2 Pull requests merged by 2 people

• [dynamo][super variable] Fix bug to use correct source

  #152774 merged May 6, 2025

• [cudagraphs] Fix issue in collecting static_input_idxs

  #152768 merged May 6, 2025

206 Pull requests opened by 117 people

• [CI] Use cmake from pip instead of conda in CI docker images

  #152537 opened Apr 30, 2025

• Use swap_tensors path in nn.Module.to for all subclasses that override __torch_dispatch__

  #152539 opened Apr 30, 2025

• [CUDA] Rest peak memory stats before running `test_set_per_process_memory_fraction`

  #152540 opened Apr 30, 2025

• strict multidimensional slicing

  #152543 opened Apr 30, 2025

• ci: Switch benchmark dependency to use pip

  #152545 opened Apr 30, 2025

• Remove Conda Instructions

  #152546 opened Apr 30, 2025

• Implemented `Size.__radd__`

  #152554 opened Apr 30, 2025

• [BE] Update numba versions

  #152557 opened Apr 30, 2025

• xpu: rely on sycl/sycl.hpp to include bfloat16.hpp

  #152562 opened Apr 30, 2025

• [c10d][fr] Make FR vendor neutral so that other backends can use it

  #152563 opened Apr 30, 2025

• [ROCm] Update spack includes

  #152569 opened Apr 30, 2025

• [Hierarchical Compile] Replace tracing alias and mutation check with dynamo impl

  #152570 opened Apr 30, 2025

• [Dynamo] Fix typing in graph_deduplication.py

  #152572 opened May 1, 2025

• Allow decomposeK to fuse

  #152573 opened May 1, 2025

• [2/N] Use std::filesystem

  #152586 opened May 1, 2025

• [WIP] suggest whitelist for dynamic shape recompilations

  #152588 opened May 1, 2025

• [Dynamo] Optimize dedupe region ancestor tracking

  #152589 opened May 1, 2025

• Fix #152280: add Literal[…] PaddingMode to Conv modules

  #152590 opened May 1, 2025

• Fix: promote scalar to MPS device in exec_binary_kernel

  #152591 opened May 1, 2025

• [c10d] Add support for ReduceOp::AVG in ProcessGroupMPI for FSDP2

  #152594 opened May 1, 2025

• [not for review] benchmark script

  #152596 opened May 1, 2025

• [multigraph] add backend_specialization kwarg to mark_dynamic

  #152597 opened May 1, 2025

• [multigraph] use backend specializations in compile_and_call_fx_graph

  #152601 opened May 1, 2025

• [BE] Delete `Module_CUDA_fix`

  #152603 opened May 1, 2025

• [Testing] Is FindCUDA.cmake from `Modules_CUDA_fix` called at all?

  #152604 opened May 1, 2025

• [Environment Variable] Use thread-safe getenv functions

  #152609 opened May 1, 2025

• Update padding_mode type annotation to use Literal type (PaddingMode)

  #152610 opened May 1, 2025

• Makefile: refactor build, setup and lint rules

  #152611 opened May 1, 2025

• Revert "Cleanup VS 2019 refs in pytorch (#145863)"

  #152613 opened May 1, 2025

• [WIP] Make FR vendor generic and try to enable it for gloo

  #152614 opened May 1, 2025

• Stop proxy-ing autograd.Function.ctx into the graph

  #152621 opened May 1, 2025

• Parameterized CUDA Graph Launch

  #152622 opened May 1, 2025

• [pytree] make `tree_*` functions accept both Python and C++ `PyTreeSpec`

  #152624 opened May 1, 2025

• [ROCm] Initial AITER Integration for mha_bwd asm kernels

  #152630 opened May 1, 2025

• [ca] wrap flex attention tests with compiled autograd

  #152633 opened May 1, 2025

• [CUTLASS][WIP] Gate rowwise matmul CUTLASS kernels by compute capability

  #152642 opened May 1, 2025

• [BE]remove vulkan test

  #152643 opened May 1, 2025

• [do-not-land][ca] default on for CI

  #152646 opened May 1, 2025

• Add assert_fp8_close helper for FP8 tensor comparisons

  #152651 opened May 2, 2025

• Refactor some common autotune-related utils into a new file

  #152652 opened May 2, 2025

• cleanup, refactor and add missing self._dde_suppressed checks

  #152657 opened May 2, 2025

• Fix the basic description of torch.min(), torch.max(), torch.all(), torch.any()

  #152658 opened May 2, 2025

• Fix evaluate_expr to include suppress_guards_tls in cache key

  #152661 opened May 2, 2025

• Re-enable FakeTensor caching for SymInts

  #152662 opened May 2, 2025

• Raise error when no record on extra_files

  #152664 opened May 2, 2025

• MXFP8 Fix broken bias support for mxfp8

  #152665 opened May 2, 2025

• Added documentation for nonzero_static function (#152347)

  #152669 opened May 2, 2025

• [export] Dynamo symint support

  #152677 opened May 2, 2025

• Fix signature of torch.sparse_coo_tensor()

  #152681 opened May 2, 2025

• Update the signature and test of torch.hamming_window()

  #152682 opened May 2, 2025

• [ca][dtensor] run real PG dtensor tests under CA

  #152689 opened May 2, 2025

• set CUDA_MODULE_LOADING for older drivers only

  #152695 opened May 2, 2025

• Removing conda references from PyTorch Docs

  #152702 opened May 2, 2025

• [Memento] Add PT2 to Memory Snapshot

  #152707 opened May 2, 2025

• Scheduler Flops refactor

  #152708 opened May 2, 2025

• remove conda from devcontainer

  #152713 opened May 2, 2025

• [caffe2] Make c10::str works with scoped enum (#152705)

  #152714 opened May 2, 2025

• [BE][CI] Merge regular and MPS test config shards

  #152719 opened May 2, 2025

• try something

  #152722 opened May 2, 2025

• [Dynamo] Guard serialization for NN_MODULE

  #152725 opened May 2, 2025

• [Dynamo] Guard serialization for FUNCTION_MATCH

  #152727 opened May 2, 2025

• [Dynamo] Guard serialization for CLOSURE_MATCH

  #152728 opened May 2, 2025

• [Dynamo] Guard serialization for BUILTIN_MATCH

  #152729 opened May 2, 2025

• [Dynamo] Guard serialization for SEQUENCE_LENGTH

  #152730 opened May 2, 2025

• [Cutlass] Handle broadcasting in EVT python codegen

  #152733 opened May 2, 2025

• docs: fix dead link in torch.compile docs

  #152734 opened May 2, 2025

• [ca][ddp] loud error instead of silent incorrectness under C++ Reducer

  #152735 opened May 2, 2025

• [BE][Cleanup][Dynamo] Stop logging entire_frame_compile_time_s

  #152738 opened May 2, 2025

• [export] add serialized_artifact test

  #152739 opened May 2, 2025

• [export][cond] support merging constant ints as unbacked symint

  #152742 opened May 2, 2025

• [CUDA][cuDNN] Fix handling of `CPU` side input and target length tensors in `CTCLoss`

  #152745 opened May 3, 2025

• Conditionally support experimental filesystem include in jit_opt_limit

  #152748 opened May 3, 2025

• Handle less functions than number of segments

  #152753 opened May 3, 2025

• Allow ATen ops overloading

  #152759 opened May 3, 2025

• [Easy][BE] update recommanded VS Code settings

  #152760 opened May 3, 2025

• added short integer for repeat_interleave_cpu, Fixes #151311

  #152762 opened May 3, 2025

• [aoti] Add grid_sampler_3d to cshim

  #152771 opened May 4, 2025

• [Inductor] Pattern matcher support for mutable ops with non-view inputs

  #152775 opened May 4, 2025

• [WIP] Pattern matcher support for mutable ops with view inputs

  #152776 opened May 4, 2025

• [BE]: Update cudnn to 9.9 for cu128

  #152782 opened May 4, 2025

• test that guard_or_true change can only make valid results null but does not change result or make invalid valid

  #152784 opened May 4, 2025

• Fix negative dim issue in for parallel loss context manager

  #152785 opened May 4, 2025

• Update CMakeLists.txt

  #152786 opened May 4, 2025

• Implement DeviceType.h as header-only

  #152787 opened May 4, 2025

• Fixed rerr computation in lobpcg

  #152789 opened May 4, 2025

• same test for guard_or_false 1

  #152802 opened May 5, 2025

• same test for guard_or_false 2

  #152803 opened May 5, 2025

• [invoke_subgraph] Force the output stride to be same as eager

  #152806 opened May 5, 2025

• wip

  #152807 opened May 5, 2025

• another try

  #152808 opened May 5, 2025

• Upgrade to NCCL 2.26.5 for CUDA 12

  #152810 opened May 5, 2025

• [Quant][X86] add an op to compute uint8 batch norm 2d

  #152811 opened May 5, 2025

• [TEST][ATen][CUDA] Skip row-wise scaled matrix mmultiplication tests on sm_120+

  #152814 opened May 5, 2025

• [Cutlass] E2E Tests for EVT

  #152815 opened May 5, 2025

• [TEST][Quantization] Skip test_learnable due to hypothesis

  #152819 opened May 5, 2025

• [DO NOT MERGE] update build tools version

  #152820 opened May 5, 2025

• [Easy][Inductor] Adds safety checks in get_estimated_runtime

  #152821 opened May 5, 2025

• Use gcc13 in Manylinux 2.28 images

  #152825 opened May 5, 2025

• [MSVC] Enable updated lambda processor by setting compiler flag /Zc:lambda globally

  #152828 opened May 5, 2025

• [BE]: Improve aten formatter with fmtlib

  #152830 opened May 5, 2025

• Allow to set custom PYTHONPATH for torch.inductor

  #152832 opened May 5, 2025

• [c10d] Fix extra CUDA context created by barrier

  #152834 opened May 5, 2025

• [DRAFT] Test nccl

  #152835 opened May 5, 2025

• [nativert] Move MPMCQueue to torch/nativert.

  #152837 opened May 5, 2025

• [precompile] Add BundledAOTAutogradCacheEntry

  #152840 opened May 5, 2025

• Add memory reporting for XPU to Memory Profiler

  #152842 opened May 5, 2025

• ci: Remove conda-env-macOS-ARM64, prefer pip

  #152843 opened May 5, 2025

• Fix HF loading when there's no metadata file to work with fsspec

  #152856 opened May 5, 2025

• Add torch._C.Tag.needs_contiguous_strides

  #152859 opened May 5, 2025

• [Graph Partition] remove weak dep from `partition_input_names`

  #152863 opened May 5, 2025

• [Dynamo] Guard serialization for TUPLE_ITERATOR_LEN

  #152865 opened May 5, 2025

• [Dynamo] Guard serialization for RANGE_ITERATOR_MATCH

  #152872 opened May 5, 2025

• Clarify wrap_triton doc about optional triton_op usage

  #152874 opened May 5, 2025

• [Graph Partition][Flex Attention] analyze symints from subgraph inputs and outputs

  #152878 opened May 5, 2025

• xpu: support custom ops with torch.library on xpu backend

  #152879 opened May 5, 2025

• [dynamo] Fix bug in hasattr(tensor, "size")

  #152883 opened May 6, 2025

• [SDPA] Add testing to ensure stride order exactly matches

  #152894 opened May 6, 2025

• Move link check jobs to pull to go with doc build

  #152896 opened May 6, 2025

• [Inductor] Set correct baseline for decomposek test

  #152897 opened May 6, 2025

• Remove `property` from python_type function

  #152900 opened May 6, 2025

• [Set] Add set.symmetric_difference(_update)

  #152901 opened May 6, 2025

• [Set] Add `set.issubset` and `set.issuperset`

  #152902 opened May 6, 2025

• [Set] Raise `KeyError` if elem not contained in the set

  #152903 opened May 6, 2025

• [Set] Raise TypeError if number of arguments mismatch

  #152904 opened May 6, 2025

• [Set] Add `set.difference(_update)`

  #152905 opened May 6, 2025

• [Set] Add `set.intersection(_update)`

  #152906 opened May 6, 2025

• [Set] Raise KeyError on empty `set.pop()`

  #152907 opened May 6, 2025

• [Set] Add correct set/frozenset __init__ behavior

  #152908 opened May 6, 2025

• Add overall tensor similarity comparison (#152647)

  #152920 opened May 6, 2025

• Upgrade to cuda 12.8.1 for docker builds

  #152923 opened May 6, 2025

• include user stacks with constraint violation error message

  #152924 opened May 6, 2025

• [WIP] Add unified memory APIs for torch.accelerator

  #152932 opened May 6, 2025

• Allow Inductor backends to attest their own availability

  #152933 opened May 6, 2025

• [Dynamo] Allow inlining into AO quantization modules

  #152934 opened May 6, 2025

• Fix doc cosineannealinglr 152081

  #152936 opened May 6, 2025

• [feature] Channel Wise Parallel API for Conv layers

  #152937 opened May 6, 2025

• [Pipelining] Fix _batch_p2p bug for non-NCCL backends (#132644)

  #152938 opened May 6, 2025

• get right function declaration on windows inductor

  #152939 opened May 6, 2025

• [Don't merge] Debug

  #152940 opened May 6, 2025

• Clean up of CUTLASS_VERSION

  #152947 opened May 6, 2025

• [Linter] Add linter to detect device-bias hard code in test cases.

  #152948 opened May 6, 2025

• [dtensor] add privateuse1 SDPA op support to DTensor

  #152949 opened May 6, 2025

• Add NestedTensorHPU to to_padded_tensor in native_functions.yaml

  #152950 opened May 6, 2025

• [ROCm] Ck gemm architecture guard

  #152951 opened May 6, 2025

• [nativert] Move Placement to pytorch core

  #152953 opened May 6, 2025

• [ROCm] unkip test_non_standard_bool except for failings ops

  #152956 opened May 6, 2025

• Follow up to #152209, remove compat patch

  #152958 opened May 6, 2025

• docs: Improve documentation for NCCL timeout / watchdog variables

  #152959 opened May 6, 2025

• Change aoti cpp tests to run serially within file

  #152960 opened May 6, 2025

• [Dynamo] Remove unused guard PYMODULE_MATCH

  #152961 opened May 6, 2025

• WIP so many changes to generate non-as strided view

  #152965 opened May 6, 2025

• [Memento] On-demand mode using without torch api

  #152966 opened May 6, 2025

• [ATen][CUDA] Optimize 128 bit vectorization

  #152967 opened May 6, 2025

• [inductor] Generate synthetic offsets appropriately for autotuning _scaled_grouped_mm

  #152968 opened May 6, 2025

• [nativert] Move GraphSignature to pytorch core

  #152969 opened May 6, 2025

• Adding XPU support to DTensor examples.

  #152973 opened May 6, 2025

• [hop_schema] add HopSchemaGenerator to make it easier to create hop schema

  #152974 opened May 6, 2025

• [dtensor] Extend Partial partition of replicated tensor for min/max reduce

  #152975 opened May 6, 2025

• [MegaCache] Make MegaCache generic to allow external plugins registration

  #152977 opened May 6, 2025

• [Pytorch] Add `torch.cuda.streams.Event` to save torch functions list

  #152978 opened May 6, 2025

• Fix `'TensorBox' object has no attribute 'is_input_buffer'`

  #152980 opened May 6, 2025

• Catch TypeError from ValueRanges

  #152981 opened May 6, 2025

• [torch][ao] Properly strip tracking stats in _fold_conv_bn_qat for 1D

  #152982 opened May 6, 2025

• compile_fx: make a compile event that corresponds to the fx_compile waitcounter

  #152983 opened May 6, 2025

• [hop_schema] support gen_schema for invoke_subgraph

  #152984 opened May 6, 2025

• [WIP] Add XPU support for FlightRecorder

  #152986 opened May 6, 2025

• [Set] Handle exception in ConstantVariable operation

  #152987 opened May 6, 2025

• [Set] Raise `TypeError` if argument is unhashable

  #152988 opened May 6, 2025

• [Set] Update `set.union` and `set.update` to support *args

  #152989 opened May 6, 2025

• [Set] Raise TypeError if set is called with the wrong number of arguments

  #152990 opened May 6, 2025

• [FrozenSet] Fixes for FrozenSet

  #152991 opened May 6, 2025

• [inductor] Fix ModularIndexing assumptions

  #152993 opened May 6, 2025

• [inductor] dtype promotion error in cat decomp

  #152995 opened May 6, 2025

• [export] Unflatten None

  #153000 opened May 6, 2025

• [cutlass backend][test] re-enable test_cuda_compile_command for fbcode

  #153001 opened May 6, 2025

• [CI] Use sccache installed in docker image in xla build

  #153002 opened May 6, 2025

• [cutlass backend] Skip cuda lib path if it is torch/lib

  #153003 opened May 6, 2025

• [WIP][Inductor-CPU] int8 WoQ concat linear

  #153004 opened May 6, 2025

• [autograd][docs] Add more details on why save_for_backward is important in extending autograd note

  #153005 opened May 6, 2025

• [cutlass backend] Use src code to generate cutlass gemm name

  #153006 opened May 6, 2025

• Remove redundant type aliases of _device_t for torch.Device (#152952)

  #153007 opened May 7, 2025

• Detect NVSHMEM location

  #153010 opened May 7, 2025

• [WIP][dynamic shapes] unbacked safer cat, repeat

  #153011 opened May 7, 2025

• c10d/gloo: add ibverbs backend

  #153015 opened May 7, 2025

• Add a project section to pyproject.toml, making uv sync work

  #153020 opened May 7, 2025

• Adding a generic attribute for easier checkpoint discrepancy debugging.

  #153021 opened May 7, 2025

• [Typing] Apply `torch.types.Device` in `torch/cuda/memory.py`

  #153027 opened May 7, 2025

• [Typing] Improve device typing for `torch.set_default_device()`

  #153028 opened May 7, 2025

• WIP: Fix caching when output has unbacked

  #153034 opened May 7, 2025

• Allow zero sized dimensions in padding operations

  #153037 opened May 7, 2025

• Add CUDA support for Adagrad(fused=True)

  #153038 opened May 7, 2025

• [Dynamo] Replace `unimplemented` with `unimplemented_v2` in `torch/_dynamo/variables/misc.py` [2/2]

  #153039 opened May 7, 2025

• [AOTInductor] Generate kernels separately for const graph and main graph

  #153040 opened May 7, 2025

• [Typing] Remove redundant type aliases of `_device_t` for `torch.types.Device` in `torch/_dynamo/device_interface.py`

  #153043 opened May 7, 2025

• 🌠 Add Muon optimizer

  #153048 opened May 7, 2025

• Update docs of saved_tensors_hooks to avoid ref cycle

  #153049 opened May 7, 2025

• [Intel GPU] empty-size tensor case handling in addmm, baddmm

  #153051 opened May 7, 2025

• [BE]: Use undocumented temp shim to restore setuptools compat

  #153052 opened May 7, 2025

• [BE]: Blacklist broken setuptools until we upgrade MSVC API

  #153053 opened May 7, 2025

• [HOP] Reworked HOPs to use FunctionalizeCtxWrapper

  #153054 opened May 7, 2025

• [BE]: Add PEP621 project section to pyproject.toml

  #153055 opened May 7, 2025

• [BE] Update ruamel to 0.18.10

  #153057 opened May 7, 2025

• Fix misleadingly high AOT Inductor dashboard performance

  #153060 opened May 7, 2025

• Keep raw cubin file around in case it gets deleted underneath us

  #153064 opened May 7, 2025

• [ONNX] dynamic_shapes uses DYNAMIC

  #153065 opened May 7, 2025

• fix bug with TORCHINDUCTOR_DUMP_LAUNCH_PARAMS

  #153066 opened May 7, 2025

96 Issues closed by 42 people

• manylinux_2_28 support

  #114232 closed May 7, 2025

• [graph pickler] [inductor compile async] imprecise filter for non standard op?

  #151904 closed May 7, 2025

• [inductor] cudagraph error for individually compiled transformer blocks

  #152887 closed May 7, 2025

• DISABLED test_sdpa_compile_cuda_bfloat16 (__main__.TestNestedTensorSubclassCUDA)

  #119903 closed May 7, 2025

• DISABLED test_host_memory_stats (__main__.TestCuda)

  #148607 closed May 7, 2025

• [XPU] test_tensordot_out_kernel_errors_with_autograd_xpu_float32 UT failure

  #152090 closed May 7, 2025

• Failed visualized 1D DTensor

  #152848 closed May 7, 2025

• [RFC] A device-agnostic Python runtime API design for stream-based accelerators

  #128403 closed May 7, 2025

• DISABLED test_comprehensive_scatter_xpu_bool (__main__.TestInductorOpInfoXPU)

  #153018 closed May 7, 2025

• DISABLED test_comprehensive_scatter_xpu_int64 (__main__.TestInductorOpInfoXPU)

  #153017 closed May 7, 2025

• Add metal-flash-attention for MPS backend

  #139668 closed May 7, 2025

• [MPS] Binary kernels produce incorrect results when one of the tensor arguments is from a wrapped scalar

  #152582 closed May 7, 2025

• [ONNX] Create a message to suggest users setting dynamo=True when exporting

  #152025 closed May 6, 2025

• [dynamo] register_module_forward_pre_hook lead to compiled model produce wrong inference results

  #149502 closed May 6, 2025

• [CI] [anaconda] Review Devcontainer anaconda usage

  #148341 closed May 6, 2025

• addmv bfloat16 accuracy issues on cpu

  #147860 closed May 6, 2025

• Optimize printing sympy expressions during logging and cache key computation

  #151823 closed May 6, 2025

• UNSTABLE pull / linux-docs / build-docs-functorch-false

  #152955 closed May 6, 2025

• Parameters between models don't copy in the C++ Pytroch Frontend under windows

  #114485 closed May 6, 2025

• Unexpected result from `torch.xpu.is_bf16_supported()` when XPU is unavailable

  #152301 closed May 6, 2025

• DISABLED test_dynamo_timed (__main__.TestDynamoTimed)

  #148093 closed May 6, 2025

• Running `LazyModuleMixin` example throw errors

  #150404 closed May 6, 2025

• can't build torch on WSL

  #152763 closed May 6, 2025

• [AOTI] Package lowered with package_constants_in_so=False still uses lots of memory when loaded

  #152356 closed May 6, 2025

• Throwing more specific errors for CrossEntropyLoss weights being on a different device than the input/target

  #122757 closed May 6, 2025

• DISABLED test_matmul_layer_norm_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)

  #151835 closed May 6, 2025

• DISABLED test_tmp_not_defined_issue2_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)

  #135219 closed May 6, 2025

• [XPU] Get [ZE]: 0x78000011 on torch.compile with new driver

  #151898 closed May 6, 2025

• flex attention does not leverage masking, memory error

  #152528 closed May 5, 2025

• Flex attention: batch-index-dependent block mask causes error with changing batch size

  #152297 closed May 5, 2025

• torch.compile LLMs on MPS progress tracker

  #150710 closed May 5, 2025

• [ued] Investigate diffuser pipeline transformer recompilations due to different width/height

  #150702 closed May 5, 2025

• [binary builds] Anaconda. Remove dependency on conda libuv module in MacOS and Windows nightly builds

  #145872 closed May 5, 2025

• Dynamo Unsupported: call_method UserDefinedObjectVariable(dict_items) __iter__ () {}

  #147440 closed May 5, 2025

• [binary builds] Anaconda. Remove dependency on conda environment for Windows nightly builds

  #146048 closed May 5, 2025

• torch.multinomial is not deterministic for large number of input probabilities when replacement=True

  #152854 closed May 5, 2025

• AsyncCollectiveTensor doesn't trigger wait upon dtype cast

  #152534 closed May 5, 2025

• Poor performance of torch.dot with float16 & bfloat16

  #152798 closed May 5, 2025

• Segmentation fault (core dumped) in torch.nn.functional.max_unpool2d

  #152804 closed May 5, 2025

• Mention of nondeterministic index_add when deterministic implementation is being used

  #152817 closed May 5, 2025

• There is some discrepancy between the document explain and code implement of optim.SGD() when use maximize=True

  #149476 closed May 5, 2025

• [RFC][PGNCCL] Add Float8 support

  #148344 closed May 5, 2025

• False INTERNAL ASSERT FAILED

  #152805 closed May 5, 2025

• [inductor] [triton] the generated triton code throws `NameError('rindex is not defined')` when using `torch.cummin`

  #151738 closed May 5, 2025

• The docstring linter should not force overridden methods to be documented

  #151692 closed May 5, 2025

• [RFC] Intel GPU ATen Operations Upstreaming Options

  #119682 closed May 5, 2025

• DISABLED test_variant_consistency_eager_nn_functional_conv3d_cuda_complex64 (__main__.TestCommonCUDA)

  #114592 closed May 5, 2025

• [CPU][UT] 16 UT of test/inductor/test_cpu_select_algorithm.py failed with PyTorch 2025-04-028 nightly wheel

  #152398 closed May 5, 2025

• Error with nccl + multiple RTX5090 in ddp training. CUDA error: an illegal memory access was encountered

  #152780 closed May 5, 2025

• Kw argument `dtype` less relative with the functions themselves

  #145607 closed May 5, 2025

• cpu - gpu calculation results differs by far with torch.nn.functional.linear

  #69969 closed May 4, 2025

• Inconsistent behavior between CPU and GPU implementations of `torch.Tensor.put_` method

  #152755 closed May 4, 2025

• Performance Regression nightly 02/14→02/15, on nanogpt speedrun

  #152761 closed May 4, 2025

• Note some limit in docstring of `padding` in Poolnd

  #152156 closed May 4, 2025

• AOTInductor package can only be loaded on the first GPU (cuda:0) in C++ via AOTIModelPackageLoader

  #152087 closed May 4, 2025

• Make scaler.step() return if step was skipped or not

  #152279 closed May 3, 2025

• [Torch Profiler] Only two streams captured in CUDA graph but multiple streams shown in Torch Profiler

  #152114 closed May 3, 2025

• The device_id parameter of distributed.init_process_group will cause each process to occupy video memory on the first accessible GPU

  #149119 closed May 3, 2025

• [inductor][triton] Inductor is not compatible with the latest upstream Triton

  #152531 closed May 2, 2025

• Performance Regression nightly 2025/02/08→02/09, on nanogpt speedrun

  #147463 closed May 2, 2025

• DISABLED test_captured_scale_float16_cuda_float16 (__main__.TestFlexAttentionCUDA)

  #152083 closed May 2, 2025

• DISABLED test_builtin_score_mods_float32_score_mod4_cuda_float32 (__main__.TestFlexAttentionCUDA)

  #152082 closed May 2, 2025

• Update documentation to include insert and + methods to add layers in sequential

  #146892 closed May 2, 2025

• test_reference_numerics_normal fails with certain versions of numpy/scipy

  #148143 closed May 2, 2025

• inductor-periodic failures 5/2/2025

  #152691 closed May 2, 2025

• static cuda launcher causes `RuntimeError: CUDA driver error: invalid device context` in torchtitan CI

  #152639 closed May 2, 2025

• Triton Error [CUDA]: invalid device context when autograd.backward a triton kernel

  #124565 closed May 2, 2025

• DISABLED test_cat_max_autotune_triton (__main__.TestMaxAutotune)

  #145830 closed May 2, 2025

• DISABLED test_sparse_add_cuda_complex64 (__main__.TestSparseCSRCUDA)

  #145069 closed May 2, 2025

• Some Performance Bug in `tol` of `torch.lobpcg()`

  #152154 closed May 2, 2025

• DISABLED test_nvshmem

  #152649 closed May 2, 2025

• py_limited_api=True in PyTorch2.7 will break the build of extensions

  #152243 closed May 2, 2025

• [ONNX] Improve and sort out fallback mechanism

  #151703 closed May 2, 2025

• Should make the doc of `nn.CrossEntropyLoss()` more clear

  #134853 closed May 1, 2025

• torch.compile should not recompiles when `.requires_grad=True` under `torch.no_grad()` context

  #131975 closed May 1, 2025

• compiled autograd + dynamic shapes fails with constraint violation

  #133575 closed May 1, 2025

• [torch.compile] Dynamic shape behavior is different between using torch.compile with and without compiled_autograd.enable

  #113129 closed May 1, 2025

• Export QAT model is not performing as expected when compared to the original model and FX Graph QAT

  #150746 closed May 1, 2025

• `torch.export` fails on `InstanceNorm1d`

  #152467 closed May 1, 2025

• [CI] [anaconda] CI Perf Tests

  #148342 closed May 1, 2025

• [Inductor] Dynamo hangs when processing an operator, seemingly depending on a logical argument value

  #151743 closed May 1, 2025

• [export] Warn users when 0/1 specialization happens

  #151582 closed May 1, 2025

• The test 'test_host_memory_stats' is failing in torch2.7.0+cu118

  #152422 closed May 1, 2025

• How does torch.cudagraph capture a hybrid graph?

  #152584 closed May 1, 2025

• `torch.randint` can't handle large `high` argument (and in general high range of `torch.uint64`)

  #152564 closed Apr 30, 2025

• torch.randint should accept high=2**63

  #81446 closed Apr 30, 2025

• pytorch index_select is too slow

  #111247 closed Apr 30, 2025

• cuda graphs produce two additional kernel calls

  #143572 closed Apr 30, 2025

• [regression] Not getting `CUDA error: device-side assert triggered` on main for CUDA_KERNEL_ASSERT2

  #107396 closed Apr 30, 2025

• [CI] [anaconda] Benchmarks anaconda removal

  #152123 closed Apr 30, 2025

• More logs to show why fx graph cache isn't hit / created?

  #152065 closed Apr 30, 2025

• Mr

  #152549 closed Apr 30, 2025

• Add Description of `validate_args` in `torch.distributions.`

  #152165 closed Apr 30, 2025

• [ROCm] "No available kernel" when running EFFICIENT_ATTENTION sdpa

  #138864 closed Apr 30, 2025

• difficulty creating magma tarball when new rocm or cuda versions are deployed

  #151707 closed Apr 30, 2025

• [CUDA Graph tree] Cannot capture buffer allocation on side CUDA Streams

  #151199 closed Apr 30, 2025

151 Issues opened by 78 people

• [FlexAttention] export fails to trace with functorch

  #153063 opened May 7, 2025

• non-strict export should detect fake tensor leakage

  #153062 opened May 7, 2025

• register_constant doesn't work on simple types

  #153061 opened May 7, 2025

• DISABLED test_input_hooks_same (__main__.HooksTests)

  #153059 opened May 7, 2025

• `cuda.Event` handling in dynamo is broken

  #153058 opened May 7, 2025

• Export doesn't move embedding to correct device

  #153056 opened May 7, 2025

• Process never ends when sending tensors through multiprocessing queues in Python 3.12+ on macOS

  #153050 opened May 7, 2025

• DISABLED test_comprehensive_special_ndtri_cuda_int64 (__main__.TestInductorOpInfoCUDA)

  #153047 opened May 7, 2025

• DISABLED test_comprehensive_trunc_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #153046 opened May 7, 2025

• DISABLED test_hook_with_nested_closure (__main__.HooksTests)

  #153045 opened May 7, 2025

• Unexpected float32 overflow for amp training with torch.compile

  #153044 opened May 7, 2025

• Pytorch 2.7 crashes when using flex attention with torch.amp

  #153042 opened May 7, 2025

• using as_strided in torch compile generate wrong result.

  #153041 opened May 7, 2025

• Add Split Softmax

  #153035 opened May 7, 2025

• missalignment with differenet shape in F.linear with bf16 dtype

  #153033 opened May 7, 2025

• DISABLED test_hook_with_closure (__main__.HooksTests)

  #153032 opened May 7, 2025

• DISABLED test_comprehensive_svd_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #153031 opened May 7, 2025

• DISABLED test_comprehensive_amin_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #153030 opened May 7, 2025

• DISABLED test_comprehensive_asinh_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #153029 opened May 7, 2025

• Multiple CUDA graphs utilizing multiple CUDA GPUs encounter illegal memory access during replay

  #153025 opened May 7, 2025

• [RFC] Enable XPU+FlexAttention on Intel GPU

  #153024 opened May 7, 2025

• XPU inference output abnormal with device 'XPU:1'

  #153022 opened May 7, 2025

• [RFC][API-Unstable]Enable A16W4 on XPU Device

  #153019 opened May 7, 2025

• inconsistent grads between two types of `allgather`s

  #153016 opened May 7, 2025

• Operations on different precision tensors in CPU lead to different outputs

  #153014 opened May 7, 2025

• DISABLED test_comprehensive_scatter_xpu_bool (__main__.TestInductorOpInfoXPU)

  #153009 opened May 7, 2025

• DISABLED test_comprehensive_scatter_xpu_int64 (__main__.TestInductorOpInfoXPU)

  #153008 opened May 7, 2025

• `lintrunenr init` fails

  #152999 opened May 6, 2025

• DISABLED test_comprehensive_rsub_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152996 opened May 6, 2025

• [dynamo] Actually support functools.lru_cache

  #152994 opened May 6, 2025

• conv2d with int8 on CUDA: GET was unable to find an engine to execute this computation

  #152992 opened May 6, 2025

• `torch.load` can't deserialize `datetime` objects, even with the appropriate `safe_globals`

  #152985 opened May 6, 2025

• FPE when using `torch.lcm_` with int32 tensor and int16 scalar

  #152979 opened May 6, 2025

• Refactor MegaCache to make it generic

  #152976 opened May 6, 2025

• avoid falling back to as_strided for non-contiguous in-place reshape.

  #152972 opened May 6, 2025

• DISABLED test_comprehensive_scatter_xpu_int32 (__main__.TestInductorOpInfoXPU)

  #152971 opened May 6, 2025

• DISABLED test_comprehensive_gather_xpu_int64 (__main__.TestInductorOpInfoXPU)

  #152970 opened May 6, 2025

• [FSDP2] need dummy forward/backward to stay SPMD

  #152964 opened May 6, 2025

• DTensor support for dynamic shapes is soft

  #152963 opened May 6, 2025

• TestNestedTensorOpInfoCUDA.test_compile_backward_matmul_cuda_float32 Test Failure

  #152962 opened May 6, 2025

• DTensor placement propagation for `slice` fails during recompile due to SymInts

  #152954 opened May 6, 2025

• Remove redundant type aliases of _device for torch.Device

  #152952 opened May 6, 2025

• DISABLED test_compiler_collectives_automatic_dynamic_tensor (__main__.TestMultiProc)

  #152944 opened May 6, 2025

• DISABLED test_comprehensive_ormqr_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152943 opened May 6, 2025

• aten._scaled_dot_product_efficient_attention returns LSE padded to next highest multiple of 32

  #152942 opened May 6, 2025

• ROCm: no HIP device available if device is already initialized

  #152941 opened May 6, 2025

• DISABLED test_comprehensive_gather_xpu_bool (__main__.TestInductorOpInfoXPU)

  #152931 opened May 6, 2025

• DISABLED test_comprehensive_gather_xpu_int32 (__main__.TestInductorOpInfoXPU)

  #152930 opened May 6, 2025

• DISABLED test_comprehensive_gather_xpu_float16 (__main__.TestInductorOpInfoXPU)

  #152929 opened May 6, 2025

• DISABLED test_comprehensive_triu_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152928 opened May 6, 2025

• DISABLED test_comprehensive_rot90_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152927 opened May 6, 2025

• DISABLED test_comprehensive_scatter_xpu_float16 (__main__.TestInductorOpInfoXPU)

  #152925 opened May 6, 2025

• Enable 12.8.1

  #152922 opened May 6, 2025

• [inductor][cpu] pytorch_CycleGAN_and_pix2pix AMP multiple thread performance regression in 2025-04-27 nightly release

  #152921 opened May 6, 2025

• We should include where specialization happens when we throw a constraint violation error

  #152918 opened May 6, 2025

• UNSTABLE inductor / unit-test / cuda12.6-py3.10-gcc9-sm86 / test (inductor_cpp_wrapper)

  #152916 opened May 6, 2025

• Segmentation fault (core dumped) in torch.nn.functional.max_unpool2d

  #152913 opened May 6, 2025

• DISABLED test_comprehensive_scatter_xpu_float32 (__main__.TestInductorOpInfoXPU)

  #152912 opened May 6, 2025

• DISABLED test_comprehensive_gather_xpu_float32 (__main__.TestInductorOpInfoXPU)

  #152911 opened May 6, 2025

• DISABLED test_comprehensive_gather_xpu_float64 (__main__.TestInductorOpInfoXPU)

  #152910 opened May 6, 2025

• DISABLED test_comprehensive_scatter_xpu_float64 (__main__.TestInductorOpInfoXPU)

  #152898 opened May 6, 2025

• DISABLED test_comprehensive_nn_functional_conv3d_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152893 opened May 6, 2025

• DISABLED AotInductorTest.BasicPackageLoaderTestCpu (build.bin.test_aoti_inference)

  #152891 opened May 6, 2025

• DISABLED test_comprehensive_sort_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152892 opened May 6, 2025

• DISABLED test_comprehensive_diagonal_copy_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152890 opened May 6, 2025

• DISABLED AotInductorTest.BasicTestCpu (build.bin.test_aoti_inference)

  #152889 opened May 6, 2025

• DISABLED AotInductorTest.BasicTestCuda (build.bin.test_aoti_inference)

  #152888 opened May 6, 2025

• UNSTABLE Lint / Link checks / Lint URLs / linux-job

  #152884 opened May 6, 2025

• Motivate Pytorch's forward mode AD APIs with training examples

  #152877 opened May 5, 2025

• Incorporate CUDA Memory Trimming Into DeviceCachingAllocator

  #152875 opened May 5, 2025

• Docs Update `wrap_triton`

  #152870 opened May 5, 2025

• DISABLED testAssertNotRegex (__main__.CPythonTest_Assertions)

  #152869 opened May 5, 2025

• Have WrapTriton work w/ `TRITON_INTERPRET=1` in eager

  #152868 opened May 5, 2025

• [dynamo] Improve final traceback frame format

  #152867 opened May 5, 2025

• inductor-periodic rocm tests failing since at least 4/10

  #152866 opened May 5, 2025

• torch.cuda.use_mem_pool is not thread safe

  #152861 opened May 5, 2025

• DISABLED test_comprehensive___rmul___cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152851 opened May 5, 2025

• Clarification on build_stage usage with DistributedDataParallel: the code example in torch.distributed.piplining

  #152849 opened May 5, 2025

• Clean up CUTLASS_VERSION post cutlass version update

  #152847 opened May 5, 2025

• torch.distributions.Beta.entropy returns negative values

  #152845 opened May 5, 2025

• Don't hardcoded support for DTensor to_local/from_local/redistribute into dynamo

  #152829 opened May 5, 2025

• Pipeline Parallelism Fails when stage input does not produce gradients in all stages.

  #152827 opened May 5, 2025

• `mypy` stage of `lintrunner -a` has intermittent but continuing crashes

  #152824 opened May 5, 2025

• Performance Regression nightly 03/11→03/12, on nanogpt speedrun

  #152823 opened May 5, 2025

• TorchRun: Option to specify which GPUs to run on

  #152822 opened May 5, 2025

• Depthwise Separable Convolutions with Large Tensors (> 2**31) Elements) Fail Despite cuDNN 64-bit Indexing Support

  #152816 opened May 5, 2025

• Mismatch in dynamic quantization performance for torchao and torch.quantization

  #152813 opened May 5, 2025

• DISABLED test_comprehensive_fliplr_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152797 opened May 5, 2025

• DISABLED test_comprehensive_rot90_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #152796 opened May 5, 2025

• DISABLED test_comprehensive_unbind_copy_cuda_int32 (__main__.TestInductorOpInfoCUDA)

  #152795 opened May 5, 2025

• DISABLED test_comprehensive_linalg_pinv_singular_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152793 opened May 5, 2025

• DISABLED test_comprehensive_slice_scatter_cuda_bool (__main__.TestInductorOpInfoCUDA)

  #152794 opened May 5, 2025

• Inconsistent export behavior for nonzero+grid_sample between CUDA and CPU/MPS backends

  #152791 opened May 4, 2025

• [CXX11ABI] torch 2.6.0-cu126 and cu124 have different exported symbols

  #152790 opened May 4, 2025

• undefined symbol: __nvJitLinkCreate_12_8, version libnvJitLink.so.12

  #152783 opened May 4, 2025

• Segmentation fault (core dumped) in torch.nn.functional.alpha_dropout

  #152777 opened May 4, 2025

• RuntimeError: creation_meta == CreationMeta::DEFAULT INTERNAL ASSERT FAILED at "/build/pytorch/torch/csrc/autograd/variable.cpp":224, please report a bug to PyTorch.

  #152773 opened May 4, 2025

• Cuda-12.9 removed libnvToolsExt.so.* and is now purely header nvtx3

  #152756 opened May 3, 2025

• Checkpoint sequential doesn't raise clear error when segments is greater than number of functions

  #152752 opened May 3, 2025

• Error on padding 0-sized tensors

  #152750 opened May 3, 2025

• torch.compile causes stride mismatch in SDPA with non-contiguous query in torch 2.7

  #152747 opened May 3, 2025

• [FSDP2] fully_shard(mesh=(shard, shard)) for intra and inter node all-gathers

  #152746 opened May 3, 2025

• [MPS] TensorIterator and accuracy

  #152736 opened May 2, 2025

• Inconsistent float16 overflow behavior between CPU and CUDA devices

  #152731 opened May 2, 2025

• When scoped_libary is destroyed the fake impls are not cleared

  #152720 opened May 2, 2025

• [Mergebot] Adding ciflow/pull in PR without pull and lint workflows

  #152718 opened May 2, 2025

• Cannot mask a DTensor

  #152717 opened May 2, 2025

• dtensors TP+DP issues

  #152712 opened May 2, 2025

• [FSDP2] NO_SHARD as fully_shard(mesh=(Replicate, Shard)) with shard of world size 1

  #152710 opened May 2, 2025

• Gradient can be backpropagated through only certain distributions

  #152703 opened May 2, 2025

• MPS internal assertion with jacfwd and concatenation

  #152701 opened May 2, 2025

• DISABLED test_2d_mlp_with_nd_mesh (__main__.TestFullyShardNDTraining)

  #152700 opened May 2, 2025

• [CI] [anaconda] Triton windows build

  #152699 opened May 2, 2025

• CI workflows being skipped on PR

  #152697 opened May 2, 2025

• torch._foreach_pow(DTensor, float) and torch._foreach_pow_(DTensor, float) do not work

  #152696 opened May 2, 2025

• Add sm_86 (Ampere) and sm_89 (Ada) SASS in aarch64 builds

  #152690 opened May 2, 2025

• torch.library.custom_op string support

  #152685 opened May 2, 2025

• DISABLED test_comprehensive_select_scatter_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152684 opened May 2, 2025

• Flex attention strides

  #152683 opened May 2, 2025

• [RFC] Universal Device Context and Safe GPU/CPU Execution Decorators

  #152679 opened May 2, 2025

• DISABLED AotInductorTest.BasicPackageLoaderTestCuda (build.bin.test_aoti_inference)

  #152674 opened May 2, 2025

• DISABLED test_comprehensive_std_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152673 opened May 2, 2025

• DISABLED test_comprehensive_cummin_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152672 opened May 2, 2025

• DISABLED test_comprehensive_polygamma_polygamma_n_0_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152671 opened May 2, 2025

• Torch BF16 group gemm hangs in backward pass - core issue isolated, needs proper resolution.

  #152668 opened May 2, 2025

• DISABLED test_comprehensive_nansum_cuda_int32 (__main__.TestInductorOpInfoCUDA)

  #152666 opened May 2, 2025

• Add explicit error message for def infer_size(a, b): that specificy that non broadcast path was picked due to unbacked existing in both inputs.

  #152656 opened May 2, 2025

• UNSTABLE docker-cache-mi300 / docker-cache

  #152655 opened May 2, 2025

• Check for if two tensors are overall similar instead of bitwise similar?

  #152647 opened May 2, 2025

• ProcessGroupGloo.allgather_into_tensor_coalesced crashes with CUDA tensors

  #152645 opened May 1, 2025

• TestFlexAttentionCUDA.test_GQA_score_mod7_cuda_float16 fails on h100

  #152635 opened May 1, 2025

• Incorrect strides for `nonzero_static` compilation

  #152634 opened May 1, 2025

• DISABLED test_torchvision_models_efficientnet_v2_l (__main__.TestVisionTracing)

  #152632 opened May 1, 2025

• [v2.7.1] Release Tracker

  #152627 opened May 1, 2025

• modded-nanogpt flaky NCCL hang starting 3/30 nightly

  #152623 opened May 1, 2025

• Pytorch Profiler crashes while using it with Pytorch Lightning module

  #152617 opened May 1, 2025

• Enable AOTI for Metal inductor

  #152612 opened May 1, 2025

• [triton pin update] Run Inductor CI on pin updates for Triton and the PyTorch nightly branch

  #152608 opened May 1, 2025

• Loops impacting output when utilizing hooks

  #152607 opened May 1, 2025

• AOTI regression on SAM and tts-angular

  #152606 opened May 1, 2025

• ROCm, 7900 XTX: Pytorch FLASH_ATTENTION SDPA is 2.5x slower than MATH (fp16, head_dim 256, seqlen 4360, 12 heads)

  #152595 opened May 1, 2025

• Flex Attention doesn't scale with custom bias

  #152593 opened May 1, 2025

• [ratter-build] Cannot detect CUDA when build from source

  #152592 opened May 1, 2025

• [Benchmark] High compilation time variance on benchmark dashboards

  #152566 opened Apr 30, 2025

• DISABLED test_graph_partition_reorder_cpu_and_gpu_interleave (__main__.CudaGraphTreeTests)

  #152561 opened Apr 30, 2025

• DISABLED test_pending_fusion_pro_and_epi (__main__.TestPrologueFusion)

  #152560 opened Apr 30, 2025

• DISABLED test_comprehensive_signal_windows_hamming_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #152559 opened Apr 30, 2025

• DISABLED test_comprehensive_amin_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152558 opened Apr 30, 2025

• PGO does not work on jobs for frameworks that copy code to different dirs at different attempts.

  #152555 opened Apr 30, 2025

• MPS varying seq len SDPA memory leak

  #152550 opened Apr 30, 2025

• FakeTensorUpdater does not trace nodes correctly

  #152548 opened Apr 30, 2025

534 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.

• auto functionalize base_hop

  #151067 commented on May 6, 2025 • 22 new comments

• [Accelerator] Fix Python typing in accelerator

  #152394 commented on May 6, 2025 • 16 new comments

• Random Batch Sampler Speedup

  #147706 commented on May 6, 2025 • 14 new comments

• [1/n][Optimus][Auto-AC] Support activation quantization without scaling

  #148380 commented on May 1, 2025 • 12 new comments

• [Intel GPU] Support f32 intermediate dtype, headdim size <=576 and f32 causal mask for SDPA

  #152091 commented on May 7, 2025 • 11 new comments

• convert guard_size_oblivious to runtime check in infer_size_impl

  #148872 commented on May 6, 2025 • 11 new comments

• Cache code generation during triton template expansion and enable it for mm_template.

  #151773 commented on May 5, 2025 • 9 new comments

• [cp] dispatch flex_attention to CP impl in TorchDispatchMode

  #151497 commented on May 3, 2025 • 9 new comments

• Mini tutorial for provenance tracking

  #152211 commented on May 6, 2025 • 8 new comments

• [Inductor-CPU] Faster int8 WoQ GEMM for small M with explicit prefetching

  #149373 commented on May 7, 2025 • 8 new comments

• [Cutlass] Integrate EVT into CUDACPPScheduling

  #150906 commented on May 6, 2025 • 6 new comments

• [pytree] Add public pytree module `torch.utils.pytree`

  #137400 commented on May 3, 2025 • 6 new comments

• [WIP] DeadCodeEliminator Mark(block) improvement

  #152348 commented on May 2, 2025 • 5 new comments

• [Cutlass] Changes to gemm template for EVT

  #150907 commented on May 7, 2025 • 5 new comments

• [device_mesh] improve device selection logic

  #150897 commented on May 1, 2025 • 4 new comments

• [ROCm][Inductor][CK] Add ck-tile based universal gemm kernels to torch.mm autotune choices

  #152341 commented on May 7, 2025 • 4 new comments

• Fix `lr_scheduler` unexpectedly calls `step()` when init argument last_epoch is larger than -1

  #149312 commented on May 6, 2025 • 3 new comments

• [FP8][CUTLASS] xFail `honor_sm_carveout` on `sm100`

  #152378 commented on May 6, 2025 • 3 new comments

• [Hierarchical Compilation] Use universal flatten APIs

  #152505 commented on May 5, 2025 • 3 new comments

• `torch.tensordot`: performance improvements when contracting to a scalar.

  #145936 commented on May 5, 2025 • 3 new comments

• Add differentiable ops hint message in Module docs

  #150291 commented on May 6, 2025 • 2 new comments

• [BE]: Follow detach().clone() pattern for SGD

  #144468 commented on May 5, 2025 • 2 new comments

• [Quant][X86] add ops to compute uint8 pointwise add/add_relu

  #152411 commented on May 7, 2025 • 2 new comments

• Temp test

  #148424 commented on May 6, 2025 • 2 new comments

• [CI][CUDA] Move cu118 distributed pull jobs to cu126, move cu124-sm75 to cu126-sm75

  #151594 commented on May 2, 2025 • 2 new comments

• [pytorch][triton] flex attention fwd kernel with TMA loads (#151923)

  #152460 commented on May 7, 2025 • 2 new comments

• cpu: enable gemm-bf16f32 for SDPA BF16

  #140159 commented on May 7, 2025 • 2 new comments

• remove guard_size_oblivious from unbind.

  #148815 commented on May 2, 2025 • 2 new comments

• [export] Refactor pt2 save/load

  #152495 commented on May 6, 2025 • 2 new comments

• autograd: Add VJP and JVP rules for aten::aminmax

  #151186 commented on May 3, 2025 • 2 new comments

• Horizontal

  #151780 commented on May 7, 2025 • 1 new comment

• Move prologue_supported_inputs computations to def_kernal

  #150869 commented on May 2, 2025 • 1 new comment

• [Graph Partition] Pass all cudagraph tree tests

  #152048 commented on May 7, 2025 • 1 new comment

• Move mps_linear forward to use MPS kernels directly instead of MPSGraph

  #152210 commented on May 6, 2025 • 1 new comment

• [dynamic shapes] guard_or_false for infer_size

  #152146 commented on May 6, 2025 • 1 new comment

• update get_default_device to also respect torch.device ctx manager

  #148621 commented on May 6, 2025 • 1 new comment

• [inductor] lowering for fractional_max_pool3d

  #148630 commented on May 7, 2025 • 1 new comment

• Fix `InstanceNorm` wrong suggestion in warning message

  #151534 commented on May 2, 2025 • 1 new comment

• Change unsafe_marked_cacheable_functions to a dictionary, so that you can specify a static cache key

  #152486 commented on May 6, 2025 • 1 new comment

• Parallelize sort using libstdc++ parallel mode

  #150195 commented on May 1, 2025 • 1 new comment

• add device generalisation support for distributed tests

  #152471 commented on May 2, 2025 • 1 new comment

• [WIP][dynamic shapes] rewrite should_swap with guard_or_false

  #150164 commented on May 2, 2025 • 1 new comment

• [aotd] Support saved tensors hooks in aot_autograd

  #150032 commented on May 6, 2025 • 1 new comment

• Deprecate DataLoader pin_memory_device param

  #146821 commented on May 6, 2025 • 1 new comment

• [torchgen] Refactor `torchgen.utils.FileManager` to accept `pathlib.Path`

  #150726 commented on May 2, 2025 • 1 new comment

• Make `Adam`, `AdamW` work with nonzero-dim Tensor betas

  #149939 commented on May 3, 2025 • 1 new comment

• Add is_pinned to host allocator

  #151439 commented on May 6, 2025 • 1 new comment

• DRAFT: Add TMA opt for concat function target hopper and blackwell arch

  #149893 commented on May 7, 2025 • 1 new comment

• Skip fuse attention on fp32 if not tf32

  #151924 commented on May 5, 2025 • 1 new comment

• Use gather in index_select

  #151715 commented on May 7, 2025 • 1 new comment

• Inductor logging + analysis of torch.profile

  #149697 commented on May 3, 2025 • 1 new comment

• [ca] mark scalar int sizes as dynamic via tensor wrapping

  #151731 commented on May 7, 2025 • 1 new comment

• [DTensor] enable SimpleFSDP's composability with Tensor Parallel

  #152286 commented on May 6, 2025 • 1 new comment

• Rewrite autograd producer consumer stream sync logic

  #151079 commented on May 7, 2025 • 1 new comment

• Fix CUPTI lookup to include target directory

  #148668 commented on May 5, 2025 • 0 new comments

• [SGD] Add SGD capturable API and tests

  #148647 commented on May 5, 2025 • 0 new comments

• [dynamic shapes] guard_or_false for computeStorageNbytes

  #150483 commented on May 6, 2025 • 0 new comments

• Adjust CMake code for Eigen

  #148628 commented on May 7, 2025 • 0 new comments

• Optimize AOTInductor: Caching, Reduced Decompositions, and Improved JSON Handling

  #148616 commented on May 2, 2025 • 0 new comments

• [BE][pytree] cleanup parameterized pytree tests

  #148569 commented on May 5, 2025 • 0 new comments

• [triton hash update] update the pinned triton hash

  #148492 commented on May 7, 2025 • 0 new comments

• [BE][pytree] rename argument name in register function to match the type annotations: `*_fn -> *_func`

  #148484 commented on May 2, 2025 • 0 new comments

• [BE][pytree] rename `NodeDef` member to match the type annotations: `*_fn -> *_func`

  #148474 commented on May 2, 2025 • 0 new comments

• Remove `torch.testing` from `MOD_SKIPLIST`

  #148459 commented on May 5, 2025 • 0 new comments

• Add 'x in {...}' patterns to perf_linter

  #148417 commented on May 4, 2025 • 0 new comments

• Add perf_linter to auto-fix some anti-patterns

  #148416 commented on May 3, 2025 • 0 new comments

• softmax: add device check for xpu with half_to_float

  #150278 commented on May 7, 2025 • 0 new comments

• Add cmake variable USE_ROCM_CK

  #150245 commented on May 6, 2025 • 0 new comments

• [c10d] Test multiple CUDA Graph captures

  #150040 commented on May 7, 2025 • 0 new comments

• Fixes detection of ArmPL on Linux platform

  #150031 commented on May 7, 2025 • 0 new comments

• AOTI freezing: fix test issues and enable by default

  #149961 commented on May 2, 2025 • 0 new comments

• [inductor] Add typing to _inductor/ir.py

  #149958 commented on May 7, 2025 • 0 new comments

• Enable XPU distributed test for PT2.8

  #149916 commented on May 5, 2025 • 0 new comments

• Refactoring FSDP2 (_composable/fsdp) test cases to be device agnostic

  #149848 commented on May 5, 2025 • 0 new comments

• Add SWA with a cyclical scheduler example

  #149847 commented on May 2, 2025 • 0 new comments

• Add x86-simd-sort accelerated sorting

  #149362 commented on May 7, 2025 • 0 new comments

• [cuDNN][SDPA] cuDNN SDPA refactor/cleanup, nested tensor backward, test priority bump for `sm90`, `sm100`

  #149282 commented on May 1, 2025 • 0 new comments

• Make Subset dataset a true wrapper

  #149272 commented on May 5, 2025 • 0 new comments

• [Easy] update pip sources for CUDA in nightly pull tool

  #149143 commented on May 1, 2025 • 0 new comments

• Update the heuristic for AArch64 bmm/baddbmm

  #149122 commented on May 7, 2025 • 0 new comments

• [test] bigger runnner

  #149003 commented on Apr 30, 2025 • 0 new comments

• [Inductor] Record Triton’s Base32 Cache Key in .best_config for Debugging

  #148981 commented on May 6, 2025 • 0 new comments

• Move token linter code into tools/linter/adaptors/_linter/

  #148959 commented on May 5, 2025 • 0 new comments

• Fix AttributeError for `_get_vc_env` with setuptools>=75.9.0

  #148847 commented on May 7, 2025 • 0 new comments

• C++ support to print symbolic tensors as `Symbolic tensor: size=(...)`

  #148846 commented on May 4, 2025 • 0 new comments

• Set specialized representation string for meta/fake tensor with empty construction

  #148794 commented on May 7, 2025 • 0 new comments

• cpp_wrapper: build non-performance-sensitive code at O1

  #148773 commented on May 2, 2025 • 0 new comments

• Trunk workflow for Windows Arm64

  #148753 commented on May 2, 2025 • 0 new comments

• Fix calling torch.compile inside of a `__torch_dispatch__`

  #148712 commented on May 6, 2025 • 0 new comments

• [Just SCRTCH] no review

  #148710 commented on May 6, 2025 • 0 new comments

• Automated perf_linter changes: x in (...)

  #148415 commented on May 4, 2025 • 0 new comments

• Define USE_C10D_XCCL and USE_XCCL in pytorch

  #147593 commented on May 7, 2025 • 0 new comments

• [import][inductor] Simplify grid handling

  #147583 commented on May 3, 2025 • 0 new comments

• [ONNX][demo] Rotary embedding

  #147576 commented on May 4, 2025 • 0 new comments

• Update pybind11 submodule to 3.0.0-dev test

  #147524 commented on Apr 30, 2025 • 0 new comments

• removed zero dim cpu logic from fake_tensor.py

  #147501 commented on May 1, 2025 • 0 new comments

• [test] sccache log

  #147470 commented on May 6, 2025 • 0 new comments

• [Inductor] Avoid tensor slice overflow for large step

  #147433 commented on May 6, 2025 • 0 new comments

• [Inductor][CPP] Add float16 support for CppMicroGemmAMX

  #147368 commented on May 6, 2025 • 0 new comments

• Add the memory and dispatch to the logging module.

  #147262 commented on May 3, 2025 • 0 new comments

• Fix clang-tidy warnings in torch/jit

  #147253 commented on May 2, 2025 • 0 new comments

• logging: close handler after removing it

  #147235 commented on May 4, 2025 • 0 new comments

• Record the XPU and XCCL build settings in the compiled binary

  #147161 commented on May 7, 2025 • 0 new comments

• [fsdp] add an experimental allocator hook for buffers that participate in collective communication

  #147146 commented on May 5, 2025 • 0 new comments

• fake_tensor: Handle op errors more gracefully

  #147049 commented on May 7, 2025 • 0 new comments

• experimental proposal DCP v2

  #146999 commented on May 2, 2025 • 0 new comments

• [BE]: Try to remove unused type ignores - attempt 1

  #146989 commented on May 2, 2025 • 0 new comments

• [DO NOT MERGE][cuDNN][SDPA] Testing sm90/sm100 priority for cuDNN SDPA

  #146947 commented on May 6, 2025 • 0 new comments

• Support pin_memory() during CUDA stream capture.

  #146924 commented on May 7, 2025 • 0 new comments

• [DO NOT MERGE] ROCm sandbox PR

  #146903 commented on May 7, 2025 • 0 new comments

• Fix non-bitwise type annotations for Tensor operators (see #145838)

  #146845 commented on May 1, 2025 • 0 new comments

• Enable Windows tests

  #146695 commented on May 4, 2025 • 0 new comments

• Optimize LRScheduler docs

  #146684 commented on May 2, 2025 • 0 new comments

• [HOP] Mutation and alias rework

  #146658 commented on May 7, 2025 • 0 new comments

• [WIP] BaseSubclass

  #146612 commented on May 5, 2025 • 0 new comments

• clang-format CUDASymmetricMemory.cu

  #146592 commented on May 5, 2025 • 0 new comments

• Update quantile doc

  #146485 commented on May 2, 2025 • 0 new comments

• [WIP][dynamic shapes] mark backed size symbols as size-like

  #146335 commented on May 5, 2025 • 0 new comments

• [dcp] Minor improvements to filesystem writer

  #146273 commented on May 6, 2025 • 0 new comments

• Format tests by PYFMT

  #146267 commented on May 2, 2025 • 0 new comments

• docs: change log to ln in Softplus function and class

  #146199 commented on May 2, 2025 • 0 new comments

• Automated perf_linter changes: list constructors

  #148414 commented on May 4, 2025 • 0 new comments

• Automated perf_linter changes: generators

  #148413 commented on May 3, 2025 • 0 new comments

• Add api info for torch._C._nn.pyi [1/N]

  #148410 commented on May 3, 2025 • 0 new comments

• Enable `_lazy_clone` between CPU and MPS

  #148408 commented on May 1, 2025 • 0 new comments

• [Utilization] Add utilization monitor for linux build

  #148375 commented on May 6, 2025 • 0 new comments

• test index_put

  #148357 commented on May 3, 2025 • 0 new comments

• [ROCm][CI] Add support for gfx1100 in rocm workflow + test skips

  #148355 commented on May 7, 2025 • 0 new comments

• [pytree] simplify public API exposition with `__module__`

  #148328 commented on May 3, 2025 • 0 new comments

• Checking for cuda version to see if bf16 is natively supported or emulated

  #148322 commented on May 6, 2025 • 0 new comments

• ```torch.as_strided``` negative stride SIGSEV fix when using ```torch.compile```

  #148301 commented on May 3, 2025 • 0 new comments

• Treat CUDA warnings as errors

  #148294 commented on May 3, 2025 • 0 new comments

• handle jk for emulation runs

  #148240 commented on May 2, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `torch/ao/` to `ruff format`

  #148185 commented on May 1, 2025 • 0 new comments

• [pytree] add another simplified pytree module `torch.pytree`

  #148180 commented on May 3, 2025 • 0 new comments

• [Don't merge]Upgrade submodule oneDNN to v3.7 (#147498)(ZI)

  #148173 commented on May 4, 2025 • 0 new comments

• Disable cudnn to avoid creating guards that denies exporting

  #148140 commented on May 4, 2025 • 0 new comments

• Checks kv pair indexing in OrderedPreservingDictTest.test_range_insert

  #148136 commented on May 5, 2025 • 0 new comments

• [torch/elastic][upstream] Fix the wrong order when start_index is not 0

  #147967 commented on May 4, 2025 • 0 new comments

• [Don't merge]Upgrade submodule oneDNN to v3.7 (#147498)(Zi)

  #147917 commented on May 2, 2025 • 0 new comments

• [Draft] Enable cpu_offload for _distribute_state_dict

  #147916 commented on May 2, 2025 • 0 new comments

• [export][dynamic shapes] add Dim._OBLIVIOUS, _mark_oblivious()

  #147881 commented on May 2, 2025 • 0 new comments

• Set disable_clone=True when running opt_gm

  #147845 commented on May 2, 2025 • 0 new comments

• Use /permissive- for torch libraries in MSVC builds

  #147825 commented on May 4, 2025 • 0 new comments

• remove asserttion in expand_to_full_mesh_op_strategy

  #147823 commented on May 6, 2025 • 0 new comments

• [WIP][ptd][nccl] use current-stream as nccl-stream under async=False mode

  #147820 commented on May 2, 2025 • 0 new comments

• [cuda] Added a correctness test for layernorm backwards

  #147763 commented on May 5, 2025 • 0 new comments

• [DCP][OSS] Rank local checkpointing in DCP without collectives

  #147758 commented on Apr 30, 2025 • 0 new comments

• Modifications to RuntimeEstimator and SACEstimator

  #147750 commented on May 6, 2025 • 0 new comments

• Skip test_dtypes xpu test on bmm and addbmm

  #147721 commented on May 3, 2025 • 0 new comments

• [Dtensor] Pass device information in OffsetBasedRNGTracker

  #147594 commented on May 4, 2025 • 0 new comments

• Use std::apply for CPU code

  #152526 commented on May 2, 2025 • 0 new comments

• Add `padding="same"` for transposed convolution

  #152228 commented on May 5, 2025 • 0 new comments

• Add support for torch.cuda.FloatTensor()

  #152208 commented on May 1, 2025 • 0 new comments

• [submodule] Update ONNX to 1.18

  #152200 commented on May 7, 2025 • 0 new comments

• [inductor] propagate shapes in CSEVariable

  #152198 commented on May 7, 2025 • 0 new comments

• IGNORE: Testing OIDC

  #152181 commented on May 1, 2025 • 0 new comments

• Extend compute_global_tensor_shape to multi dimension sharding

  #152166 commented on May 3, 2025 • 0 new comments

• Add dynamo config to HOP-ify context managers

  #152159 commented on Apr 30, 2025 • 0 new comments

• Add runtime asserts to AOTI

  #152125 commented on May 6, 2025 • 0 new comments

• [dynamo][ca] support dynamic annotations on tensors in ListVariables/TupleVariables

  #152119 commented on May 7, 2025 • 0 new comments

• Update _torch_docs.py to Fix torch.bernoulli()

  #152104 commented on May 5, 2025 • 0 new comments

• Switch to standard pep517 sdist generation

  #152098 commented on May 6, 2025 • 0 new comments

• unbreak fb:operator_benchmark_test

  #152049 commented on May 1, 2025 • 0 new comments

• [map] always turn on dynamo for map

  #152041 commented on May 6, 2025 • 0 new comments

• Add CPython complex tests

  #152015 commented on May 6, 2025 • 0 new comments

• [Kineto] Upgrade the kineto commit to fb36cce

  #152007 commented on May 7, 2025 • 0 new comments

• [UniformValueConstantFolder] deduce value on CPU rather than on device

  #151998 commented on May 6, 2025 • 0 new comments

• Add torchcheck for replication_pad3d_backward

  #151986 commented on May 2, 2025 • 0 new comments

• Make `aten.embedding` do not wrap negative index

  #151967 commented on May 5, 2025 • 0 new comments

• [ca] hide unused scalar int sizes from dynamo

  #151962 commented on May 7, 2025 • 0 new comments

• [ROCm][CI] Update dockerfile to use centos9

  #151929 commented on May 6, 2025 • 0 new comments

• [BE] Upgrade XPU support package to 2025.1 in CICD

  #151899 commented on May 7, 2025 • 0 new comments

• Avoid differing results in `linalg.(tensor_)solve`

  #151896 commented on May 6, 2025 • 0 new comments

• [aot][ca] save bw_module in AOTAutogradCache

  #151860 commented on May 7, 2025 • 0 new comments

• [reland][ROCm] remove caffe2 from hipify

  #151845 commented on May 6, 2025 • 0 new comments

• [2/n][Optimus][Auto-AC] Support activation quantization with scaling

  #151770 commented on May 1, 2025 • 0 new comments

• Add adaptive_avg_pool2d input and output_size check

  #151769 commented on May 2, 2025 • 0 new comments

• [Don't merge] Upgrade oneDNN to v3.8 for XPU build

  #151767 commented on May 7, 2025 • 0 new comments

• Implement avg_pool3d for MPS backend

  #151742 commented on May 6, 2025 • 0 new comments

• [ROCm] Maxpool forward NHWC Perf Improvement targeting Resnet scenarios

  #151727 commented on May 7, 2025 • 0 new comments

• elastic: do not shutdown rendezvous on leaving workers

  #152525 commented on May 2, 2025 • 0 new comments

• [compile async] [cache] testing

  #152523 commented on May 7, 2025 • 0 new comments

• [inductor] [compile async] Don't compile in eager

  #152507 commented on May 6, 2025 • 0 new comments

• [Hierarchical Compile] Take into account mutation deps in cycle detection

  #152506 commented on May 1, 2025 • 0 new comments

• Fix flaky test in test_custom_ops

  #152484 commented on May 6, 2025 • 0 new comments

• [IR] Input Adapter refactor prototype

  #152459 commented on Apr 30, 2025 • 0 new comments

• fix: Update padding_mode to use Literal for type checking

  #152458 commented on May 2, 2025 • 0 new comments

• Add epoch to fake tensor cache key

  #152453 commented on May 2, 2025 • 0 new comments

• [ROCm] cpp_extension allow user to override default flags

  #152432 commented on May 4, 2025 • 0 new comments

• Relax tolerance for test_quick_baddbmm_cpu_complex64

  #152424 commented on May 6, 2025 • 0 new comments

• [Inductor][CPP] Enable vectorized fp8 quant dequant

  #152418 commented on Apr 30, 2025 • 0 new comments

• [Hierarchical Compile] Add mutation dependencies to topological sorting

  #152410 commented on May 1, 2025 • 0 new comments

• [Hierarchical Compilation] Track node mutations

  #152389 commented on May 1, 2025 • 0 new comments

• Add vec_reduce_all specialization for std::plus on AArch64

  #152388 commented on Apr 30, 2025 • 0 new comments

• fix: outdated contents in dynamo overview

  #152382 commented on May 6, 2025 • 0 new comments

• complex.pow(2) on GPU by replacing with complex * complex to avoid numerical instability

  #152373 commented on May 2, 2025 • 0 new comments

• [Relandx2] Rewrite the guts of torch::jit::Lexer to speed it up

  #152372 commented on May 1, 2025 • 0 new comments

• vec::map: directly process reduced-precision floats when reasonable

  #152366 commented on Apr 30, 2025 • 0 new comments

• add is_vec_specialized_for

  #152365 commented on May 2, 2025 • 0 new comments

• Format all headers under ATen/cpu/vec, not just top-level

  #152364 commented on May 6, 2025 • 0 new comments

• Add codeowner for merge rules

  #152354 commented on May 6, 2025 • 0 new comments

• [inductor][dynamo] Include operator name in size/stride/alignment assertion

  #152353 commented on May 7, 2025 • 0 new comments

• [cp] dispatch flex_attention_backward to CP impl in TorchDispatchMode

  #152311 commented on May 2, 2025 • 0 new comments

• Enable the AMP precision with freezing for CPU nightly test

  #152298 commented on May 6, 2025 • 0 new comments

• [CI] Add xpu inductor test into periodic workflow

  #152281 commented on May 7, 2025 • 0 new comments

• [ROCm] Maxpool backward NHWC Perf Improvement targeting Resnet scenarios

  #152267 commented on Apr 30, 2025 • 0 new comments

• [executorch hash update] update the pinned executorch hash

  #152238 commented on May 7, 2025 • 0 new comments

• At least one of ROCM_HOME or CUDA_HOME must be None

  #152236 commented on Apr 30, 2025 • 0 new comments

• _get_total_norm should use float64 to avoid rounding errors

  #152234 commented on May 1, 2025 • 0 new comments

• flex attention: fix dispatch order for tensor subclasses, avoid hardcoding call to faketensor impl in dynamo

  #151719 commented on May 2, 2025 • 0 new comments

• Do not cover up `__dunder`__ method type-hints from `.pyi` file

  #150875 commented on May 6, 2025 • 0 new comments

• Add CPython tests for iter/sort

  #150797 commented on May 6, 2025 • 0 new comments

• Add CPython generator/contextlib tests

  #150796 commented on May 6, 2025 • 0 new comments

• Add CPython int/float tests

  #150795 commented on May 6, 2025 • 0 new comments

• Add CPython math/cmath tests

  #150794 commented on May 6, 2025 • 0 new comments

• Add CPython string tests

  #150793 commented on May 6, 2025 • 0 new comments

• [Set] Add CPython set tests

  #150792 commented on May 7, 2025 • 0 new comments

• Add CPython dict tests

  #150791 commented on May 6, 2025 • 0 new comments

• Add CPython list/tuple tests

  #150790 commented on May 6, 2025 • 0 new comments

• Add CPython exception tests

  #150789 commented on May 6, 2025 • 0 new comments

• Add CPython tests for unittest

  #150788 commented on May 6, 2025 • 0 new comments

• Make device check error message more descriptive

  #150750 commented on May 7, 2025 • 0 new comments

• [BE][CI][Easy] Run `lintrunner` on generated `.pyi` stub files

  #150732 commented on May 2, 2025 • 0 new comments

• [BE] Resolve lint errors in `.pyi` stub files

  #150731 commented on May 2, 2025 • 0 new comments

• [BE] Ensure generated stub files by `gen_pyi` are properly formatted

  #150730 commented on May 2, 2025 • 0 new comments

• [BE] Add `__all__` to `torch/nn/functional.pyi` and `torch/return_types.pyi`

  #150729 commented on May 2, 2025 • 0 new comments

• [BE] Update `.pyi` stub template to use Generic TypeAlias (PEP 585) and Union Type (PEP 604)

  #150728 commented on May 2, 2025 • 0 new comments

• [torchgen] Refactor and simplify `gen_pyi.py` to use Generic TypeAlias (PEP 585) and Union Type (PEP 604)

  #150727 commented on May 2, 2025 • 0 new comments

• Avoid overwriting COW data in MPS code

  #150721 commented on May 2, 2025 • 0 new comments

• [export] add runtime assert messages to python torch checks

  #150719 commented on May 6, 2025 • 0 new comments

• Support XPU in memory tracker

  #150703 commented on May 7, 2025 • 0 new comments

• [draft][distributed] add into 3d composability test at AMD CI test

  #150694 commented on May 6, 2025 • 0 new comments

• AOTI: add all fallback ops that are missing from C-shim

  #150673 commented on May 2, 2025 • 0 new comments

• [Inductor] Fix CUDA memory usage for CPU only compile

  #150669 commented on May 2, 2025 • 0 new comments

• Refactor `torch/utils/data/datapipes/gen_pyi.py` with `torchgen`

  #150626 commented on May 2, 2025 • 0 new comments

• Make LazyModuleMixin materialize after load_state_dict

  #150593 commented on May 2, 2025 • 0 new comments

• fix dynamic shapes for kwargs

  #150583 commented on Apr 30, 2025 • 0 new comments

• Enable lazy cloning in `Tensor.to` between CPU and MPS

  #150569 commented on May 2, 2025 • 0 new comments

• API change for new enum in cusparseltsplitkmode-t for cusparseLT 0.7.0+

  #150536 commented on May 5, 2025 • 0 new comments

• Add a custom profiler configuration option

  #151656 commented on May 6, 2025 • 0 new comments

• [Intel GPU] Use user-friendly err msg in mm

  #151655 commented on May 7, 2025 • 0 new comments

• [Intel GPU][Inductor] Fallback embedding_dense_backward on XPU

  #151637 commented on May 7, 2025 • 0 new comments

• Add OIDC perms to windows-[build|test] workflows

  #151596 commented on May 1, 2025 • 0 new comments

• Add OIDC permissions to linux-test workflow

  #151585 commented on May 1, 2025 • 0 new comments

• Add OIDC permissions to linux-build workflow

  #151581 commented on May 1, 2025 • 0 new comments

• [bazel] Fix unusual reference to cpuinfo workspace

  #151578 commented on May 5, 2025 • 0 new comments

• Add device agnostic support for distributed tests

  #151560 commented on May 6, 2025 • 0 new comments

• Update OpenBLAS commit

  #151547 commented on May 6, 2025 • 0 new comments

• Add hint message when parameters is empty in clip_grad_norm_

  #151529 commented on May 6, 2025 • 0 new comments

• Allow to byteswap data when reading saved torch jit data

  #151447 commented on May 1, 2025 • 0 new comments

• [ez] Rewrite comment to be more friendly to non haskellers

  #151421 commented on May 4, 2025 • 0 new comments

• [Cutlass] Add epilogue inputs/outputs to def_kernel

  #151406 commented on May 5, 2025 • 0 new comments

• [ROCm] Upgrade ROCm CI to ROCm6.4

  #151368 commented on May 7, 2025 • 0 new comments

• Use Allocator API raw_allocate & raw_dealloc in CUDAAllocator

  #151305 commented on May 6, 2025 • 0 new comments

• [WIP] Generalize device caching allocator

  #151298 commented on May 6, 2025 • 0 new comments

• Remove outdated Android workarounds of nearbyintf

  #151292 commented on May 3, 2025 • 0 new comments

• [dynamo] Avoid unnecessary `.detach()` call in `_make_subclass` polyfill

  #151265 commented on May 6, 2025 • 0 new comments

• [aot autograd][logging] Profile large missing gaps in compile time tracing

  #151256 commented on May 6, 2025 • 0 new comments

• NCCL: Fix cmake file when cross compiling.

  #151234 commented on May 5, 2025 • 0 new comments

• [dynamo] keep C++ symbolic shape guards disabled for benchmarks

  #151225 commented on May 1, 2025 • 0 new comments

• Implement MKLGenerator

  #151218 commented on May 1, 2025 • 0 new comments

• Update slow tests

  #151207 commented on May 5, 2025 • 0 new comments

• [dynamo] Prevent lazy variable realization on STORE_FAST

  #151184 commented on May 2, 2025 • 0 new comments

• TESTING: IGNORE

  #151116 commented on May 1, 2025 • 0 new comments

• Update auto-tuning support for _scaled_grouped_mm

  #150944 commented on May 6, 2025 • 0 new comments

• [CI] Enable XCCL in XPU CI build

  #150927 commented on May 7, 2025 • 0 new comments

• fix shard tensor gather when a local tensor on certain ranks has zero elements

  #150914 commented on May 7, 2025 • 0 new comments

• [torch.compile] handle a custom __delattr__ method correctly

  #150899 commented on May 5, 2025 • 0 new comments

• An exported onnx model can't reduce on dim with value of 0 if 'keepdims' is false

  #66200 commented on May 4, 2025 • 0 new comments

• Install pytorch from pypi using local CUDA build

  #150742 commented on May 4, 2025 • 0 new comments

• pytorch pip install instructions: always include the cuda index

  #150432 commented on May 4, 2025 • 0 new comments

• [compile] DDPOptimizer + activation checkpointing not supported

  #104674 commented on May 4, 2025 • 0 new comments

• [ROCm] PyTorch slow on TTS

  #150168 commented on May 4, 2025 • 0 new comments

• torch.lobpcg producing different largest eigenvalue than scipy and np.linalg.eig

  #101075 commented on May 4, 2025 • 0 new comments

• Inconsistent `sum`/`dot`/`norm` behavior

  #151761 commented on May 5, 2025 • 0 new comments

• Logging when executing fx.Interpreter

  #117351 commented on May 5, 2025 • 0 new comments

• DISABLED test_cublas_addmm_reduced_precision_fp16_accumulate_size_10000_cuda_float16 (__main__.TestMatmulCudaCUDA)

  #151661 commented on May 5, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_int8 (__main__.TestForeachCUDA)

  #150407 commented on May 5, 2025 • 0 new comments

• [C10D] Allow NCCL single P2P ops to use parent/collective communicator

  #152220 commented on May 5, 2025 • 0 new comments

• Unexpected behavior when using dist.all_reduce(x, op=dist.ReduceOp.SUM)

  #152300 commented on May 5, 2025 • 0 new comments

• can't reconstruct the communication group using PyTorch.

  #152527 commented on May 5, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_inplace_cuda_complex64 (__main__.TestForeachCUDA)

  #150960 commented on May 5, 2025 • 0 new comments

• DISABLED test_cublas_addmm_reduced_precision_fp16_accumulate_size_1000_cuda_float16 (__main__.TestMatmulCudaCUDA)

  #151675 commented on May 5, 2025 • 0 new comments

• General MPS op coverage tracking issue

  #77764 commented on May 5, 2025 • 0 new comments

• DISABLED test_comprehensive_lu_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #152520 commented on May 5, 2025 • 0 new comments

• The 2.7.0 release tarball is missing `.ci/docker/ci_commit_pins/nccl-cu12.txt` required for building

  #152532 commented on May 1, 2025 • 0 new comments

• Newly added lint-urls jobs are very flaky

  #152439 commented on May 2, 2025 • 0 new comments

• associative_scan not composable with vmap in eager-mode

  #134000 commented on May 2, 2025 • 0 new comments

• Multiple Learning Rate Scheduler for Specific Parameters Groups

  #101082 commented on May 2, 2025 • 0 new comments

• Adam (fused=True) issues

  #90752 commented on May 2, 2025 • 0 new comments

• Unwanted Warning in lr_scheduler.step()

  #117540 commented on May 2, 2025 • 0 new comments

• DISABLED test_comprehensive_bitwise_right_shift_cuda_int32 (__main__.TestInductorOpInfoCUDA)

  #152057 commented on May 2, 2025 • 0 new comments

• DISABLED test_comprehensive_nansum_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #139710 commented on May 2, 2025 • 0 new comments

• AdamW(fused=True) slower than unfused AdamW

  #121857 commented on May 2, 2025 • 0 new comments

• DISABLED test_comprehensive_floor_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152058 commented on May 2, 2025 • 0 new comments

• DISABLED test_comprehensive_nansum_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #140693 commented on May 2, 2025 • 0 new comments

• `einsum` is about 40x slower on CUDA than manually multiplying and summing

  #101249 commented on May 2, 2025 • 0 new comments

• torch.nn.functional.ctc_loss raises cuDNN error in PyTorch versions >=2.5.0

  #152421 commented on May 3, 2025 • 0 new comments

• Deprecation of NVTX 2 (`nvToolsExt`): Recommended to move to NVTX 3

  #147011 commented on May 3, 2025 • 0 new comments

• A problem discovered when computing complex matrices in deep neural networks

  #151182 commented on May 3, 2025 • 0 new comments

• [dynamic shapes] data-dependent error when backed + unbacked expression resolves statically

  #151491 commented on May 3, 2025 • 0 new comments

• MPS operator coverage tracking issue (2.6+ version)

  #141287 commented on May 4, 2025 • 0 new comments

• [dynamo] guard code generation triggers attribute error on DeviceMesh object

  #152447 commented on May 5, 2025 • 0 new comments

• [PT2] torch.layer_norm errors in eager but runs fine in backend=aot_eager_decomp_partition

  #151478 commented on May 5, 2025 • 0 new comments

• [inductor] [aot] `torch.linalg.lu` can't accept `slice operation`, behaving differently with eager

  #151401 commented on May 5, 2025 • 0 new comments

• [dynamo] Try tracing into einops

  #152480 commented on May 5, 2025 • 0 new comments

• [dynamo] `torch.compile` prevents fsdp warning from getting generated

  #152451 commented on May 5, 2025 • 0 new comments

• `torch.compile` causes assertion error in distributed checkpoint wrapper test

  #152442 commented on May 5, 2025 • 0 new comments

• [inductor] Improve codegen for argmax+max

  #146643 commented on May 5, 2025 • 0 new comments

• TORCH_COMPILE_DEBUG=1 does not consistently generate debug logs

  #152374 commented on May 5, 2025 • 0 new comments

• [ued] Slow start up time for `torch.compile` on GGUF Auraflow

  #150706 commented on May 5, 2025 • 0 new comments

• [inductor] [assertion error] `torch.select_scatter` crashes on inductor but passes on eager

  #151296 commented on May 5, 2025 • 0 new comments

• DISABLED AotInductorTest.FreeInactiveConstantBufferRuntimeConstantFoldingCuda (build.bin.test_aoti_inference)

  #150299 commented on May 5, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_inplace_cuda_float16 (__main__.TestForeachCUDA)

  #150985 commented on May 5, 2025 • 0 new comments

• Cannot override __add__ in NamedTuple with __new__ + torch.compile

  #133762 commented on May 5, 2025 • 0 new comments

• `randint(max)` causes a graph break, but not `rand().mul(max).floor().to(torch.long)` (on CPU)

  #135664 commented on May 5, 2025 • 0 new comments

• Investigate FlexAttention performance degradation on low precision inputs

  #147336 commented on May 5, 2025 • 0 new comments

• [AOTAutograd] tweak min-cut partitioner to avoid saving softmax output

  #126348 commented on May 6, 2025 • 0 new comments

• upstream `apex.normalization.FusedRMSNorm`

  #72643 commented on May 5, 2025 • 0 new comments

• [XPU] Upgrade the XPU support packages version to 2025.1 in CI/CD

  #151097 commented on May 5, 2025 • 0 new comments

• RFC: Torch Native Runtime

  #152034 commented on May 5, 2025 • 0 new comments

• [TensorDict - compile] dynamo generator compatibility

  #129658 commented on May 5, 2025 • 0 new comments

• DISABLED test_binary_op_with_scalar_self_support__foreach_pow_is_fastpath_True_cuda_float32 (__main__.TestForeachCUDA)

  #149409 commented on May 5, 2025 • 0 new comments

• [discussion] "TensorList" as first-class python-world abstraction (to mirror existing dispatcher's `TensorList` and `torch::autograd::tensor_list`) and as key for dispatch for merging `torch._foreach_*` into regular `torch.*` functions

  #100249 commented on May 5, 2025 • 0 new comments

• RFC: The State of Custom CUDA extensions in PyTorch

  #152032 commented on May 5, 2025 • 0 new comments

• Illegal Instruction Caused by `grid_sample` Under Windows

  #152385 commented on May 5, 2025 • 0 new comments

• optree package status in PyTorch

  #152535 commented on May 5, 2025 • 0 new comments

• DISABLED test_pattern_matcher_multi_user_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)

  #134433 commented on May 5, 2025 • 0 new comments

• DISABLED test_cublas_addmm_reduced_precision_fp16_accumulate_size_100_cuda_float16 (__main__.TestMatmulCudaCUDA)

  #151712 commented on May 5, 2025 • 0 new comments

• Update quantization to make source files complient with /Zc:lambda

  #92600 commented on May 5, 2025 • 0 new comments

• Stop special-casing einops in Dynamo

  #142486 commented on May 5, 2025 • 0 new comments

• DISABLED test_comprehensive_index_select_cuda_int32 (__main__.TestInductorOpInfoCUDA)

  #152416 commented on May 5, 2025 • 0 new comments

• DISABLED test_comprehensive_polygamma_polygamma_n_0_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152469 commented on May 5, 2025 • 0 new comments

• DISABLED test_comprehensive_polygamma_polygamma_n_1_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #152470 commented on May 5, 2025 • 0 new comments

• DISABLED test_comprehensive_repeat_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #152500 commented on May 5, 2025 • 0 new comments

• [Async TP] all-gather-matuls not fusing properly when rowwise scales are used

  #149990 commented on May 1, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_float16 (__main__.TestForeachCUDA)

  #150173 commented on May 1, 2025 • 0 new comments

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

  #145374 commented on May 1, 2025 • 0 new comments

• Ability to do aot/inductor compilation from a jit model (or torch.exported model)

  #127928 commented on May 1, 2025 • 0 new comments

• DISABLED test_int64_upsample3d_cuda_bfloat16 (__main__.TestTorchDeviceTypeCUDA)

  #146007 commented on May 1, 2025 • 0 new comments

• Update `torch/nn/modules/conv.py` to use Literal for support padding modes

  #152280 commented on May 1, 2025 • 0 new comments

• AOTI packaged model fails with generic error when run in for loop but succeeds on individual sample

  #146524 commented on May 1, 2025 • 0 new comments

• `nn.CrossEntropyLoss` accepts negative target probabilities

  #152437 commented on May 1, 2025 • 0 new comments

• RuntimeError: "_amp_foreach_non_finite_check_and_unscale_cuda" not implemented for 'BFloat16'

  #127176 commented on May 1, 2025 • 0 new comments

• DISABLED test_remove_noop_view_default_cpu (__main__.CpuTests)

  #151512 commented on May 1, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_float32 (__main__.TestForeachCUDA)

  #150208 commented on May 1, 2025 • 0 new comments

• DISABLED test_binary_op_with_scalar_self_support__foreach_pow_is_fastpath_True_cuda_complex128 (__main__.TestForeachCUDA)

  #149323 commented on May 1, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_inplace_cuda_complex128 (__main__.TestForeachCUDA)

  #150933 commented on May 1, 2025 • 0 new comments

• DTensor slicing on sharded dimension leads to replication

  #149447 commented on May 1, 2025 • 0 new comments

• Status of pip wheels with _GLIBCXX_USE_CXX11_ABI=1

  #51039 commented on May 1, 2025 • 0 new comments

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

  #145371 commented on May 2, 2025 • 0 new comments

• [ROCm] sdpa group query attention bf16 numeric error

  #139352 commented on Apr 30, 2025 • 0 new comments

• Profiler doesn't seem to work on AMD CPUs

  #150052 commented on Apr 30, 2025 • 0 new comments

• MPS: Conv1d fails with NotImplementedError for output_channels > 65536

  #152278 commented on Apr 30, 2025 • 0 new comments

• [ROCm] MI300X FP8 scaled_mm is extremely slow on nightly

  #143465 commented on Apr 30, 2025 • 0 new comments

• torch.compile on MPS progress tracker

  #150121 commented on Apr 30, 2025 • 0 new comments

• NotImplementedError: Could not run 'aten::index.Tensor' with arguments from the 'SparseCUDA' backend.

  #152226 commented on Apr 30, 2025 • 0 new comments

• Training/Fine-tuning fails with PyTorch 2.8 + 4x 5090 GPUs using DDP/FSDP/DeepSpeed

  #150734 commented on Apr 30, 2025 • 0 new comments

• enhance documentation around the developer build

  #108406 commented on Apr 30, 2025 • 0 new comments

• Quantile is limited to 16 million elements and have poor performance.

  #64947 commented on Apr 30, 2025 • 0 new comments

• missing docs for torch.Tag

  #126518 commented on Apr 30, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_complex128 (__main__.TestForeachCUDA)

  #150141 commented on Apr 30, 2025 • 0 new comments

• DISABLED test_is_isnot (__main__.TestScript)

  #120694 commented on May 1, 2025 • 0 new comments

• DISABLED test_remove_noop_slice_cpu (__main__.CpuTests)

  #151384 commented on May 1, 2025 • 0 new comments

• DISABLED test_inductor_all_gather_into_tensor_coalesced (__main__.CompileTest)

  #146806 commented on May 1, 2025 • 0 new comments

• DISABLED test_comprehensive_nanmean_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #140339 commented on May 1, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_complex64 (__main__.TestForeachCUDA)

  #150161 commented on May 1, 2025 • 0 new comments

• [Tracker] Nested tensor op coverage requests

  #118107 commented on May 1, 2025 • 0 new comments

• Signature should be extended for `torch.hamming_window()`

  #146590 commented on May 2, 2025 • 0 new comments

• DISABLED test_comprehensive_special_xlog1py_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #140648 commented on May 2, 2025 • 0 new comments

• DISABLED AotInductorTest.FreeInactiveConstantBufferCuda (build.bin.test_aoti_inference)

  #149495 commented on May 2, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_int32 (__main__.TestForeachCUDA)

  #150350 commented on May 2, 2025 • 0 new comments

• [inductor] nan_asserts doesn't work for FP8, "RuntimeError: "isinf" not implemented for 'Float8_e4m3fn'"

  #149002 commented on May 2, 2025 • 0 new comments

• Major perf regression with `BatchNorm2d` + `torch.compile` with `reduce-overhead` + DDP

  #139207 commented on May 2, 2025 • 0 new comments

• DISABLED test_binary_op_with_scalar_self_support__foreach_pow_is_fastpath_True_cuda_complex64 (__main__.TestForeachCUDA)

  #149199 commented on May 2, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_int64 (__main__.TestForeachCUDA)

  #150392 commented on May 2, 2025 • 0 new comments

• module.cuda() doesn't work under FakeTensorMode

  #148977 commented on May 2, 2025 • 0 new comments

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

  #141958 commented on May 2, 2025 • 0 new comments

• [CI] [anaconda] CI Build and Test scripts Windows

  #148338 commented on May 2, 2025 • 0 new comments

• Continuous calls to nn.Linear in fp32 on the 5090D cause severe performance degradation

  #150725 commented on May 2, 2025 • 0 new comments

• Label tracking meta-issue (edit me to get automatically CC'ed on issues! cc bot)

  #24422 commented on May 2, 2025 • 0 new comments

• Pytorch DDP across nodes: self._store = TCPStore( # type: ignore[call-arg] RuntimeError: Stop_waiting response is expected

  #114357 commented on May 2, 2025 • 0 new comments

• [CUDA][Compex] `test_reference_numerics_large_jiterator_unary_cuda_complex64` broken after updating to `numpy >= 1.25.0`

  #125198 commented on May 2, 2025 • 0 new comments

• fully_shard() for huggingface 72B model: pytorch caches too much GPU memory

  #151936 commented on May 2, 2025 • 0 new comments

• Dynamo unsupported: dynamic padding

  #123855 commented on May 1, 2025 • 0 new comments

• When using torch to convert to oxxn model, testing the inference results with actual images shows tensor mismatch

  #152097 commented on May 1, 2025 • 0 new comments

• dynamo cannot trace global op_set .__contains__

  #145761 commented on May 1, 2025 • 0 new comments

• [RFC] : Dynamically Quantized 8-bit Matrix Multiplication support

  #149500 commented on May 1, 2025 • 0 new comments

• LoadHIP.cmake should find_package(composable_kernel)

  #149809 commented on May 1, 2025 • 0 new comments

• `view()` + modify-in-place fails silently with DTensor

  #147570 commented on May 1, 2025 • 0 new comments

• DISABLED test_remove_noop_view_default_cuda (__main__.GPUTests)

  #151511 commented on May 1, 2025 • 0 new comments

• Context Parallel -- unsharded output doesn't match output without CP.

  #152261 commented on May 1, 2025 • 0 new comments

• [Feature request] Exclusive prefix sum, `torch.cumsum(input, dim=0, exclusive=True)`

  #76191 commented on May 1, 2025 • 0 new comments

• The state of sparse Tensors

  #9674 commented on May 1, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_float64 (__main__.TestForeachCUDA)

  #150298 commented on May 2, 2025 • 0 new comments

• Add description of several params in the basic usage of `torch.min()`, `torch.max()`, `torch.all()` and `torch.any()`

  #152176 commented on May 2, 2025 • 0 new comments

• Raise an Error when File Not Found in `torch.jit.load()`

  #152178 commented on May 2, 2025 • 0 new comments

• DISABLED test_remove_noop_view_dtype_cuda (__main__.GPUTests)

  #151541 commented on May 2, 2025 • 0 new comments

• DISABLED test_remove_noop_view_dtype_cpu (__main__.CpuTests)

  #151540 commented on May 2, 2025 • 0 new comments

• DISABLED test_foreach_copy_with_multi_dtypes__foreach_copy_cuda_int16 (__main__.TestForeachCUDA)

  #150309 commented on May 2, 2025 • 0 new comments

• [NCCL] Unordered destruction of `ProcessGroupNCCL` no longer supported

  #137507 commented on May 2, 2025 • 0 new comments

• [inductor] enable bf32 test for mkldnn conv

  #127293 commented on May 7, 2025 • 0 new comments

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

  #127294 commented on May 7, 2025 • 0 new comments

• [1/N] Update CI jobs to use CMake >= 3.25

  #130522 commented on May 6, 2025 • 0 new comments

• Make IPC features extendable on third-party devices

  #133222 commented on May 6, 2025 • 0 new comments

• add ranking for grouped benchmarks

  #133287 commented on May 2, 2025 • 0 new comments

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

  #136835 commented on May 6, 2025 • 0 new comments

• Add TORCH_CHECK_INDEX in convert_indices_from_coo_to_csr_cpu

  #138068 commented on May 6, 2025 • 0 new comments

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

  #138202 commented on May 2, 2025 • 0 new comments

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

  #138214 commented on May 5, 2025 • 0 new comments

• [WIP] Add DeviceAllocator as the base device allocator

  #138222 commented on May 6, 2025 • 0 new comments

• Always produce XML

  #138513 commented on May 6, 2025 • 0 new comments

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

  #139699 commented on May 5, 2025 • 0 new comments

• Fix warnings and simplify code in TensorShape

  #141971 commented on May 4, 2025 • 0 new comments

• Fix platform detection in MKLDNN CMake file

  #142067 commented on May 2, 2025 • 0 new comments

• [Draft][WIP] Enable XPU path for FlexAttention

  #143553 commented on May 7, 2025 • 0 new comments

• Replacing explicit backend search with api call

  #144944 commented on May 5, 2025 • 0 new comments

• Wrong formula for CosineAnnealingLR

  #152081 commented on May 6, 2025 • 0 new comments

• `Aborted (core dumped)` in `torch.cuda.nccl.reduce`

  #150836 commented on May 7, 2025 • 0 new comments

• [RFC] zentorch Integration

  #150296 commented on May 7, 2025 • 0 new comments

• DISABLED test_comprehensive_signal_windows_general_cosine_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #139682 commented on May 7, 2025 • 0 new comments

• [inductor] [silent incorrectness] `torch.nn.PairwiseDistance(p=2)` outputs incorrect results with eager

  #151198 commented on May 7, 2025 • 0 new comments

• PyTorch VS2022 official build Windows binary illegal instruction on AVX2(max ISA level) CPU

  #145702 commented on May 7, 2025 • 0 new comments

• DISABLED test_foreach_check_stride_ignore_dims_of_one_cuda_float32 (__main__.TestForeachCUDA)

  #150026 commented on May 7, 2025 • 0 new comments

• `setup.py develop` command is disappearing soon from `setuptools`

  #152276 commented on May 7, 2025 • 0 new comments

• ImportError: dlopen: cannot load any more object with static TLS

  #2575 commented on May 7, 2025 • 0 new comments

• [ATen][Sparse] Use Third-Party Eigen for sparse addmm

  #101814 commented on May 5, 2025 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on May 2, 2025 • 0 new comments

• [pytree] support PyStructSequence types for Python pytree

  #113258 commented on May 5, 2025 • 0 new comments

• Automated submodule update: FBGEMM

  #115316 commented on May 7, 2025 • 0 new comments

• [vision hash update] update the pinned vision hash

  #125806 commented on May 7, 2025 • 0 new comments

• refine fp32 precision api

  #125888 commented on May 7, 2025 • 0 new comments

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

  #126050 commented on May 7, 2025 • 0 new comments

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

  #126054 commented on May 7, 2025 • 0 new comments

• [test] fix unit test

  #144977 commented on May 2, 2025 • 0 new comments

• removed check for ConvTranspose3D on MPS

  #145366 commented on May 7, 2025 • 0 new comments

• Open up PT UTs to cover additional devices

  #145589 commented on May 6, 2025 • 0 new comments

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

  #145594 commented on May 5, 2025 • 0 new comments

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

  #145595 commented on May 5, 2025 • 0 new comments

• [c10d] implement ReduceOp.unbox()

  #145652 commented on May 5, 2025 • 0 new comments

• Avoid data-dependent errors by runtime assert substitution.

  #145681 commented on May 2, 2025 • 0 new comments

• [Easy] update pip sources for ROCm in nightly pull tool

  #145685 commented on May 1, 2025 • 0 new comments

• [Async-TP] Port _fused_all_gather_matmul_native to cpp to reduce launching overhead

  #145794 commented on May 5, 2025 • 0 new comments

• [AsyncMM] preliminary tuning

  #145795 commented on May 5, 2025 • 0 new comments

• [Async-TP] _pipelined_multi_all_gather_and_consume reduce overhead

  #145796 commented on May 5, 2025 • 0 new comments

• [Async-TP] improve algo selection

  #145797 commented on May 5, 2025 • 0 new comments

• [will-not-merge] tuning

  #145798 commented on May 5, 2025 • 0 new comments

• Replace distutils.version with copied looseversion

  #145819 commented on May 6, 2025 • 0 new comments

• [CUDAEvent.h] support external cuda events in cudagraphs

  #146145 commented on May 5, 2025 • 0 new comments

• [CI] Get rid of UCC builds

  #146173 commented on May 2, 2025 • 0 new comments

• Add where_ ops

  #143636 commented on May 6, 2025 • 0 new comments

• Defaults to C++20 in CMake torch targets

  #143959 commented on May 7, 2025 • 0 new comments

• [Intel GPU] add tf32 support for matmul on XPU

  #144240 commented on May 7, 2025 • 0 new comments

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

  #144293 commented on May 2, 2025 • 0 new comments

• [pytree][1/N] change pytree usages to implementation agnostic: `torch.distributed`

  #144332 commented on May 2, 2025 • 0 new comments

• [TorchInductor] Add ALiBi (Attention with Linear Biases) Fused Attention Pattern

  #144338 commented on May 5, 2025 • 0 new comments

• [BE][pytree][Easy] change imports `torch.utils._pytree` -> `torch.utils.pytree.python`

  #144405 commented on May 2, 2025 • 0 new comments

• Remove the `_stacklevel` arg from `log_softmax`, `softmax` and `softmin`

  #144451 commented on May 6, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `{torch,test}/{nn,optim}/**` to `ruff format`

  #144548 commented on May 1, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `torch/_[a-h]*/` to `ruff format`

  #144551 commented on May 1, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `torch/[p-z]*/` to `ruff format`

  #144552 commented on May 1, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `torch/[e-n]*/` to `ruff format`

  #144553 commented on May 1, 2025 • 0 new comments

• [BE][PYFMT] migrate PYFMT for `torch/[a-c]*/` to `ruff format`

  #144554 commented on May 1, 2025 • 0 new comments

• [BE][PYFMT] remove `black`: finish `black -> ruff format` migration

  #144557 commented on May 1, 2025 • 0 new comments

• [Intel CPU] Fix issue #143482.

  #144760 commented on May 2, 2025 • 0 new comments

• [Intel CPU] Fix issue #143483.

  #144854 commented on May 2, 2025 • 0 new comments

• [export] check non-negative modulus, avoid unnecessary congruences, in export solver

  #144925 commented on May 3, 2025 • 0 new comments

• DISABLED test_comprehensive_rot90_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #140773 commented on May 6, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_inplace_cuda_float64 (__main__.TestForeachCUDA)

  #151019 commented on May 6, 2025 • 0 new comments

• DISABLED test_cublas_addmm_size_1000_cuda_bfloat16 (__main__.TestMatmulCudaCUDA)

  #151834 commented on May 6, 2025 • 0 new comments

• Semi-Structured Sparsity unsupported for Windows

  #125302 commented on May 6, 2025 • 0 new comments

• Add switch to disable truncation to long list print

  #152427 commented on May 6, 2025 • 0 new comments

• DISABLED test_per_sample_api_compute_batch_size_not_pytreeable_cpu (__main__.TestExpandedWeightModuleCPU)

  #146972 commented on May 6, 2025 • 0 new comments

• DISABLED test_cublas_addmm_size_1000_cuda_float16 (__main__.TestMatmulCudaCUDA)

  #151862 commented on May 6, 2025 • 0 new comments

• Segmentation error for torch==2.2.1 on MacOs

  #121101 commented on May 6, 2025 • 0 new comments

• Expand Tag Set: views & reductions

  #129020 commented on May 6, 2025 • 0 new comments

• at::BlasBackend::Ck does not handle all ROCm BLAS gpus

  #150187 commented on May 6, 2025 • 0 new comments

• compile generates inefficient code for mutations on small slices of inputs

  #152346 commented on May 6, 2025 • 0 new comments

• DeepSeek: mixed precision optimizers (BF16AdamW)

  #146542 commented on May 6, 2025 • 0 new comments

• DISABLED test_inductor_all_gather_into_tensor_single (__main__.CompileTest)

  #147707 commented on May 6, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_outplace_cuda_bfloat16 (__main__.TestForeachCUDA)

  #151054 commented on May 6, 2025 • 0 new comments

• DISABLED test_binary_op_with_scalar_self_support__foreach_pow_is_fastpath_True_cuda_uint8 (__main__.TestForeachCUDA)

  #149858 commented on May 6, 2025 • 0 new comments

• [t.compile][Functools] Cache decorator support for dynamo

  #146598 commented on May 6, 2025 • 0 new comments

• Attributeless FakeRootModule

  #135696 commented on May 5, 2025 • 0 new comments

• torch.onnx.export causes floating point exception with core dump for empty slice assignment

  #110056 commented on May 5, 2025 • 0 new comments

• torch.export with dynamic shapes on Static Cache HF LLama model fails

  #152465 commented on May 5, 2025 • 0 new comments

• torch.matrix_exp gets stuck on GPU

  #149335 commented on May 6, 2025 • 0 new comments

• welfordreduce slows down forward layernorm in a bunch of cases

  #120184 commented on May 6, 2025 • 0 new comments

• dist.barrier() hangs after calling async_save

  #123447 commented on May 6, 2025 • 0 new comments

• [inductor] `proxy_tensor.py` throws `SyntaxError` when using `.random_`

  #151432 commented on May 6, 2025 • 0 new comments

• [inductor] [silence] `nn.ConvTranspose2d-F.dropout` outputs inconsistent results with eager

  #148061 commented on May 6, 2025 • 0 new comments

• [inductor] [cuda] [fake tensor] `torch.triu_indices` throws `pointer argument` error when using `[0, 0]`

  #151737 commented on May 6, 2025 • 0 new comments

• Device Error on vmap

  #151591 commented on May 6, 2025 • 0 new comments

• Support Delay Loading of c10.dll in when using libtorch as a thirdparty library.

  #105058 commented on May 6, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_inplace_cuda_float32 (__main__.TestForeachCUDA)

  #151003 commented on May 6, 2025 • 0 new comments

• [dynamo] Dynamo fails to run torch.cat() with FakeTensors because it can't confirm 's0 + s1*u0' is nonzero

  #152473 commented on May 6, 2025 • 0 new comments

• [DTensor] Calling .item() on DTensor with Partial placement results in local value

  #152406 commented on May 6, 2025 • 0 new comments

• Simplification of pruned models

  #58846 commented on May 6, 2025 • 0 new comments

• torch.compile fails in FSDP due to .data assignment with different floating type

  #152162 commented on May 6, 2025 • 0 new comments

• TorchInductor CPU Performance Dashboard

  #93531 commented on May 6, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_outplace_cuda_complex128 (__main__.TestForeachCUDA)

  #151093 commented on May 6, 2025 • 0 new comments

• inductor `full_like` decompositions give incorrect strides

  #144699 commented on May 6, 2025 • 0 new comments

• [CI] No workflows scheduled on PRs

  #151322 commented on May 6, 2025 • 0 new comments

• NCCL out of memory error after updating to PyTorch 2.7

  #152302 commented on May 6, 2025 • 0 new comments

• [ued] HF diffusers pipeline `enable_cpu_offload` errors or graph breaks with a `torch.compile`-ed transformer

  #150711 commented on May 6, 2025 • 0 new comments

• Question about that support of torch.compile for a custom CUDA operator?

  #152270 commented on May 7, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_outplace_cuda_complex64 (__main__.TestForeachCUDA)

  #151099 commented on May 7, 2025 • 0 new comments

• AttributeError: type object 'torch._C._distributed_c10d.BackendType' has no attribute 'XCCL'.

  #147059 commented on May 7, 2025 • 0 new comments

• Windows inductor genarated code without function declaration, and compile failed on MSVC.

  #152251 commented on May 7, 2025 • 0 new comments

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

  #142135 commented on May 7, 2025 • 0 new comments

• `context_parallel` fails for training with `RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation`

  #149306 commented on May 7, 2025 • 0 new comments

• Request to cherrypick a fix into v1.13.1 (v1.8 has a CVE)

  #98115 commented on May 7, 2025 • 0 new comments

• `RuntimeError: UR error` with XPU

  #149953 commented on May 7, 2025 • 0 new comments

• DISABLED test_slice_scatter_reinplace_cuda (__main__.GPUTests)

  #145189 commented on May 7, 2025 • 0 new comments

• DISABLED test_cublas_and_lt_reduced_precision_fp16_accumulate_cuda (__main__.TestMatmulCudaCUDA)

  #151890 commented on May 7, 2025 • 0 new comments

• DISABLED test_parity__foreach_acos_fastpath_outplace_cuda_float16 (__main__.TestForeachCUDA)

  #151114 commented on May 7, 2025 • 0 new comments

• [Manylinux 2.28] Migrate Docker container to use gcc 13, CUDA 12.6 and gcc14 CUDA 12.8

  #152426 commented on May 6, 2025 • 0 new comments

• [dynamo] Replace `unimplemented` with `unimplemented_v2`

  #147913 commented on May 6, 2025 • 0 new comments

• [dynamo] torch._dynamo crashes on `self.value.__module__` inside SkipFunctionVariable.call_function() (PyTorch 2.7, works 2.6)

  #152316 commented on May 6, 2025 • 0 new comments

• `torch.compile()` produces incorrect results for `asinh_()` operation on large/small values

  #152299 commented on May 6, 2025 • 0 new comments

• Unusually slow draft_export time

  #152337 commented on May 6, 2025 • 0 new comments

• Silent incorrectness between static torch.compile vs eager

  #152425 commented on May 6, 2025 • 0 new comments

• Softmax Decomp Causes Incorrect Gradients when Using `torch.compile` with `F.multi_head_attention_forward`

  #152309 commented on May 6, 2025 • 0 new comments

• RMS norm causes NaNs when used with torch.compile + float8 with rowwise scales

  #150859 commented on May 6, 2025 • 0 new comments

• aot_eager produces wrong output with all_gather_tensor_autograd

  #148701 commented on May 6, 2025 • 0 new comments

• Significant precision error from torch.compile

  #145213 commented on May 6, 2025 • 0 new comments

• Composition of torch.compile and torch.func.grad silently produces a wrong result.

  #136662 commented on May 6, 2025 • 0 new comments

• Fx Graph cache hit generates guards that does not exists in the original cached program causing recompilations only at cache hit.

  #152435 commented on May 6, 2025 • 0 new comments

• Invalid handling of nans in compiled torch.quantile / torch.nanquantile on cuda

  #152423 commented on May 6, 2025 • 0 new comments

• OptimizedModule __getattr__ may causes dead recursive call loop

  #138157 commented on May 6, 2025 • 0 new comments

• Graph break on .t() when Tensor._make_subclass

  #151771 commented on May 6, 2025 • 0 new comments

• torch.compile on MPS fails: generated Metal kernel uses loop-local variable out of scope

  #152155 commented on May 6, 2025 • 0 new comments

• MPS Error on sequoia 15.3: NDArray dimension length > INT_MAX'

  #146769 commented on May 6, 2025 • 0 new comments