform based on the Blackwell architecture?

  #145471 opened Jan 23, 2025

• [inductor][torchbench] Unsupported operator issue when running the torch_multimodal_clip model with batch size 4.

  #145468 opened Jan 23, 2025

• Mark Dynamic does not work for nn module constructor inputs

  #145463 opened Jan 23, 2025

• Incomplete check of LR as a tensor in Optimizer

  #145461 opened Jan 23, 2025

• Flex Attention not support score_mod with gradients

  #145460 opened Jan 23, 2025

• DISABLED test_tensor_subclass_basic (__main__.TestCompiledAutograd)

  #145457 opened Jan 23, 2025

• torch._dynamo.exc.Unsupported: Graph break due to unsupported builtin torch._C._dynamo.eval_frame.set_eval_frame.

  #145456 opened Jan 23, 2025

• [CUDA] Illegal Memory Access with `AdaptiveMaxPool2d`

  #145453 opened Jan 23, 2025

• [Dynamo]while_loop raise an exception

  #145451 opened Jan 23, 2025

• [dynamo] fix graph break on random.random

  #145446 opened Jan 23, 2025

• [dynamo] `random.Random` gives wrong result on second call

  #145445 opened Jan 23, 2025

• seg fault in aot_inductor_package on arm GPU with 2.6.0 RC

  #145441 opened Jan 23, 2025

• Crash in wrapper_benchmark.py with --profile enabled

  #145434 opened Jan 23, 2025

• XPU - UserWarning: Failed to initialize XPU devices. when run on Machine without without Intel GPU Driver

  #145433 opened Jan 23, 2025

• aot inductor intermediate tensor debug printing (setting 2) not working

  #145425 opened Jan 22, 2025

• torch.compile has different numerics for var_mean

  #145401 opened Jan 22, 2025

• [EXPORT AOTI] `aoti_compile_and_package` custom_ops dependecies

  #145394 opened Jan 22, 2025

• create DISABLED issues for specific runner labels

  #145388 opened Jan 22, 2025

• DISABLED test_view_of_slice_cuda (__main__.TestUnbackedSymintsCUDA)

  #145386 opened Jan 22, 2025

• Windows Pytorch compiler crash some version of cl.exe. Fix provided

  #145383 opened Jan 22, 2025

• flaky test issues should close themselves if the test doesn't exist anymore

  #145382 opened Jan 22, 2025

• torch.logit works incorrectly when input < eps after torch.compile

  #145379 opened Jan 22, 2025

• Loading weights using `torch.distributed.checkpoint` leads to large loss values

  #145378 opened Jan 22, 2025

• Inductor autograd raises an error in the second run may because of fx graph cache

  #145377 opened Jan 22, 2025

• distributed.new_group with backend GLOO hangs when distributed.split_group was called before

  #145376 opened Jan 22, 2025

• [XPU] torch.nn.functional.pad brings wrong results with torch.compile on Intel GPU

  #145372 opened Jan 22, 2025

• Set `size` when `is_coalesced` is set in `torch.sparse_coo_tensor()`

  #145371 opened Jan 22, 2025

• The possible error in the pytorch documentation of RNN.

  #145368 opened Jan 22, 2025

• DISABLED test_cache_hot_load_device_cuda_bfloat16_dynamic_False (__main__.TestFxGraphCache)

  #145364 opened Jan 22, 2025

• DISABLED test_cache_load_function_device_cuda_float32_dynamic_False_bundle_triton_True_grad_False (__main__.TestFxGraphCache)

  #145363 opened Jan 22, 2025

• DISABLED test_max_autotune_remote_caching_dynamic_False (__main__.TestMaxAutotuneRemoteCache)

  #145361 opened Jan 22, 2025

• DISABLED test_comprehensive_svd_lowrank_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #145362 opened Jan 22, 2025

• DISABLED test_linear_and_cel_max_autotune (__main__.InplacePaddingTest)

  #145359 opened Jan 22, 2025

• FP8: E5M2: The FP8 E5M2 result is not `inf` when casting a FP32 value larger than max normal value of FP8 E5M2 (57344)

  #145357 opened Jan 22, 2025

• [autograd] inconsistent jvp results

  #145356 opened Jan 22, 2025

• Missing docs for `torch._foreach_copy_`

  #145355 opened Jan 22, 2025

• DISABLED test_extern (__main__.NumBytesMetricTests)

  #145352 opened Jan 22, 2025

• [CUDA] Illegal Memory Access with `ReplicationPad2D`

  #145350 opened Jan 22, 2025

• [CUDA] Illegal Memory Access with `AdaptiveAvgPool2d`

  #145349 opened Jan 22, 2025

• DISABLED test_graph_break_inside_ctx_with_side_effects (__main__.ContextlibContextManagerTests)

  #145346 opened Jan 22, 2025

• DISABLED test_partitioning_with_view (__main__.MinCutPartitioningTests)

  #145345 opened Jan 22, 2025

• DISABLED test_cat (__main__.NumBytesMetricTests)

  #145344 opened Jan 22, 2025

• DISABLED test_partitioning_unremat_bw (__main__.MinCutPartitioningTests)

  #145343 opened Jan 22, 2025

• internal compiler error: in extract_insn when compiling pytorch with xpu with gcc 12

  #145340 opened Jan 22, 2025

• [libTorch] Model initialization on multi-device is slow. It seems to run sequentially in multi-thread

  #145337 opened Jan 22, 2025

• DISABLED test_cache_load_function_device_cuda_float32_dynamic_False_bundle_triton_True_grad_True (__main__.TestFxGraphCache)

  #145336 opened Jan 22, 2025

• DISABLED test_mm_plus_mm (__main__.TestPatternMatcher)

  #145335 opened Jan 22, 2025

• DISABLED test_reorder_peak_memory (__main__.TestOperatorReorderForPeakMemory)

  #145332 opened Jan 22, 2025

• DISABLED test_warn_on_invalid_torch_function_standalone_class (__main__.TestTorchFunctionWarning)

  #145333 opened Jan 22, 2025

• DISABLED test_cache_hot_load_device_cuda_bfloat16_dynamic_False (__main__.AOTAutogradCacheTests)

  #145334 opened Jan 22, 2025

• [dynamo] Save/restore system random state more carefully

  #145329 opened Jan 22, 2025

• [inductor][triton] refactor ASTSource.make_ir integration

  #145326 opened Jan 21, 2025

• Using torch.cond in a model intended for onnx.export(dynamo=True,...) has issues with the functions provided.

  #145300 opened Jan 21, 2025

• [dynamo] `torch.compile` ICE on using a sourceless unspecialized NN module as branching condition

  #145284 opened Jan 21, 2025

• Torch Compile edge case with != versus is not

  #145277 opened Jan 21, 2025

• AttributeError: '_OpNamespace' 'aten' object has no attribute 'momentum'

  #145274 opened Jan 21, 2025

• `torch.ops.aten.embedding_dense_backward` Crashes with Out-of-Bounds Indices On CPU

  #145267 opened Jan 21, 2025

• [Pipelining] Problem using `torch.distributed.pipelining` on `Gemma2ForCausalLM`

  #145263 opened Jan 21, 2025

• Missing create_graph arguments in torch.func apis

  #145262 opened Jan 21, 2025

• Custom symbolic functions for ONNX export with None args causes SEGFAULT

  #145261 opened Jan 21, 2025

• No Range Check for `storage_offset` in `as_strided` Function

  #145259 opened Jan 21, 2025

• Missing Length Check for `reflection_pad3d` `padding` Argument

  #145258 opened Jan 21, 2025

• PyObject preservation does not prevent weakrefs being cleared by Python garbage collector

  #145253 opened Jan 21, 2025

• [Break XPU] device type in test_aot_inductor.py is not passed correctly to cpp_builder.

  #145247 opened Jan 21, 2025

• Expose configurable path instead of using fixed path in the inductor module for serialized pattern generation

  #145242 opened Jan 21, 2025

• Not using set_num_threads results in very slow .all()

  #145233 opened Jan 21, 2025

• Flaky Dynamo test: TestAutograd.test_gradcheck_nondeterministic

  #145231 opened Jan 21, 2025

• The `sympy` dependency spec for pytorch on PyPi wheel is still unchanged.

  #145225 opened Jan 20, 2025

• Regression in the compilation of the torch.all operation in PyTorch version 2.6.0 compared to 2.5.1

  #145220 opened Jan 20, 2025

• `torch.compile` may produce wrong result with `Linear+MaxPool2d+BatchNorm2d`.

  #145219 opened Jan 20, 2025

• getting different results when adding `torch.Tensor` or python number to a DTensor - Is that expected?

  #145218 opened Jan 20, 2025

• DISABLED test_cache_load_function_device_cuda_bfloat16_dynamic_False_bundle_triton_True_grad_False (__main__.TestFxGraphCache)

  #145217 opened Jan 20, 2025

• [ARM] - test_quantized_module.py test_lstm_api fails on Aarch64

  #145216 opened Jan 20, 2025

• Nested tensor support for pointwise matrix multiplication of nested tensor and normal tensor

  #145214 opened Jan 20, 2025

• Significant precision error from torch.compile

  #145213 opened Jan 20, 2025

• DISABLED test_remote_cache_load_function_device_cuda_float32_dynamic_False_bundle_triton_False (__main__.TestFxGraphCache)

  #145212 opened Jan 20, 2025

• DISABLED test_aoti (__main__.TestMemoryPlanning)

  #145211 opened Jan 20, 2025

• DISABLED test_reorder_peak_memory_lpmf (__main__.TestOperatorReorderForPeakMemory)

  #145210 opened Jan 20, 2025

• Some FlexAttention learned bias bugs/limitations

  #145208 opened Jan 20, 2025

• Indexed ^= (XOR in-place) operation doesn't work as expected on MPS backend

  #145203 opened Jan 20, 2025

• DISABLED test_reuse_kernel_cuda (__main__.AOTInductorTestABICompatibleGpu)

  #145193 opened Jan 20, 2025

• DISABLED test_mixed_mm (__main__.TestPatternMatcher)

  #145192 opened Jan 20, 2025

• DISABLED test_remote_cache_load_function_device_cuda_float32_dynamic_False_bundle_triton_True (__main__.TestFxGraphCache)

  #145191 opened Jan 20, 2025

• DISABLED test_remote_cache_load_function_device_cuda_bfloat16_dynamic_False_bundle_triton_True (__main__.TestFxGraphCache)

  #145190 opened Jan 20, 2025

• DISABLED test_slice_scatter_reinplace_cuda (__main__.GPUTests)

  #145189 opened Jan 20, 2025

• DISABLED test_sdpa_rewriter_12_cuda (__main__.SDPAPatternRewriterCudaDynamicTests)

  #145188 opened Jan 20, 2025

• DISABLED test_sdpa_rewriter_12_cuda (__main__.SDPAPatternRewriterCudaTests)

  #145187 opened Jan 20, 2025

• DISABLED test_mm_concat_cuda (__main__.FreezingGpuTests)

  #145186 opened Jan 20, 2025

• DISABLED test_aoti_eager_cache_hit_dynamic_shapes_cuda (__main__.DynamicShapesCodegenGPUTests)

  #145184 opened Jan 20, 2025

• DISABLED test_reorder_peak_memory_dfs (__main__.TestOperatorReorderForPeakMemory)

  #145183 opened Jan 20, 2025

• DISABLED test_remote_cache_load_function_device_cuda_bfloat16_dynamic_False_bundle_triton_False (__main__.TestFxGraphCache)

  #145182 opened Jan 20, 2025

• DISABLED test_cache_load_function_device_cuda_bfloat16_dynamic_False_bundle_triton_True_grad_True (__main__.TestFxGraphCache)

  #145181 opened Jan 20, 2025

• CUDA initialization error with vLLM 0.5.4 and PyTorch 2.4.0+cu121

  #145170 opened Jan 19, 2025

• empty_cache does not work for CUDAPluggableAllocator + MemPool

  #145168 opened Jan 19, 2025

• Pytorch matmul for nested 4D tensors in jagged layout doesn't work

  #145158 opened Jan 18, 2025

• The latest PyTorch XPU wheel 2.7.0.dev20250117+xpu does not work on Windows

  #145155 opened Jan 18, 2025

• Driver Allocated Memory grows unrestricted when using torch.unique on MPS device

  #145151 opened Jan 18, 2025

• [RFC] Improve performance for layer_norm op for cuda with revectorized

  #145145 opened Jan 18, 2025

• Please add fp16 to MPS devices.

  #145144 opened Jan 18, 2025

• Bracket indexing not working

  #145143 opened Jan 18, 2025

• Obey sm_carveout (limit on number of SMs) in inductor persistent kernel

  #145115 opened Jan 17, 2025

• torch._C._IncludeDispatchKeyGuard is very broken?

  #145108 opened Jan 17, 2025

• `torch.onnx.export` (dynamo=False) fails with uninformative error when exporting `apply_rotary_pos_emb`/`repeat_interleave`

  #145100 opened Jan 17, 2025

• Tracking issue: Incorrect Meta Strides / Turn On PyDispatcher in FakeTensor Mode

  #145094 opened Jan 17, 2025

• Inductor aten.clone lowering ignores Conjugate and Negative dispatch keys

  #145093 opened Jan 17, 2025

• [torchbench] torch._dynamo.exc.Unsupported: Graph break due to unsupported builtin None.morphologyEx

  #145088 opened Jan 17, 2025

• partitioner hangs for some long chains of ops with many users

  #145081 opened Jan 17, 2025

• list comprehension in SkipFiles are always skipped with no way to override

  #145079 opened Jan 17, 2025

• Negative values in stride causing error in `avg_pool2d` (on both CPU and CUDA)

  #145077 opened Jan 17, 2025

• AssertionError: increase TRITON_MAX_BLOCK['X'] to 4096 Again!

  #145074 opened Jan 17, 2025

• Segmentation fault when passing an empty tensor to `_local_scalar_dense`

  #145072 opened Jan 17, 2025

• Illegal memory access and segmentation fault due to large `storage_offset` in `as_strided`

  #145071 opened Jan 17, 2025

• Segment fault on CPU and IndexError on CUDA for `_adaptive_avg_pool2d_backward`

  #145070 opened Jan 17, 2025

• DISABLED test_sparse_add_cuda_complex64 (__main__.TestSparseCSRCUDA)

  #145069 opened Jan 17, 2025

• SIGSEGV error when passing a 0-sized tensor to `_local_scalar_dense`

  #145066 opened Jan 17, 2025

• SIGFPE error when passing very large kernel_size to `avg_pool1d`

  #145065 opened Jan 17, 2025

430 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [inductor] Kernel memory analysis for use in heuristics

  #142026 commented on Jan 24, 2025 • 19 new comments

• [Inductor] Unifiy Low Precision FP Legalization for to_dtype_bitcast & constant

  #144646 commented on Jan 24, 2025 • 14 new comments

• [PGNCCL] Add an API to get the status/error code at the PG level

  #144498 commented on Jan 24, 2025 • 12 new comments

• Use `typing.IO[bytes]` instead of `io.BytesIO` in annotations

  #144994 commented on Jan 24, 2025 • 11 new comments

• Update test_c10d_object_collectives.py with DistributedTestBase class

  #145056 commented on Jan 24, 2025 • 10 new comments

• [inductor] Add type annotations to _inductor/utils.py

  #144108 commented on Jan 23, 2025 • 9 new comments

• [inductor] [cpp] Support vectorization for score and mask in FlexAttention CPU

  #143638 commented on Jan 24, 2025 • 8 new comments

• pickler for GraphModule

  #141659 commented on Jan 24, 2025 • 8 new comments

• Add noncontiguous OpInfo tests for MPS

  #142202 commented on Jan 24, 2025 • 6 new comments

• [dcp] Add ZStandard transformer

  #143360 commented on Jan 24, 2025 • 6 new comments

• serde unbacked bindings

  #144894 commented on Jan 23, 2025 • 6 new comments

• Parallelize epilogue/prologue benchmarking

  #143408 commented on Jan 21, 2025 • 5 new comments

• [torch.special] Adding betainc, betaincc, betaincinv, betainccinv, betaln and beta with backward operation

  #132135 commented on Jan 23, 2025 • 4 new comments

• Fix Throughputbenchmark issue

  #144669 commented on Jan 24, 2025 • 4 new comments

• Align CPU behavior with CUDA for `ConvTranspose` when `out_channels=0`

  #142859 commented on Jan 24, 2025 • 4 new comments

• Introduce the public API for all_gather_scaled_matmul

  #141053 commented on Jan 20, 2025 • 4 new comments

• [Intel GPU] qconv_pointwise.binary XPU support

  #135189 commented on Jan 23, 2025 • 4 new comments

• Add prepacking for linear weights

  #139387 commented on Jan 23, 2025 • 4 new comments

• [CI] enable operator benchmark on CPU

  #143733 commented on Jan 23, 2025 • 4 new comments

• [Intel GPU] qlinear at XPU backend

  #133307 commented on Jan 23, 2025 • 3 new comments

• Add test cases of fp8 datatypes in pt2e

  #144388 commented on Jan 24, 2025 • 3 new comments

• [Intel CPU] Fix issue #143483.

  #144854 commented on Jan 24, 2025 • 3 new comments

• [Dynamo][autograd.Function] Relax backward speculation strict mode

  #142830 commented on Jan 22, 2025 • 3 new comments

• [Intel CPU] Fix issue #143484.

  #144950 commented on Jan 24, 2025 • 2 new comments

• OpenReg: fix issue of pin_memory

  #145046 commented on Jan 24, 2025 • 2 new comments

• [Inductor] Fix starvation issue when threads attempt to acquire write…

  #144460 commented on Jan 23, 2025 • 2 new comments

• add fp8 scaled_mm for XPU

  #140972 commented on Jan 20, 2025 • 2 new comments

• Add overloads to diagonal docs

  #144214 commented on Jan 22, 2025 • 2 new comments

• [compiled autograd] Always proxy autograd.Function nodes; handle AOT backwards

  #143405 commented on Jan 24, 2025 • 2 new comments

• [Do NOT merge] Enable inductor-periodic testing for ROCm on MI300

  #144594 commented on Jan 24, 2025 • 2 new comments

• Add option to limit number of SMs used by matmul kernels

  #144974 commented on Jan 22, 2025 • 2 new comments

• [inductor] Fix an aten.squeeze stride computation issue

  #143683 commented on Jan 23, 2025 • 1 new comment

• Implement cuda graphs implementation of torch.cond and torch.while_loop

  #140979 commented on Jan 23, 2025 • 1 new comment

• inductor_config_logging: Don't drop keys

  #144700 commented on Jan 24, 2025 • 1 new comment

• Avoid data-dependent errors in NJT tests via capture_scalar_outputs=True

  #144588 commented on Jan 21, 2025 • 1 new comment

• [compiled autograd] support Tensor Subclasses in AOTBackward

  #144115 commented on Jan 24, 2025 • 1 new comment

• [BE] Add stride check in `torch.max_pool1d()`

  #144023 commented on Jan 23, 2025 • 1 new comment

• Exclude upsample_bilinear2d.vec from default core ATen decomposition table

  #141791 commented on Jan 18, 2025 • 1 new comment

• Fix a number of flexattention issues (cse, cudagraph, etc.)

  #145059 commented on Jan 22, 2025 • 1 new comment

• [Inductor changes] Invoke Quant

  #139102 commented on Jan 17, 2025 • 1 new comment

• [Intel CPU] Fix issue #143482.

  #144760 commented on Jan 24, 2025 • 1 new comment

• Fix fft jit ops cpu

  #143894 commented on Jan 21, 2025 • 1 new comment

• Enable CPP Extension Open Registration tests on Arm

  #144774 commented on Jan 21, 2025 • 1 new comment

• Update ck

  #144799 commented on Jan 24, 2025 • 1 new comment

• Support narrow() on batch dim for NJT

  #142063 commented on Jan 17, 2025 • 1 new comment

• Fix flash attention seed/offset overflow when seed/offset larger than int64

  #144844 commented on Jan 23, 2025 • 1 new comment

• Enable SVE ACLE implementation for tanH Aten op for FP32 dType.

  #143741 commented on Jan 21, 2025 • 1 new comment

• Support remaining *_like factory functions for NJT

  #144889 commented on Jan 22, 2025 • 1 new comment

• [Inductor UT] Refactor FlexAttention UT and add CPU tests

  #144953 commented on Jan 22, 2025 • 1 new comment

• [ATen][CUDA] Implement 128 bit vectorization

  #141959 commented on Jan 24, 2025 • 0 new comments

• Add AOT inductor support for _scaled_mm for CPU

  #141961 commented on Jan 21, 2025 • 0 new comments

• Permute test

  #140261 commented on Jan 19, 2025 • 0 new comments

• support condition branch in ao debug handler

  #141302 commented on Jan 21, 2025 • 0 new comments

• [sympy] Make solve of Mul for Eq int replacement friendly

  #141347 commented on Jan 23, 2025 • 0 new comments

• Save models in OCI registry

  #141354 commented on Jan 21, 2025 • 0 new comments

• Use std::string_view in torchgen

  #141735 commented on Jan 17, 2025 • 0 new comments

• [Don't merge] test only

  #141468 commented on Jan 24, 2025 • 0 new comments

• [Store log]Test log struct

  #141439 commented on Jan 23, 2025 • 0 new comments

• WIP delta graph logging

  #141416 commented on Jan 23, 2025 • 0 new comments

• [hop] fix unbacked_bindings meta for while_loop

  #143559 commented on Jan 24, 2025 • 0 new comments

• Fix FSDP hanging

  #143540 commented on Jan 23, 2025 • 0 new comments

• [hop][inductor] track the dependency on unbacked symbols correctly with constant_args for hops

  #143456 commented on Jan 24, 2025 • 0 new comments

• [compiled autograd] stop specializing on metadata during initial trace

  #143417 commented on Jan 24, 2025 • 0 new comments

• [compiled autograd] Proxy nodes for user-defined C++ torch::autograd::Function

  #143387 commented on Jan 24, 2025 • 0 new comments

• [EXPERIMENTAL][dynamo] Turn on `inline_inbuilt_nn_modules` for fbcode

  #143313 commented on Jan 17, 2025 • 0 new comments

• [compiled autograd] Proxy a node for CopyBackwards into the graph

  #143304 commented on Jan 24, 2025 • 0 new comments

• [compiled autograd] Proxy opaque nodes for built-in autograd nodes

  #143296 commented on Jan 24, 2025 • 0 new comments

• Set proper `LD_LIBRARY_PATH` on Linux in nightly venv in nightly pull tool

  #143262 commented on Jan 17, 2025 • 0 new comments

• [TorchGen] Simplify argumenttype_type

  #143254 commented on Jan 20, 2025 • 0 new comments

• [Testing only] Add python cycle detection

  #143204 commented on Jan 22, 2025 • 0 new comments

• Unify use of `enableCollectiveHashDebug_` and trivial updates

  #142865 commented on Jan 18, 2025 • 0 new comments

• Fix RMSNorm epsilon value type for BF16 or FP16

  #142848 commented on Jan 22, 2025 • 0 new comments

• Set `enable_faithful_generator_behavior` flag to True

  #142513 commented on Jan 23, 2025 • 0 new comments

• parallelize sort

  #142391 commented on Jan 23, 2025 • 0 new comments

• Fix type annotation of `Linear.bias`

  #142326 commented on Jan 23, 2025 • 0 new comments

• [inductor] Decide cooperative RSPLIT with same algorithm as split reductions

  #142295 commented on Jan 24, 2025 • 0 new comments

• [scan] Refactoring of input checking and dynamo invocation

  #142125 commented on Jan 23, 2025 • 0 new comments

• support condition branch in ao debug handler

  #140256 commented on Jan 19, 2025 • 0 new comments

• [Environment Variable][7/N] Use thread-safe getenv functions

  #140211 commented on Jan 21, 2025 • 0 new comments

• [Environment Variable][6/N] Use thread-safe getenv functions

  #140200 commented on Jan 19, 2025 • 0 new comments

• Cleanup stale Dynamo Feature Flags

  #140147 commented on Jan 18, 2025 • 0 new comments

• [export] Add custom op profiles and generate meta kernel

  #140048 commented on Jan 21, 2025 • 0 new comments

• [associative_scan] Lifted arguments

  #140043 commented on Jan 23, 2025 • 0 new comments

• Add torch._scaled_mm for CPU

  #139975 commented on Jan 22, 2025 • 0 new comments

• [Don't Review] Test CI

  #139971 commented on Jan 21, 2025 • 0 new comments

• [associative_scan] scan dim handling in user-facing associative_scan()

  #139864 commented on Jan 24, 2025 • 0 new comments

• Add Windows Arm64 Nightly Builds

  #139760 commented on Jan 23, 2025 • 0 new comments

• Add support for loading model `state_dict()`in C++ which are OrderedDicts

  #139750 commented on Jan 21, 2025 • 0 new comments

• [cuDNN] Add an option to force cuDNN usage (incl. SDPA)

  #139699 commented on Jan 23, 2025 • 0 new comments

• [export] Serialize draft export report

  #139384 commented on Jan 18, 2025 • 0 new comments

• Allow BUILD for classes with types built via functions/classes allowed for REDUCE (i.e. not GLOBALs in checkpoint)

  #139302 commented on Jan 17, 2025 • 0 new comments

• [do not review] saving things for NJT metadata cache

  #139247 commented on Jan 20, 2025 • 0 new comments

• Use the device interface for detecting Triton availability

  #139171 commented on Jan 22, 2025 • 0 new comments

• Unify shallow_copy_and_detach overloads by passing c10::VariableVersion

  #138999 commented on Jan 24, 2025 • 0 new comments

• [c10d] Remove ProcessGroupGloo + CUDA tests

  #138998 commented on Jan 23, 2025 • 0 new comments

• Tensor .cuda() very slow with specific array sizes

  #138964 commented on Jan 21, 2025 • 0 new comments

• Fix bug of torch.nn.functional.kl_div when broadcast happened

  #138810 commented on Jan 23, 2025 • 0 new comments

• [Docs] Optimize parameter description to declare allowed type (3/N)

  #138798 commented on Jan 19, 2025 • 0 new comments

• Switch back to the default checkout action

  #138739 commented on Jan 22, 2025 • 0 new comments

• inductor `full_like` decompositions give incorrect strides

  #144699 commented on Jan 21, 2025 • 0 new comments

• Expose torch.autograd.graph.is_backward_executing

  #141276 commented on Jan 21, 2025 • 0 new comments

• Fix InductorLower when attribute is shape ()

  #141226 commented on Jan 21, 2025 • 0 new comments

• [Inductor] be able to disable cache for test

  #141195 commented on Jan 24, 2025 • 0 new comments

• [CI] Reduce distributed test timeout to 60s

  #141168 commented on Jan 21, 2025 • 0 new comments

• Specific attribute for device DTensor RNG support indication.

  #141141 commented on Jan 20, 2025 • 0 new comments

• [WIP][Inductor XPU] Support mkldnn fusion in freezing for XPU.

  #141096 commented on Jan 19, 2025 • 0 new comments

• [pytree] Save namedtuple fields

  #141084 commented on Jan 19, 2025 • 0 new comments

• Suport generators

  #141055 commented on Jan 23, 2025 • 0 new comments

• [EZ] Remove TODO because it already works

  #141002 commented on Jan 18, 2025 • 0 new comments

• dynamo: Support custom attributes in tensor subclasses

  #140978 commented on Jan 21, 2025 • 0 new comments

• Add option to split Linear gates for Quantizable LSTM into separate ops

  #140868 commented on Jan 21, 2025 • 0 new comments

• [aoti] Avoid DCE unbacked symint node

  #140858 commented on Jan 18, 2025 • 0 new comments

• Fix TORCH_CUDA_ARCH_LIST for SBSA+CUDA build

  #140844 commented on Jan 18, 2025 • 0 new comments

• Enable CUDA 12.6 OSS CI

  #140793 commented on Jan 24, 2025 • 0 new comments

• Enable C++ dynamic shape guards by default

  #140756 commented on Jan 23, 2025 • 0 new comments

• fix torchrec on inductor

  #140747 commented on Jan 18, 2025 • 0 new comments

• [Intel GPU] Enable fp64 GEMM

  #140677 commented on Jan 21, 2025 • 0 new comments

• Add boolean conversion support for SymNodeVariable

  #140621 commented on Jan 17, 2025 • 0 new comments

• add NHWC support to GroupNorm backward pass and optimize NHWC GroupNorm kernels

  #140440 commented on Jan 18, 2025 • 0 new comments

• Auto SAC - Automated SAC (Selective Activation Checkpointing) Policy Construction and Wrapping

  #140410 commented on Jan 22, 2025 • 0 new comments

• remove redundant assign

  #140399 commented on Jan 24, 2025 • 0 new comments

• Fix inconsistent results from integral linspace on MPS

  #140371 commented on Jan 18, 2025 • 0 new comments

• [WIP] SimpleFSDP prototype frontend changes

  #140360 commented on Jan 21, 2025 • 0 new comments

• [scan] Refactored testcases

  #140321 commented on Jan 18, 2025 • 0 new comments

• [Inductor] optimize welford reduction

  #145061 commented on Jan 24, 2025 • 0 new comments

• optimize the decomposition of aten.native_group_norm

  #144733 commented on Jan 21, 2025 • 0 new comments

• [Intel GPU] Support SparseCsrXPU codegen

  #144722 commented on Jan 18, 2025 • 0 new comments

• functional compiled autograd

  #144707 commented on Jan 24, 2025 • 0 new comments

• Output of nonzero is transposed, fix fake tensor

  #144695 commented on Jan 23, 2025 • 0 new comments

• Generalize poison fork logic for each device backend

  #144664 commented on Jan 17, 2025 • 0 new comments

• Fix torch.logsumexp dim description

  #144661 commented on Jan 22, 2025 • 0 new comments

• [MPS] lu factor ex implementation

  #144651 commented on Jan 21, 2025 • 0 new comments

• remove Windows XPU build workaround.

  #144644 commented on Jan 23, 2025 • 0 new comments

• [Not4Land] test `optree` version compatibility

  #144642 commented on Jan 17, 2025 • 0 new comments

• [inductor] Add features to docstring_linter (see #142496)

  #144620 commented on Jan 23, 2025 • 0 new comments

• Collect packages with importlib in collect_env

  #144616 commented on Jan 24, 2025 • 0 new comments

• [device_mesh] improve device selection logic

  #144600 commented on Jan 18, 2025 • 0 new comments

• Fix DTensorTestBase to barrier with device ids

  #144599 commented on Jan 18, 2025 • 0 new comments

• Implemented dropout usage for RNN with MIOpen backend

  #144572 commented on Jan 23, 2025 • 0 new comments

• [BE][CI] bump `ruff` to 0.9.0: string quote styles

  #144569 commented on Jan 24, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `torch.{distributed,distributions}` to `ruff format`

  #144547 commented on Jan 18, 2025 • 0 new comments

• [BE][CI] bump `ruff` to 0.9.2: multiline `assert` statements

  #144546 commented on Jan 24, 2025 • 0 new comments

• Fix clang-tidy warnings of performance from uncovered files

  #144542 commented on Jan 21, 2025 • 0 new comments

• Save integral tensor data for ET

  #144508 commented on Jan 24, 2025 • 0 new comments

• blocked benchmarking to avoid queue limit

  #144507 commented on Jan 18, 2025 • 0 new comments

• better overlapping of sleep and memory warmup

  #144505 commented on Jan 18, 2025 • 0 new comments

• add most basic event packing

  #144501 commented on Jan 18, 2025 • 0 new comments

• patch for block-wise quantization + pt2e

  #144492 commented on Jan 17, 2025 • 0 new comments

• prov logging

  #145047 commented on Jan 20, 2025 • 0 new comments

• [Easy] Replace paper description with link to make a concise description.

  #145031 commented on Jan 23, 2025 • 0 new comments

• Made partitioning more(?) deterministic

  #145024 commented on Jan 23, 2025 • 0 new comments

• [CD] Annotate linux/arm64 cuda wheels with consistent nvidia dependencies

  #145021 commented on Jan 21, 2025 • 0 new comments

• [dynamo/export] call local_scalar_dense when full() value is scalar tensor

  #144999 commented on Jan 24, 2025 • 0 new comments

• Introduce new template heuristic for triton autotune configs

  #144985 commented on Jan 24, 2025 • 0 new comments

• [inductor] Fix for pattern file contains 'getitem' fails during impor…

  #144980 commented on Jan 23, 2025 • 0 new comments

• [test] fix unit test

  #144977 commented on Jan 17, 2025 • 0 new comments

• Let `tensor_a.new_tensor()` be on `tensor_a.device` by default

  #144958 commented on Jan 22, 2025 • 0 new comments

• [Dynamo] Allow `format()` to handle int

  #144956 commented on Jan 24, 2025 • 0 new comments

• Replacing explicit backend search with api call

  #144944 commented on Jan 24, 2025 • 0 new comments

• [ROCm][TunableOp] Improve selection criteria for fastest solution

  #144942 commented on Jan 23, 2025 • 0 new comments

• Binary upload checksum

  #144887 commented on Jan 23, 2025 • 0 new comments

• update guard_size_oblivious comment

  #144880 commented on Jan 22, 2025 • 0 new comments

• [64-bit] Int64 casting for UpSampleNearest3D

  #144865 commented on Jan 23, 2025 • 0 new comments

• WIP pp_cp test

  #144834 commented on Jan 18, 2025 • 0 new comments

• Added swizzle searching, disabled fp16 accum, and enabled ping-pong for cutlass

  #144829 commented on Jan 24, 2025 • 0 new comments

• Test of RST to MD

  #144804 commented on Jan 22, 2025 • 0 new comments

• [c10d][NCCL] Implement ncclCommInitRankScalable (merging #136789)

  #144794 commented on Jan 24, 2025 • 0 new comments

• [caffe2] Use the manifold cache backend as the default

  #144773 commented on Jan 24, 2025 • 0 new comments

• Unconditional dependency on setuptools

  #144763 commented on Jan 18, 2025 • 0 new comments

• [Reopen] [Intel GPU] Set higher tolerance for some models only on XPU Device

  #144756 commented on Jan 21, 2025 • 0 new comments

• [cherry-pick][dtensor] expose the __create_chunk_list__ in the doc (#144100)

  #144741 commented on Jan 18, 2025 • 0 new comments

• c10::string_view -> std::string_view in torchgen

  #144177 commented on Jan 18, 2025 • 0 new comments

• [Submodule] Turning flash-attention integration into 3rd party submod

  #144120 commented on Jan 23, 2025 • 0 new comments

• Avoid overflow in vector_norm for scalar input

  #144073 commented on Jan 23, 2025 • 0 new comments

• Native channel shuffle floating point exception

  #144010 commented on Jan 18, 2025 • 0 new comments

• cpp_wrapper: Precompile device-specific header files

  #144002 commented on Jan 22, 2025 • 0 new comments

• [poc][not-ready-for-review] visualize dynamic shapes shape env mutations over time

  #143961 commented on Jan 20, 2025 • 0 new comments

• Add ability to skip compute capability checks for Triton

  #143956 commented on Jan 17, 2025 • 0 new comments

• Fix an unnecessary CPU to GPU copy within flex_attention

  #143928 commented on Jan 23, 2025 • 0 new comments

• Using acc_t for log_softmax

  #143896 commented on Jan 21, 2025 • 0 new comments

• Add option to serialization config to reduce random reads from get_record_offset when loading with mmap=True

  #143880 commented on Jan 24, 2025 • 0 new comments

• Remove lexicographical sorting of storage keys in torch.save

  #143879 commented on Jan 24, 2025 • 0 new comments

• [FlexAttention] make bm creation cuda-graphable

  #143872 commented on Jan 23, 2025 • 0 new comments

• [1/N]Add Intel GPU Support to Torch Test Cases

  #143833 commented on Jan 17, 2025 • 0 new comments

• [inductor] Used fixed configs for contiguous reductions

  #143812 commented on Jan 24, 2025 • 0 new comments

• Enable clang-tidy on torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp

  #143806 commented on Jan 24, 2025 • 0 new comments

• [don't merge] use vs2022 build windows cpu wheel.

  #143791 commented on Jan 23, 2025 • 0 new comments

• [Intel GPU] Avoid copy when the input of Matmul is broadcasted

  #143784 commented on Jan 17, 2025 • 0 new comments

• Modify the tolerance level in TIMM benchmark for XPU PreCI

  #143739 commented on Jan 23, 2025 • 0 new comments

• nn.MultiheadAttention string representation

  #143724 commented on Jan 23, 2025 • 0 new comments

• Getattr access for subclasses in pre-dispatch

  #143671 commented on Jan 23, 2025 • 0 new comments

• Fix the build errors in ONEDNN+BLIS Path

  #143642 commented on Jan 17, 2025 • 0 new comments

• [triton pin 3.2] Cherry pick additional device context fix

  #143622 commented on Jan 23, 2025 • 0 new comments

• Add the max_autotune tests in the periodic jobs.

  #143560 commented on Jan 21, 2025 • 0 new comments

• Introduce cache clearing APIs for the lazy graph executor

  #144489 commented on Jan 18, 2025 • 0 new comments

• Support Swiglu for Module and functional

  #144465 commented on Jan 24, 2025 • 0 new comments

• improve WOQ first token performance on CPU

  #144463 commented on Jan 23, 2025 • 0 new comments

• Support negative values for fill with uint tensors

  #144458 commented on Jan 21, 2025 • 0 new comments

• [CUDA][cuBLAS] Add fp16 accumulate option to cuBLAS/cuBLASLt

  #144441 commented on Jan 24, 2025 • 0 new comments

• Implement `generator.throw(exception)`

  #144424 commented on Jan 23, 2025 • 0 new comments

• Implement `generator.close()`

  #144423 commented on Jan 23, 2025 • 0 new comments

• Implement `generator.send(..)`

  #144422 commented on Jan 23, 2025 • 0 new comments

• Implement `generator.__iter__()`

  #144421 commented on Jan 23, 2025 • 0 new comments

• Add `CLEANUP_THROW` bytecode

  #144420 commented on Jan 23, 2025 • 0 new comments

• fix a bug for constant_pad_nd

  #144394 commented on Jan 23, 2025 • 0 new comments

• `torch.linalg.solve`: doc update on dealing with rank-deficient systems which admit a solution

  #144390 commented on Jan 19, 2025 • 0 new comments

• Fix lowering to inductor IR for triton CPU

  #144389 commented on Jan 24, 2025 • 0 new comments

• [Intel GPU] fix memory leak in deconv backward

  #144385 commented on Jan 17, 2025 • 0 new comments

• Filter out iGPU if dGPU is found on XPU

  #144378 commented on Jan 24, 2025 • 0 new comments

• [Don't Merge] Fix poision child process issue when call getAccelerator()

  #144368 commented on Jan 22, 2025 • 0 new comments

• implement LazyInductorBenchmarker

  #144365 commented on Jan 18, 2025 • 0 new comments

• Improve torchrun documentation

  #144354 commented on Jan 24, 2025 • 0 new comments

• implement pruning for GroupedInductorBenchmarker

  #144353 commented on Jan 18, 2025 • 0 new comments

• codecache.py: Utilize precompiled headers for CPP python bindings

  #144349 commented on Jan 22, 2025 • 0 new comments

• codecache: Remove cpp_prefix.h duplication per build, then precompile it

  #144293 commented on Jan 22, 2025 • 0 new comments

• [inductor] Only call triton.compile in worker processes

  #144288 commented on Jan 22, 2025 • 0 new comments

• Cholesky mps implementation

  #144193 commented on Jan 23, 2025 • 0 new comments

• aot_inductor TIMM convit_base inference regression on dashboard

  #144772 commented on Jan 21, 2025 • 0 new comments

• TorchBench mobilenet_v2 cudagraphs_freezing inference regression

  #144891 commented on Jan 21, 2025 • 0 new comments

• TIMM cudagraphs_freezing inference regression

  #144888 commented on Jan 21, 2025 • 0 new comments

• torch.export fails for whisper tiny

  #144906 commented on Jan 21, 2025 • 0 new comments

• Dynamo is not thread safe

  #118260 commented on Jan 21, 2025 • 0 new comments

• torch.accelerator.is_available() raise RuntimeError if no available CUDA/XPU devices

  #144567 commented on Jan 21, 2025 • 0 new comments

• `torch.device(0)` makes CUDA init fail in subprocess since `2.5.0`

  #144152 commented on Jan 21, 2025 • 0 new comments

• Make flex_attention work if `score_mod`'s output doesn't require gradients at all

  #145050 commented on Jan 21, 2025 • 0 new comments

• [torchbench] stable_diffusion_unet compilation failure

  #144991 commented on Jan 21, 2025 • 0 new comments

• CUDAGraph outputs will be overwritten by a subsequent run?

  #144961 commented on Jan 21, 2025 • 0 new comments

• Inconsistency of `tensor.new_tensor(data)` between eager and dynamo

  #144957 commented on Jan 21, 2025 • 0 new comments

• UserWarning: cuDNN SDPA backward got grad_output.strides() != output.strides()

  #144913 commented on Jan 21, 2025 • 0 new comments

• Issue: Illegal Memory Access in Backward Pass of `scaled_dot_product_attention` with Custom Attention Mask

  #145040 commented on Jan 21, 2025 • 0 new comments

• [compiled autograd] It would be nice if the compiiled autograd graph was actually runnable

  #144982 commented on Jan 21, 2025 • 0 new comments

• ONNX: Wrong output shape for ceil_mode Pooling

  #71549 commented on Jan 21, 2025 • 0 new comments

• Numerical error when using torch.nn.functional.pad with a large array on MPS device

  #121961 commented on Jan 21, 2025 • 0 new comments

• [inductor][cpu]float32 dynamic shape maml_omniglot performance regression in 2025-01-13 nightly release

  #144937 commented on Jan 21, 2025 • 0 new comments

• [inductor][cpu]amp fp16 llama dynamic shape cpp wrapper performance regression in 2025-01-07 nightly release

  #144932 commented on Jan 21, 2025 • 0 new comments

• [inductor][cpu] fused attention Inductor tests fails with an error " name 'getitem' is not defined "

  #144674 commented on Jan 21, 2025 • 0 new comments

• [Inductor] Unify the data type propagation between Triton and CPP Backend

  #144246 commented on Jan 21, 2025 • 0 new comments

• Add API to detect if activation checkpointing is enabled in the current region or not

  #144928 commented on Jan 21, 2025 • 0 new comments

• Torch compile cache

  #144859 commented on Jan 21, 2025 • 0 new comments

• _pickle.UnpicklingError: pickle data was truncated - Windows multiprocessing during training

  #69611 commented on Jan 21, 2025 • 0 new comments

• Multiple CPU processes using same GPU model for inference

  #16943 commented on Jan 21, 2025 • 0 new comments

• [RFC] Disable CMake find_library(libm) on Windows, and solve libm conflict to MSVC runtime lib(ucrt.lib).

  #141946 commented on Jan 21, 2025 • 0 new comments

• DISABLED test_method_overloading (__main__.TestScript)

  #131104 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_mixed_mm_epi_works (__main__.TestPatternMatcher)

  #126489 commented on Jan 22, 2025 • 0 new comments

• DISABLED test_batch_linear_post_grad_fusion (__main__.TestPostGradBatchLinearFusion)

  #120280 commented on Jan 22, 2025 • 0 new comments

• custom gradient for int8

  #129889 commented on Jan 22, 2025 • 0 new comments

• torch.compile() within TorchDispatchMode always causes an unknown guard failure.

  #144787 commented on Jan 22, 2025 • 0 new comments

• DISABLED test_sdpa_mask_fp16_L6_S17_NH23_HS121 (__main__.TestSDPA)

  #138905 commented on Jan 22, 2025 • 0 new comments

• RFC: Dynamically Quantized 4 bit matmul API and usage

  #143289 commented on Jan 22, 2025 • 0 new comments

• Code fails with "Expected curr_block->next == nullptr to be true, but got false"

  #140419 commented on Jan 22, 2025 • 0 new comments

• DISABLED test_mixed_mm_bad_cases (__main__.TestPatternMatcher)

  #128487 commented on Jan 22, 2025 • 0 new comments

• ModuleNotFoundError: No module named 'torch.privateuseone'

  #144955 commented on Jan 22, 2025 • 0 new comments

• DISABLED test_comprehensive_argsort_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #131158 commented on Jan 22, 2025 • 0 new comments

• Adding Infiniband to RDZV Backend for optimal torch run training

  #144779 commented on Jan 21, 2025 • 0 new comments

• user-defined triton kernels + inductor stride re-ordering can lead to silent incorrectness

  #130243 commented on Jan 21, 2025 • 0 new comments

• asynchronous copies from accelerator to cpu: what should be the expected behaviour?

  #140296 commented on Jan 21, 2025 • 0 new comments

• `torch._foreach_mul` does not support autograd

  #144580 commented on Jan 21, 2025 • 0 new comments

• Connection Limitation in PyTorch Distributed (Vanilla) with c10d Rendezvous Backend

  #144856 commented on Jan 21, 2025 • 0 new comments

• CheckpointError with `torch.distributed.algorithms._checkpoint.checkpoint_wrapper` and `torch.compile`

  #144637 commented on Jan 21, 2025 • 0 new comments

• Better mergebot messages when reverting a PR

  #139680 commented on Jan 21, 2025 • 0 new comments

• [CI] Manywheel image should use hash based on `.ci/docker` directory

  #142218 commented on Jan 21, 2025 • 0 new comments

• ONNX: wrong operator for ceil_mode Pooling in case of skip the last window

  #131272 commented on Jan 21, 2025 • 0 new comments

• DataLoader hangs when object fails during pickling

  #142884 commented on Jan 21, 2025 • 0 new comments

• python-3.13t binaries are only available for Linux x86

  #144357 commented on Jan 21, 2025 • 0 new comments

• Bug when using reparameterized model evaluating with DDP

  #145043 commented on Jan 21, 2025 • 0 new comments

• torch.distributed hangs between Linux (X86) and Mac (M2 Pro)

  #144851 commented on Jan 21, 2025 • 0 new comments

• Consider making torch.cond return zero rather than None for the gradients of tensors that are in the not-taken branch of the if-else.

  #141301 commented on Jan 21, 2025 • 0 new comments

• Partitioner stores fp8 copy of all weights between fwd and bwd, causing OOM

  #141881 commented on Jan 21, 2025 • 0 new comments

• [PassRate] TorchBench training PassRate is less than 100

  #143414 commented on Jan 21, 2025 • 0 new comments

• torch.compile does not work with Flash attention 3

  #144540 commented on Jan 21, 2025 • 0 new comments

• DISABLED test_grad_scaling_autocast_cuda (__main__.TestTorchDeviceTypeCUDA)

  #119154 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_tensor_subclasses (__main__.TestScript)

  #119949 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_comprehensive_cross_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #140355 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_dunder_round_edgecases_val_2147483647_ndigits_-1 (__main__.TestNonarrayArgs)

  #116121 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_is_isnot (__main__.TestScript)

  #120694 commented on Jan 17, 2025 • 0 new comments

• Is it possible to remove NCCL submodule and use only nccl binaries from pypi instead ?

  #144768 commented on Jan 17, 2025 • 0 new comments

• Cannot create and distribute array in torch.func.grad

  #134462 commented on Jan 17, 2025 • 0 new comments

• [Dynamo] Do an audit on skipfiles and mark more files as inline

  #142395 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_profiler_mark_wrapper_call_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)

  #145002 commented on Jan 17, 2025 • 0 new comments

• DTensor RNG state for non CUDA backends

  #138329 commented on Jan 17, 2025 • 0 new comments

• amp.custom_fwd has incomplete support for library.custom_op

  #137033 commented on Jan 17, 2025 • 0 new comments

• ExpandableMemorySegments not working on H100s/A100s

  #122057 commented on Jan 17, 2025 • 0 new comments

• AMP doesn't gracefully handle optimizers for disabled regions

  #47128 commented on Jan 17, 2025 • 0 new comments

• Support loading and executing a ExportedProgram from torch.export in C++ environment

  #144663 commented on Jan 17, 2025 • 0 new comments

• [Pipelining] PP+DDP does not work for Zero Bubble

  #144530 commented on Jan 17, 2025 • 0 new comments

• `_pdist_forward` causes segmentation fault for 3D tensor with last dimension of size 0

  #145064 commented on Jan 17, 2025 • 0 new comments

• auto-grad graph replicate split_with_sizes(lengths) X times where X = len(lengths) effecting compile time

  #140835 commented on Jan 17, 2025 • 0 new comments

• `torch.compiler.disable()` on module hooks will disable `module.compile()`

  #142358 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_serialize_export_scan_simple_cuda_float32 (__main__.TestHOPCUDA)

  #139073 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_retrace_export_scan_simple_cuda_float32 (__main__.TestHOPCUDA)

  #139074 commented on Jan 17, 2025 • 0 new comments

• Support torch.func.grad for Flex Attention

  #144810 commented on Jan 17, 2025 • 0 new comments

• torch.stack for sequences

  #144671 commented on Jan 17, 2025 • 0 new comments

• nn.LSTM documentation

  #139582 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_device_mode_ops_sparse_sampled_addmm_cpu_complex64 (__main__.TestDeviceUtilsCPU)

  #132686 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_dtype_sympy_expr_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)

  #135213 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_tmp_not_defined_issue2_dynamic_shapes_cpu (__main__.DynamicShapesCodegenCpuTests)

  #135212 commented on Jan 17, 2025 • 0 new comments

• LibTorch -> TorchScript -> PyTorch (Python) fails with `AttributeError: 'RecursiveScriptModule' object has no attribute 'forward'`

  #68559 commented on Jan 20, 2025 • 0 new comments

• TorchInductor CPU Performance Dashboard

  #93531 commented on Jan 20, 2025 • 0 new comments

• Set up Mac builds with clang >= 17 even though Xcode only has at most clang 16

  #143913 commented on Jan 20, 2025 • 0 new comments

• Python 3.13 support for PyTorch

  #130249 commented on Jan 20, 2025 • 0 new comments

• tts_angular: fail_to_run, torch._dynamo.exc.Unsupported: call_method NNModuleVariable() flatten_parameters [] {}

  #105532 commented on Jan 20, 2025 • 0 new comments

• Illegal Memory Access With `torch.compile`

  #139628 commented on Jan 20, 2025 • 0 new comments

• TypeError: Type parameter +RV without a default follows type parameter with a default in _inductor/utils.py

  #140914 commented on Jan 20, 2025 • 0 new comments

• DISABLED test_matmul_triton_kernel_benchmark (__main__.TestKernelBenchmark)

  #115002 commented on Jan 20, 2025 • 0 new comments

• Device assert throws a runtime error in cuda backend and results in a crash in xpu backend

  #142135 commented on Jan 20, 2025 • 0 new comments

• Matmul with int32 parameters on Intel GPU leads to errors

  #144766 commented on Jan 20, 2025 • 0 new comments

• [Inductor] [CPU] `GroupNorm` triggers inconsistency when using Inductor

  #141541 commented on Jan 20, 2025 • 0 new comments

• [RFC] Intel GPU distributed Backend integration in `torch-xpu-ops`and registeration in PyTorch

  #141741 commented on Jan 20, 2025 • 0 new comments

• DISABLED test_profiler_mark_wrapper_call_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)

  #135294 commented on Jan 20, 2025 • 0 new comments

• [torch.export] _insert_copy_for_mutations can't generate proper copy nodes for pure inplace ops

  #144954 commented on Jan 20, 2025 • 0 new comments

• ARM build failed with recent XNNPACK update: third_party/XNNPACK/src/reference/unary-elementwise.cc:125:14: error: invalid ‘static_cast’ from type ‘xnn_bfloat16’ to type ‘_Float16’

  #141083 commented on Jan 20, 2025 • 0 new comments

• DISABLED test_comprehensive_pca_lowrank_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #139828 commented on Jan 20, 2025 • 0 new comments

• Torch profiler corrupted names with Python 3.11

  #121219 commented on Jan 19, 2025 • 0 new comments

• [pytree] Handling of `None` in torch.utils._pytree is inconsistent with JAX.

  #119328 commented on Jan 19, 2025 • 0 new comments

• Add BufferDict container

  #37386 commented on Jan 19, 2025 • 0 new comments

• Responses from `https://download.pytorch.org/whl/cpu` have `cache-control: no-cache` set in their headers

  #130571 commented on Jan 18, 2025 • 0 new comments

• MPS operator coverage tracking issue (2.6+ version)

  #141287 commented on Jan 18, 2025 • 0 new comments

• DISABLED test_input_mutation2_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)

  #135295 commented on Jan 18, 2025 • 0 new comments

• Make streams used for NCCL operations configurable

  #67158 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_device_mode_ops_sparse_sampled_addmm_cpu_float64 (__main__.TestDeviceUtilsCPU)

  #132737 commented on Jan 17, 2025 • 0 new comments

• Observing CUDA OOM errors in more recent versions of PyTorch nightly (post-`2.6.0.dev20241126`)

  #141904 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_torch_to (__main__.TestTEFuserStatic)

  #121876 commented on Jan 17, 2025 • 0 new comments

• DISABLED test_torch_to (__main__.TestTEFuserDynamic)

  #121875 commented on Jan 17, 2025 • 0 new comments

• [dynamo] Fix constant propagation in builtins and UserClasses

  #131354 commented on Jan 21, 2025 • 0 new comments

• [inductor] enable bf32 for mkldnn linear pointwise/binary in inductor

  #127294 commented on Jan 20, 2025 • 0 new comments

• [inductor] enable bf32 test for mkldnn conv

  #127293 commented on Jan 21, 2025 • 0 new comments

• allow to use bf16 as fp32 internal precision for mkldnn conv backward

  #126054 commented on Jan 21, 2025 • 0 new comments

• allow to use bf16 as fp32 internal precision for mkldnn conv

  #126050 commented on Jan 21, 2025 • 0 new comments

• refine fp32 precision api

  #125888 commented on Jan 21, 2025 • 0 new comments

• [vision hash update] update the pinned vision hash

  #125806 commented on Jan 24, 2025 • 0 new comments

• Automated submodule update: FBGEMM

  #115316 commented on Jan 24, 2025 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on Jan 22, 2025 • 0 new comments

• [Inductor test failure] torch:inductor/test_select_algorithm TestSelectAlgorithm.test_convolution1 with cuda 12.6.3

  #143412 commented on Jan 24, 2025 • 0 new comments

• [export] run_decompositions fails on `torch.ops.aten.index_put_`

  #141336 commented on Jan 24, 2025 • 0 new comments

• compile time regression 1/9

  #144775 commented on Jan 24, 2025 • 0 new comments

• [RFC] Add CPP INT8 SDPA Template for Inductor CPU

  #144941 commented on Jan 24, 2025 • 0 new comments

• DISABLED TCPStoreTest.testMultiTenantStores (__main__.TCPStoreTest)

  #142030 commented on Jan 24, 2025 • 0 new comments

• Performance regression in torch.compile

  #136254 commented on Jan 24, 2025 • 0 new comments

• assert size/strides for fallback kernel

  #144717 commented on Jan 24, 2025 • 0 new comments

• FFT half precision only let CUDA pass the check

  #143112 commented on Jan 24, 2025 • 0 new comments

• -fno-omit-frame-pointer by default in our builds

  #51151 commented on Jan 24, 2025 • 0 new comments

• TorchDispatchMode cann't capture the operator which name is aten::index_put_ impl_

  #145041 commented on Jan 24, 2025 • 0 new comments

• [inductor][cpu] With inductor_max_autotune, constants missing from frozen FxGraph.

  #143144 commented on Jan 24, 2025 • 0 new comments

• Support LayerNorm2d

  #144223 commented on Jan 24, 2025 • 0 new comments

• DISABLED test_aot_export_with_torch_cond (__main__.TestAOTExport)

  #139998 commented on Jan 24, 2025 • 0 new comments

• DISABLED test_flip_cpu (__main__.CpuTests)

  #142863 commented on Jan 24, 2025 • 0 new comments

• Allow generic python data structure input for torch.autograd.Function

  #144159 commented on Jan 23, 2025 • 0 new comments

• [ONNX][RFC] Migrate torchlib from onnxscript

  #139301 commented on Jan 23, 2025 • 0 new comments

• Release 2.6.0 validations checklist and cherry-picks

  #144503 commented on Jan 23, 2025 • 0 new comments

• DISABLED TCPStoreTest.testMultiTenantStoresUV (__main__.TCPStoreTest)

  #139150 commented on Jan 23, 2025 • 0 new comments

• Add overflow check for integer division

  #138684 commented on Jan 17, 2025 • 0 new comments

• [Docker] Create an independent dependecies layer

  #138612 commented on Jan 22, 2025 • 0 new comments

• Prototype Triton kernel for torch.bmm(NJT, T)

  #138555 commented on Jan 23, 2025 • 0 new comments

• Update test_function_base.py for Numpy 2.0 +

  #138463 commented on Jan 21, 2025 • 0 new comments

• Replace use of PyTorch 2.0 with torch.compile, and minor edits

  #138436 commented on Jan 23, 2025 • 0 new comments

• [pytree] add `treespec_{leaf,tuple,dict}` functions for args_spec modification

  #138214 commented on Jan 22, 2025 • 0 new comments

• [POC][FX][pytree] cleanup fx pytree implementation

  #138202 commented on Jan 23, 2025 • 0 new comments

• update _unsafe_set_version_counter to accept lists of tensors

  #137921 commented on Jan 24, 2025 • 0 new comments

• [Intel GPU] allow_tf32 context at XPU backend

  #137570 commented on Jan 20, 2025 • 0 new comments

• Add back DistributedDataParallel types that were lost when pyi was removed

  #136835 commented on Jan 20, 2025 • 0 new comments

• Avoid sqrt calculations with values ​​less than zero

  #136824 commented on Jan 24, 2025 • 0 new comments

• [Intel GPU] qlinear.pointwise with mixed dtype support

  #136753 commented on Jan 23, 2025 • 0 new comments

• [Partitioner] Reduce time consuming of partitions merger

  #136614 commented on Jan 23, 2025 • 0 new comments

• [Partitioner] Remove unnecessary upstream nodes in dependency viewer

  #136608 commented on Jan 23, 2025 • 0 new comments

• add generalized pareto distribution (GPD)

  #135968 commented on Jan 24, 2025 • 0 new comments

• [Intel GPU] qconv.pointwise with mixed dtype XPU support

  #135465 commented on Jan 23, 2025 • 0 new comments

• add supports_coalescing property in c10d::Backend to determine whether backend supports coalescing

  #135338 commented on Jan 23, 2025 • 0 new comments

• [Intel GPU] qlinear_pointwise.binary[_tensor] XPU support

  #135337 commented on Jan 23, 2025 • 0 new comments

• Pass ideep:lowp_kind to matmul_forward::compute on cache misses

  #135058 commented on Jan 23, 2025 • 0 new comments

• Add decompositions for median and nonmedian

  #134881 commented on Jan 23, 2025 • 0 new comments

• [ROCm] Add support for SymmetricMemory and Intra Node Comm

  #134817 commented on Jan 17, 2025 • 0 new comments

• add ranking for grouped benchmarks

  #133287 commented on Jan 18, 2025 • 0 new comments

• Make IPC features extendable on third-party devices

  #133222 commented on Jan 23, 2025 • 0 new comments

• basic GroupedInductorBenchmarker

  #133121 commented on Jan 18, 2025 • 0 new comments

• xpu: support sycl with torch.utils.cpp_extension APIs

  #132945 commented on Jan 18, 2025 • 0 new comments

• Added dist utility API to get backend from a device object

  #132735 commented on Jan 24, 2025 • 0 new comments

• [xla hash update] update the pinned xla hash

  #132021 commented on Jan 20, 2025 • 0 new comments

• DISABLED test_cuda_event_created_outside_of_graph (__main__.CtxManagerTests)

  #133828 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_cuda_event_created_outside_of_graph_dynamic_shapes (__main__.DynamicShapesCtxManagerTests)

  #133837 commented on Jan 23, 2025 • 0 new comments

• [Tensorboard] Problem with subfolders from SummaryWriter

  #32651 commented on Jan 23, 2025 • 0 new comments

• Runners, torchbench, & the future

  #143215 commented on Jan 23, 2025 • 0 new comments

• Need clarification on torch.nn.CrossEntropyLoss

  #137188 commented on Jan 23, 2025 • 0 new comments

• torch.library.opcheck generates gradients with strides of 0

  #132857 commented on Jan 23, 2025 • 0 new comments

• `torch.nn.function.one_hot` and `torch.Tensor.as_subclass` API not available under `torch.compile`

  #129651 commented on Jan 23, 2025 • 0 new comments

• ExportedModule default print of graph signature is unreadable

  #141243 commented on Jan 23, 2025 • 0 new comments

• TORCH_PYTHON_API contains breaking changes in same version 2.6.0a0

  #144966 commented on Jan 22, 2025 • 0 new comments

• int_mm seems broken due to Triton upgrade

  #144705 commented on Jan 22, 2025 • 0 new comments

• [feature request] Varlen indexing function for lookup and concat of varlen BPE tokens from a tensor vocab (i.e. `detokenize(...)` and arrays of strings)

  #135704 commented on Jan 22, 2025 • 0 new comments

• [triton 3.2] test_convolution_as_mm failure on A100

  #141079 commented on Jan 22, 2025 • 0 new comments

• `torch.ops.aten._local_scalar_dense` crashed on empty size tensor

  #145063 commented on Jan 22, 2025 • 0 new comments

• Method for loading a distributed checkpoint into a single state_dict is being deprecrated without alternative, request to make it possible to keep that feature

  #125777 commented on Jan 22, 2025 • 0 new comments

• [DCP]Distributed checkpoint `set_optimizer_state_dict` cause optimizer step error when optimizer contains empty param group

  #143828 commented on Jan 22, 2025 • 0 new comments

• Should coordinator_rank in class _DistWrapper be the global_rank instead of local rank in its process group?

  #141825 commented on Jan 22, 2025 • 0 new comments

• Compile error for custom op with optional mutable tensor list argument

  #144072 commented on Jan 22, 2025 • 0 new comments

• Really slow compilation times for torch.compile causing distributed training errors

  #108971 commented on Jan 22, 2025 • 0 new comments

• On Kaggle : libcusparse.so.12: undefined symbol: __nvJitLinkComplete_12_4, version libnvJitLink.so.12

  #134929 commented on Jan 22, 2025 • 0 new comments

• Support SDPA flash attention/ memory efficant attn on ROCm gfx908

  #141958 commented on Jan 22, 2025 • 0 new comments

• torch compile error with `torch.Tensor.unsqueeze_`

  #129673 commented on Jan 22, 2025 • 0 new comments

• Wrong meta function for constant_pad_nd

  #144187 commented on Jan 22, 2025 • 0 new comments

• Add ATen functions in native_functions.yaml to torch_in_graph_functions list automatically

  #145014 commented on Jan 22, 2025 • 0 new comments

• torch.nn.functional.scaled_dot_product_attention is_causal fails for kv-cache case (sequential and further parallel attention)

  #144858 commented on Jan 22, 2025 • 0 new comments

• TIMM Training cudagraphs poolformer_m36 regression

  #144893 commented on Jan 22, 2025 • 0 new comments

• PyTorch source code build failed on some Windows 11 environment caused by C++ protocol buffer compiler

  #143795 commented on Jan 22, 2025 • 0 new comments

• Fix `torch.stft` and `torch.istft` when using `center=False`, non-rectangular windows and `win_length==hop_length`

  #134323 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_reentrant_parent_error_on_cpu_cuda (__main__.TestAutogradDeviceTypeCUDA)

  #86735 commented on Jan 23, 2025 • 0 new comments

• Meta implementations of FFT operators often have incorrect strides

  #106623 commented on Jan 23, 2025 • 0 new comments

• Adding Levenberg-marquardt optimizer in PyTorch

  #83529 commented on Jan 23, 2025 • 0 new comments

• Update TorchInductor to support removed AttrsDescriptor in upstream Triton

  #144103 commented on Jan 23, 2025 • 0 new comments

• [v.2.6.0] Release Tracker

  #142814 commented on Jan 23, 2025 • 0 new comments

• compiled autograd + dynamic shapes fails with constraint violation

  #133575 commented on Jan 23, 2025 • 0 new comments

• [feature request]: Update max onnx opset to 21 for onnxruntime==1.18 compatability

  #127167 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_allocation_id_uniqueness (__main__.TestTorchTidyProfiler)

  #125021 commented on Jan 23, 2025 • 0 new comments

• AWS A100 runners reliability issue

  #140332 commented on Jan 23, 2025 • 0 new comments

• [MPS] Indexing Returns 0 if OOB

  #144824 commented on Jan 23, 2025 • 0 new comments

• capture_dynamic_output_shape_ops=True changing expected output between eager and compiled versions

  #130290 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_angle_cpu (__main__.CpuTritonTests)

  #136124 commented on Jan 23, 2025 • 0 new comments

• pytorch with xpu support fails to eval pre trained models

  #143996 commented on Jan 23, 2025 • 0 new comments

• FSDP does not work on GLOO backend

  #74041 commented on Jan 23, 2025 • 0 new comments

• AOTAutogradCache implementation

  #128234 commented on Jan 23, 2025 • 0 new comments

• FlexAttention + ROCM Issue Tracker

  #140855 commented on Jan 23, 2025 • 0 new comments

• MaxPool2D memory leakage on device MPS

  #125217 commented on Jan 23, 2025 • 0 new comments

• General MPS op coverage tracking issue

  #77764 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_crash (__main__.TestCompileWorker)

  #131064 commented on Jan 23, 2025 • 0 new comments

• [torchbench] Missing meta function for aten::_cudnn_rnn_flatten_weight

  #144989 commented on Jan 23, 2025 • 0 new comments

• [Tracker] Nested tensor op coverage requests

  #118107 commented on Jan 23, 2025 • 0 new comments

• Issues linking to libtorch on M2 mac

  #143571 commented on Jan 23, 2025 • 0 new comments

• [RFC] PyTorch - PyPi PEP-759 proposal (wheel-next)

  #139761 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_log_traced_frames (__main__.LoggingTests)

  #137461 commented on Jan 23, 2025 • 0 new comments

• Inconsistent computation of gradient in MaxUnPooling

  #80827 commented on Jan 23, 2025 • 0 new comments

• DISABLED test_new_spectral_norm_forward_swap_True (__main__.TestNNParametrization)

  #131089 commented on Jan 23, 2025 • 0 new comments