1 Release published by 1 person

• v2.5.1 PyTorch 2.5.1: bug fix release

  published Oct 29, 2024

2 Pull requests merged by 1 person

• Bump rexml from 3.3.3 to 3.3.9 in /ios/TestApp

  #139088 merged Oct 28, 2024

• Bump certifi from 2024.2.2 to 2024.7.4 in /tools/build/bazel

  #130173 merged Oct 28, 2024

213 Pull requests opened by 112 people

• [hop free symbols] replace ctx.save_for_backward to support symints/ints

  #138737 opened Oct 23, 2024

• Switch back to the default checkout action

  #138739 opened Oct 23, 2024

• [inductor] Allow unbacked symint fallback in stride reordering

  #138741 opened Oct 23, 2024

• Raise error for int64 and bool dtypes in nanmean, even for empty tensors

  #138745 opened Oct 23, 2024

• Fix custom obj being treated as input in export

  #138749 opened Oct 23, 2024

• [SymmetricMemory] expose signal_pads as tensors in Python

  #138754 opened Oct 23, 2024

• [SymmetricMemory] introduce a binding for cuMemset32Async

  #138755 opened Oct 23, 2024

• [inductor] don't fuse two nodes if likely increase peak memory

  #138756 opened Oct 23, 2024

• Create ciflow/inductor-periodic

  #138763 opened Oct 23, 2024

• [export] Add support for symbool to make it usable for torch.cond

  #138765 opened Oct 23, 2024

• config: create Config objects with JK support

  #138766 opened Oct 23, 2024

• justknobs: Remove JustKnobsConfig and justknobs_feature

  #138767 opened Oct 23, 2024

• [ONNX] Use TracedONNXFunction op signature to promote inputs to tensors

  #138770 opened Oct 23, 2024

• [dynamo] add SymNode bitwise and/or

  #138777 opened Oct 24, 2024

• AOTI Minifier demo

  #138780 opened Oct 24, 2024

• Add BFloat16 support and use a new pack method for flash attention forward kernel

  #138783 opened Oct 24, 2024

• [pytorch][inductor] Adding tf32x3 to matmul

  #138785 opened Oct 24, 2024

• Added option to control number of kernel options displayed

  #138788 opened Oct 24, 2024

• Add SwiGLU Activation class

  #138790 opened Oct 24, 2024

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

  #138798 opened Oct 24, 2024

• Use clang18 image for ASAN workflows

  #138801 opened Oct 24, 2024

• [Intel GPU] Add device guard for XPU structured operator in torchgen

  #138802 opened Oct 24, 2024

• Fix max_unpool2d op signature with stride and padding inputs

  #138805 opened Oct 24, 2024

• [inductor] support linear+binary foldinig for freezing path

  #138807 opened Oct 24, 2024

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

  #138810 opened Oct 24, 2024

• dynamo: guard on FSDP module parameters

  #138819 opened Oct 24, 2024

• [Testing] update XLA pin

  #138825 opened Oct 24, 2024

• [Testing] Update torchbench pin

  #138826 opened Oct 24, 2024

• Prioritize building with libgomp over libomp

  #138834 opened Oct 24, 2024

• [cond] make cond not throw warnings on constant pred in eager mode

  #138837 opened Oct 24, 2024

• Move inner loop of _create_symbolic_sizes_strides_storage_offset into its own method

  #138843 opened Oct 24, 2024

• Simplify _compute_symbolic_stride()

  #138844 opened Oct 24, 2024

• [export] preserve & respect dynamic decorators on input tensors

  #138850 opened Oct 24, 2024

• Improvements for associative_scan - slicing of xs

  #138858 opened Oct 24, 2024

• (wip) [CUTLASS] pull out static shapes as kernel args

  #138859 opened Oct 24, 2024

• Changed calls to yaml.load to explicitly use the safe loader.

  #138861 opened Oct 24, 2024

• [Pipelining] Relax multi-stage constraint

  #138862 opened Oct 24, 2024

• Debug test failure for separate I, W execution

  #138863 opened Oct 24, 2024

• Update tensorify pass to specialize symfloats we didn't tensorify away

  #138868 opened Oct 25, 2024

• [Dynamo] destroy old extra state

  #138870 opened Oct 25, 2024

• [AOTInductor] Fix memory access in update_constant_buffer on CPU

  #138871 opened Oct 25, 2024

• [Pipelining] Update schedules to use I, B actions.

  #138886 opened Oct 25, 2024

• [Test][Do not merge] Add BRGEMM API versioning to be compatible with different oneDNN versions

  #138887 opened Oct 25, 2024

• [Test][Do not merge] Upgrade oneDNN to v3.6

  #138889 opened Oct 25, 2024

• [Dynamo] Add explaination for not support usecase to fix todo in dynamo

  #138890 opened Oct 25, 2024

• Add some error messages for flexattention

  #138891 opened Oct 25, 2024

• Add CUDA 12.6 to Binaries Matrix

  #138899 opened Oct 25, 2024

• rocm CI fixes

  #138900 opened Oct 25, 2024

• [Intel GPU][Windows] Remove hardcoded SYCL runtime version

  #138904 opened Oct 25, 2024

• Disable c10::optional macros

  #138912 opened Oct 25, 2024

• Allow PyTorch to be built with USE_DISTRIBUTED=0

  #138913 opened Oct 25, 2024

• only unspecialize float during eager

  #138915 opened Oct 25, 2024

• unspecialize float for all torch.compile callers

  #138918 opened Oct 25, 2024

• [aoti] Create dir when using aoti_compile and package if it isn't already there

  #138919 opened Oct 25, 2024

• always unspecialize float

  #138922 opened Oct 25, 2024

• Move pippy to training IR

  #138923 opened Oct 25, 2024

• [Pipelining] Optimize ready_to_schedule logic

  #138924 opened Oct 25, 2024

• [Pipelining] Remove unused special case from simulator

  #138928 opened Oct 25, 2024

• [export] add mark_unbacked to export dynamic_shapes API

  #138929 opened Oct 25, 2024

• TEST IGNORE: async compile

  #138930 opened Oct 25, 2024

• Gaussian nll loss scalar variance support

  #138931 opened Oct 25, 2024

• Fix weights_only for BUILD instructions for user allowlisted objects with __slots__

  #138936 opened Oct 25, 2024

• [RFC] Add caching heuristic for skipping fast graphs

  #138938 opened Oct 25, 2024

• [Inductor] refactor the require_stride_order logic in GraphLowering

  #138946 opened Oct 25, 2024

• [ROCm] CK Flash Attention Backend

  #138947 opened Oct 25, 2024

• [TF32][Inductor] Account for TF32 in `test_inductor_layout_optimization_input_mutations`

  #138948 opened Oct 25, 2024

• Enable py dispatcher in FakeTensorMode.__enter__

  #138953 opened Oct 25, 2024

• config: Add env_name_default and env_name_force to Config

  #138956 opened Oct 25, 2024

• Turn off PRINT_AUTOTUNE by default

  #138963 opened Oct 26, 2024

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

  #138964 opened Oct 26, 2024

• Use -Weverything

  #138966 opened Oct 26, 2024

• Don't uselessly recompute axiom dict every static eval call

  #138967 opened Oct 26, 2024

• support nesting of suppress_guards, suppress guards when generated compiled autograd graph

  #138968 opened Oct 26, 2024

• [BE] typing for decorators - library

  #138969 opened Oct 26, 2024

• I fixed the conflict between Qt and libtorch

  #138982 opened Oct 26, 2024

• Caomengxuan666 patch 2

  #138983 opened Oct 26, 2024

• Fix `USE_STATIC_MKL` lost functionality

  #138996 opened Oct 26, 2024

• [c10d] Remove ProcessGroupGloo + CUDA tests

  #138998 opened Oct 26, 2024

• Unify shallow_copy_and_detach overloads by passing c10::VariableVersion

  #138999 opened Oct 27, 2024

• automatic_dynamic_local_pgo

  #139001 opened Oct 27, 2024

• [PGNCCL] Make sure we do not use split for P2P comm creation

  #139013 opened Oct 27, 2024

• Revert "[dynamo] Simplify creation of VariableTrackers (#135714)"

  #139014 opened Oct 27, 2024

• removed stale use_lazy_graph_module

  #139015 opened Oct 27, 2024

• Check for valid inputs to legacy _remove_batch_dim()

  #139016 opened Oct 27, 2024

• Add missing description for input parameters

  #139020 opened Oct 27, 2024

• [Intel GPU] Current libtorch_xpu.so double linked libtorch_xpu_ops.a,

  #139024 opened Oct 28, 2024

• [Intel GPU] Support RegisterXPU.cpp codegen and compile for the in-tree XPU structured GEMM OPs.

  #139025 opened Oct 28, 2024

• [WIP][AOTI XPU] Update torch-xpu-ops to extend c_shim_xpu layer with out-of-tree ATen OPs

  #139026 opened Oct 28, 2024

• Fix torch.std_mean return Nan mean value, dismatch with torch.mean

  #139035 opened Oct 28, 2024

• [rfc][dynamo] "skip_guard_eval" stance for power users

  #139038 opened Oct 28, 2024

• Update torch-xpu-ops commit pin

  #139041 opened Oct 28, 2024

• fix collective name of RECORD_PARAM_COMMS_DATA

  #139042 opened Oct 28, 2024

• enable concat linear with mkldnn linear by flag

  #139048 opened Oct 28, 2024

• Update slow tests

  #139051 opened Oct 28, 2024

• [AOTI] Use `len(serialized_weights)` when calculating `consts_size`

  #139054 opened Oct 28, 2024

• [aotd] Fuse tangents subclasses runtime traversals

  #139068 opened Oct 28, 2024

• [Flex Attention] Dynamic shapes fix

  #139069 opened Oct 28, 2024

• Type check delta value in HuberLoss

  #139070 opened Oct 28, 2024

• [aoti] Add masked_select to cshim

  #139071 opened Oct 28, 2024

• [PyTorch] Move bf16_gemv_trans to ReducedPrecisionFloatGemvFastPathKernel

  #139081 opened Oct 28, 2024

• [PyTorch] Add efficient isnan for NEON float

  #139082 opened Oct 28, 2024

• [PyTorch] Add efficient isnan for NEON half

  #139083 opened Oct 28, 2024

• [PyTorch] Extract value_type-generic NEON Vectorized<Half> functions to CRTP base class

  #139084 opened Oct 28, 2024

• [ROCm] Fix largeIndexBlockSize

  #139087 opened Oct 28, 2024

• [PyTorch] Add Vectorized<c10::BFloat16> specialization for ARM

  #139090 opened Oct 28, 2024

• [DO NOT LAND][EXPERIMENT][dynamo] Remove all instances in `dynamo_expected_failures` that are from `test/dynamo`

  #139091 opened Oct 28, 2024

• [AOTI][refactor] Move stack allocation related configs

  #139093 opened Oct 28, 2024

• [invoke_subgraph][aot_autograd_cache] Cache AC HOP

  #139094 opened Oct 28, 2024

• [aotd] coerce_same_metadata_as_tangent with expected_type for e.g.AsyncCollectiveTensor

  #139095 opened Oct 28, 2024

• [WIP] functional autograd + compiled autograd

  #139098 opened Oct 28, 2024

• Fix and update the Quantization docs

  #139100 opened Oct 28, 2024

• [Dynamo changes] Invoke Quant

  #139101 opened Oct 28, 2024

• [Inductor changes] Invoke Quant

  #139102 opened Oct 28, 2024

• TESTING

  #139105 opened Oct 28, 2024

• Add utility to get all unsafe globals in checkpoint

  #139106 opened Oct 28, 2024

• Add a hook for when AC caches a saved tensor

  #139109 opened Oct 28, 2024

• Update linter targets to Python-3.9

  #139119 opened Oct 28, 2024

• [compiled autograd][functional] partially inline into accumulate grad

  #139121 opened Oct 28, 2024

• [draft] add a c10d/PG restart UT

  #139123 opened Oct 28, 2024

• [export] Modify swap to take in lambda

  #139127 opened Oct 28, 2024

• update onnx==1.17.0

  #139128 opened Oct 28, 2024

• [invoke_subgraph] Generate fake_inputs correctly

  #139130 opened Oct 28, 2024

• Enable cppcoreguidelines-special-member-functions

  #139132 opened Oct 29, 2024

• [executorch hash update] update the pinned executorch hash

  #139133 opened Oct 29, 2024

• [profiler] Annotate triton kernels with kernel hash

  #139135 opened Oct 29, 2024

• [inductor] patterns to remove pointless view/permute pairs

  #139136 opened Oct 29, 2024

• TunableOp assume dense inputs for size calculations

  #139137 opened Oct 29, 2024

• Added aten.bernoulli.p and aten.bernoulli.default decompositions

  #139141 opened Oct 29, 2024

• TEST CHANGE TO RVALUE REF

  #139146 opened Oct 29, 2024

• Relax python linter requirements

  #139148 opened Oct 29, 2024

• [AOTI] Ignore .o files in package_aoti

  #139153 opened Oct 29, 2024

• [AOTI] Switch OSS dashboard to use aoti_compile_and_package

  #139154 opened Oct 29, 2024

• Remove const fromDLPack overload

  #139156 opened Oct 29, 2024

• [Autotune Inductor] Some clean up and dataclassing

  #139157 opened Oct 29, 2024

• [PyTorch] Migrate bf16 gemv fast path kernel from intrinsics to vec::Vectorized

  #139159 opened Oct 29, 2024

• [invoke_subgraph] User facing API to support arbitrary args and kwargs

  #139162 opened Oct 29, 2024

• Use std::string_view in get_fully_qualified_type_name

  #139164 opened Oct 29, 2024

• [Test] Refactor using make_tensor replace generate input logic

  #139168 opened Oct 29, 2024

• Use the device interface for detecting Triton availability

  #139171 opened Oct 29, 2024

• [Inductor][CPU] Enable the oneDNN Linear fusion for special case

  #139172 opened Oct 29, 2024

• Use static_assert to detect get_type_index used in device code

  #139173 opened Oct 29, 2024

• Update grad_scaler.py（Just for this bug)

  #139174 opened Oct 29, 2024

• Use cached dnnl::stream in GpuStreamManager

  #139176 opened Oct 29, 2024

• Export XPU oneDNN header to the public

  #139177 opened Oct 29, 2024

• [dim_order] raised runtime error when tensor has ambiguous dim order

  #139180 opened Oct 29, 2024

• Tests Generelization for multiple accelerator devices

  #139184 opened Oct 29, 2024

• [Tests]: replace deprecated `pkg_resources` dependency with `importlib` in stdlib

  #139186 opened Oct 29, 2024

• Bring back `io.BytesIO` type annotation for the output destination of `torch.onnx.export`

  #139187 opened Oct 29, 2024

• Revert D65030974

  #139194 opened Oct 29, 2024

• Support view_as() on NJT; allow nested int swapping

  #139196 opened Oct 29, 2024

• Add check for unsupported sprase layout to resolve false INTERNAL ASSERT FAILED

  #139198 opened Oct 29, 2024

• [Inductor][CI] Add numpy-2.X shard

  #139199 opened Oct 29, 2024

• turn off USE_MIMALLOC_ON_MKL temporary.

  #139204 opened Oct 29, 2024

• Add Gaudi support to benchmarks/dynamo

  #139205 opened Oct 29, 2024

• [triton] Update pin for PyTorch 2.6/Triton 3.2

  #139206 opened Oct 29, 2024

• [PyTorch] Build bf16 gemv fast path & entry points for non-ARM architectures too

  #139208 opened Oct 29, 2024

• Fix custom obj being input

  #139209 opened Oct 29, 2024

• Add Autograd Fallback for MTIA

  #139211 opened Oct 29, 2024

• [real_tensor_prop] Infer Fake kernels during real tensor prop

  #139213 opened Oct 29, 2024

• ModuleTracker: Add explicit garbage collection

  #139214 opened Oct 29, 2024

• [dynamo] ignore False/None callback in fail_on_recompile/force_backend stances

  #139215 opened Oct 29, 2024

• [invoke_subgraph] Re-enable fake tensor model in the fake tensor impl

  #139216 opened Oct 29, 2024

• [mergebot] Add ci-no-td label on revert

  #139218 opened Oct 29, 2024

• [export] deregister hooks and see what breaks

  #139219 opened Oct 29, 2024

• Hook up bf16_gemv_trans to x86 bf16 GEMM

  #139220 opened Oct 29, 2024

• Add utility to get all unsafe globals in checkpoint (no pickletools dependency)

  #139221 opened Oct 29, 2024

• [export] Update min_val and max_val to Optional[int] in serialization.

  #139223 opened Oct 29, 2024

• [AOTI] Update zero size computation in clone_preserve_strides

  #139224 opened Oct 29, 2024

• Update cpuinfo to 8df44962

  #139225 opened Oct 29, 2024

• [experimental] async-tp impl with cutlass-based, progress aware kernel

  #139227 opened Oct 29, 2024

• [inductor] Enable AMD cooperative reduction tests

  #139230 opened Oct 29, 2024

• [PyTorch] Fix lack of alias annotations for dropout variants

  #139231 opened Oct 29, 2024

• Move pippy to training IR

  #139233 opened Oct 29, 2024

• Fx graph always return tuple in fuse_as_graphmodule

  #139236 opened Oct 29, 2024

• Fix existing lint issues in ir.py

  #139237 opened Oct 29, 2024

• typing ir.py - Disallow untyped defs for ir.py

  #139238 opened Oct 29, 2024

• Classify miss-inplaced tensors in logs.

  #139240 opened Oct 30, 2024

• Do not print "we strongly recommend" for CPU builds

  #139243 opened Oct 30, 2024

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

  #139247 opened Oct 30, 2024

• [3/N] Fix clang-tidy warnings in python_variable_methods.cpp

  #139248 opened Oct 30, 2024

• Add conjugate method on SymFloat

  #139249 opened Oct 30, 2024

• Don't set replacement if lhs is in the free symbols of the rhs

  #139250 opened Oct 30, 2024

• [DCP] Unit Test to validate the stateful and non-stateful loads

  #139251 opened Oct 30, 2024

• [4/N] Fix Wextra-semi warning

  #139256 opened Oct 30, 2024

• [WIP] Add SYCL version control in cmake to keep BC

  #139258 opened Oct 30, 2024

• [logging] Add new utilities to record and log compilation metrics

  #139259 opened Oct 30, 2024

• [logging] Superficial use of new metrics_context in CompilationMetrics logging

  #139260 opened Oct 30, 2024

• Add sv starts/ends_with

  #139261 opened Oct 30, 2024

• [logging] Use metrics_context.timed to log CompilationMetrics time fields

  #139263 opened Oct 30, 2024

• Fix index_reduce mean error when dtype is bfloat16 (#139242)

  #139264 opened Oct 30, 2024

• [4/N] Don't skip ASAN on some tests

  #139265 opened Oct 30, 2024

• [Intel GPU] Support RegisterSparseXPU.cpp codegen.

  #139267 opened Oct 30, 2024

• [Dynamo] Fix graph break when `tensor.split()` is called within a device context manager

  #139270 opened Oct 30, 2024

• [Quant][Inductor] modify QConv/QLinear + broadcast add fusion

  #139271 opened Oct 30, 2024

• Lsan2

  #139273 opened Oct 30, 2024

• [fx] split_module subgraph should always have an output node

  #139275 opened Oct 30, 2024

• Add rvalue overload of THPVariable_Wrap

  #139278 opened Oct 30, 2024

• [dyanmo] fix `deque.maxlen` support when extending elements from left

  #139279 opened Oct 30, 2024

• Block more keys from config serialization

  #139285 opened Oct 30, 2024

• [9/N] Fix extra warnings brought by clang-tidy-17

  #139286 opened Oct 30, 2024

• [easy] Add start event metadata to collected metadata for PT2 Compile Events

  #139289 opened Oct 30, 2024

• [Easy] GraphTransformObserver Refactoring

  #139292 opened Oct 30, 2024

• [Easy] Refactor post grad application of passes

  #139293 opened Oct 30, 2024

• [Easy] Add joint graph passes, fallback_random to bisector

  #139295 opened Oct 30, 2024

• [ch] Move close_nonexistent_disable_issues.py queries to CH

  #139296 opened Oct 30, 2024

• [ROCM] Increase hipBLASLt default workspace size

  #139300 opened Oct 30, 2024

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

  #139302 opened Oct 30, 2024

• [BE] Change _marked_safe_globals_list to set

  #139303 opened Oct 30, 2024

• Add bfloat16 support for per tensor/channel cpu/cuda fake quantize ops

  #139306 opened Oct 30, 2024

• export graph for memory simulator

  #139308 opened Oct 30, 2024

• Add option to dynamo_timed and chromium_event_logger for logging pt2 compile events

  #139309 opened Oct 30, 2024

• [Profiler] Create Auto-Trace Frontend for Trace ID

  #139310 opened Oct 30, 2024

• Modified the curl|bash code to make it palatable to security-checking…

  #139311 opened Oct 30, 2024

174 Issues closed by 46 people

• [Upstream Triton] attrs_description `KeyError: 'cls'`

  #139179 closed Oct 30, 2024

• convert torch.jit.script model to ONNX get wrong result

  #88072 closed Oct 30, 2024

• Torch 1.13 - 2.3 Onnx Scope name not correct!

  #90439 closed Oct 30, 2024

• [Enhancement] Allow Dict[str, Any] format for `transforms` argument in `torchvision.transforms.v2.Compose`

  #139291 closed Oct 30, 2024

• [MPS] Typo in error message for supported autocast type

  #139190 closed Oct 30, 2024

• Adjust install_user script for Ubuntu 24.04 support

  #138812 closed Oct 30, 2024

• DISABLED test_device_mode_ops_sparse_mm_reduce_cpu_float32 (__main__.TestDeviceUtilsCPU)

  #132552 closed Oct 30, 2024

• DISABLED test_pre_dispatch_export_auto_functionalize_simple_cuda_float32 (__main__.TestHOPCUDA)

  #123267 closed Oct 30, 2024

• Fix

  #139276 closed Oct 30, 2024

• Cannot build flash-attention with torch==2.5.0

  #139067 closed Oct 30, 2024

• Bug in the description of RMSNorm

  #139274 closed Oct 30, 2024

• [dynamo] Format string with __class__

  #118675 closed Oct 30, 2024

• [DEBUG] Strange behavior observed with PyTorch 2.4.0 + Windows + CPU inference

  #131958 closed Oct 30, 2024

• DISABLED [WORKFLOW_NAME] / [PLATFORM_NAME] / [JOB_NAME]

  #139277 closed Oct 30, 2024

• DISABLED test_kineto_profiler_with_environment_variable (__main__.TestProfiler)

  #107383 closed Oct 30, 2024

• RuntimeError: Storage size calculation overflowed with sizes=[1, 4605674770382112385, 4128, 128]

  #139181 closed Oct 30, 2024

• [Compiled Autograd] Usage of AOTAutograd cause all hook functions to get pushed to the end of the graph

  #138538 closed Oct 30, 2024

• Distributed Tensor raises error with torch 2.5

  #138742 closed Oct 30, 2024

• Program does not exit when using `torch.distributed.tensor.distribute_module`.

  #139060 closed Oct 30, 2024

• [FSDP2] hsdp2 sometimes does not free grad immeidately after opt.step

  #139234 closed Oct 30, 2024

• `max_unpool2d` returns a tensor with negative dimension

  #73154 closed Oct 30, 2024

• [ONNX] Test _building.py

  #138761 closed Oct 29, 2024

• Vision Dataset : FGVC_AirCraft not a callable module in torchvision.datasets

  #139226 closed Oct 29, 2024

• Release 2.5.1 validations checklist and cherry-picks

  #138876 closed Oct 29, 2024

• MacOS runners queue again

  #138724 closed Oct 29, 2024

• Automatic linter fixing for PRs

  #133715 closed Oct 29, 2024

• [aarch64 linux ci] missing sccache binaries for aarch64 linux platform

  #121559 closed Oct 29, 2024

• Release 2.5.0 validations checklist and cherry-picks

  #137492 closed Oct 29, 2024

• NJT Embedding backward

  #138352 closed Oct 29, 2024

• DISABLED test_fsdp_unsupported_module_cls (__main__.TestFSDPMiscMultiThread)

  #137948 closed Oct 29, 2024

• xpu: huggingface levit test_retain_grad_hidden_states_attentions test hangs on exit on PVC

  #136007 closed Oct 29, 2024

• Output tensor of torch.split_with_sizes_copy has version 0 on CUDA while version 1 on CPU/XPU.

  #136303 closed Oct 29, 2024

• DISABLED test_unbacked_bindings_for_divisible_u_symint (__main__.TestExport)

  #138586 closed Oct 29, 2024

• There is no header file "torch/torch.h"file int the include folder.I doubt if that file was removed in the latest version.

  #138822 closed Oct 29, 2024

• DISABLED test_arange2_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)

  #127343 closed Oct 29, 2024

• DISABLED test_manual_with_data_parallel_dp_type_DDP_ScheduleClass0_use_new_runtime_False (__main__.ComposabilityTest)

  #139138 closed Oct 29, 2024

• Multiprocessing in Dataloaders reduces variaty

  #138989 closed Oct 29, 2024

• Bug in Maxpool2d Operator Exhibited During Metamorphic Testing

  #139078 closed Oct 29, 2024

• [ROCm] MI300X Tunable ops causes 100GB of Memory Leak leading to OOM

  #138532 closed Oct 29, 2024

• some function in torch._utils was skipped

  #138897 closed Oct 28, 2024

• Dynamo CI Shard naming proposal

  #118127 closed Oct 28, 2024

• RReLU doc doesn't specify the eval mode behaving just like LeakyReLU

  #82677 closed Oct 28, 2024

• Should `_native_batch_norm_legit_functional` be in native_functions.yaml?

  #113483 closed Oct 28, 2024

• Incorrect Type for `devices` Parameter in `benchmark_utilization` Function

  #136697 closed Oct 28, 2024

• DISABLED test_execution_trace_with_kineto_cpu (__main__.TestExecutionTraceCPU)

  #138071 closed Oct 28, 2024

• DISABLED test_source_multithreaded_close_in_scope_work_in_main_thread_False (__main__.TestProfiler)

  #119364 closed Oct 28, 2024

• DISABLED test_source_multithreaded_multiple_preexisting_work_in_main_thread_False (__main__.TestProfiler)

  #119526 closed Oct 28, 2024

• DISABLED test_source_multithreaded_complex_work_in_main_thread_False (__main__.TestProfiler)

  #119490 closed Oct 28, 2024

• DISABLED test_source_multithreaded_open_in_scope_work_in_main_thread_False (__main__.TestProfiler)

  #119537 closed Oct 28, 2024

• Missing Torch 2.5 wheels for Python 3.8

  #138979 closed Oct 28, 2024

• When I load the model data of torch_model.bin, it will go directly to GPU loading, and cannot load the whole process on CPU

  #138977 closed Oct 28, 2024

• tlparse: aot_compile doesn't setup compile ids

  #123759 closed Oct 28, 2024

• Attempting to use hipBLASLt on a unsupported architecture!

  #138067 closed Oct 28, 2024

• Error in cpuinfo: prctl(PR_SVE_GET_VL) failed

  #139052 closed Oct 28, 2024

• arange.start is annotated with scalar on inputs but got torch.SymInt

  #138827 closed Oct 28, 2024

• AttributeError: Can't pickle local object 'make_opaque_unary_fn.<locals>.OpaqueUnaryFn'

  #138070 closed Oct 28, 2024

• Changes in type annotations break downstream code

  #138478 closed Oct 28, 2024

• DISABLED test_non_blocking_with_eager_init (__main__.ProcessGroupNCCLGroupTest)

  #118827 closed Oct 28, 2024

• [c10d] ProcessGroupNCCL support setting custom stream for communication

  #138074 closed Oct 28, 2024

• RuntimeError: mat1 and mat2 shapes cannot be multiplied

  #139037 closed Oct 28, 2024

• pytorch 2.5.0 can not import torch

  #138991 closed Oct 28, 2024

• Assertion error when running example in Pippy repo

  #139028 closed Oct 28, 2024

• ValueError: Expected input batch_size (60) to match target batch_size (10)

  #139031 closed Oct 28, 2024

• DISABLED test_reorder_compute_for_overlap (__main__.TestComputeCommReorderingMultiProc)

  #113249 closed Oct 28, 2024

• DISABLED test_ddp_apply_optim_in_backward_ignored_params (__main__.TestDistBackendWithSpawn)

  #106361 closed Oct 28, 2024

• DISABLED test_reorder_compute_for_overlap_custom_runtime_estimation (__main__.TestComputeCommReorderingMultiProc)

  #121195 closed Oct 28, 2024

• [RFC] Allow DeviceMesh to use ncclCommSplit

  #137017 closed Oct 27, 2024

• [PGNCCL][BUG] mutex acquired in recursive way may deadlock

  #138995 closed Oct 27, 2024

• Weights only load numpy.float64 fails in torch 2.4.0

  #138985 closed Oct 27, 2024

• Illegal Memory Access w/ efficient Attention + compile

  #138772 closed Oct 27, 2024

• torch.cond should support omitting arguments to pass in when it is empty

  #138150 closed Oct 26, 2024

• DISABLED test_autograd_cpp_node_data_dependent (__main__.TestCompiledAutograd)

  #125579 closed Oct 26, 2024

• torch.cond would fail on eager mode

  #138664 closed Oct 26, 2024

• Dynamo fails to propagate updates to nonlocal variables across different functions

  #138112 closed Oct 26, 2024

• [dynamo] Speculation log diverges when Lazy Module is used in a certain way

  #138489 closed Oct 26, 2024

• Leverage DLPack-based construction of PyTorch tensors to avoid costly element-by-element copy

  #120614 closed Oct 26, 2024

• torch._int_mm accuracy issue on AMD CPU

  #136746 closed Oct 26, 2024

• 2.5.0 PyPI arm64 distribution logs `cpuinfo` error on import

  #138333 closed Oct 25, 2024

• 2.5.0 stable + cuda 12.4 broken; unable to to find cuda libs (works 2.4.1)

  #138324 closed Oct 25, 2024

• [inductor][cpu]drq performance failure in 2024-09-30 nightly release

  #137686 closed Oct 25, 2024

• [FlexAttention] Using FlexAttention with DDP complains about a "higher order optimizer"

  #137481 closed Oct 25, 2024

• Mergebot often fails to fetch DrCI status

  #138057 closed Oct 25, 2024

• [PTD][Doc] Review model parallel tutorial

  #138835 closed Oct 25, 2024

• [PTD][DDP] Review the DDP Video Tutorials and modify the scripts so it can be run

  #138833 closed Oct 25, 2024

• Pipelining zero bubble and activation checkpointing bug

  #136766 closed Oct 25, 2024

• DISABLED test_ddp_apply_optim_in_backward (__main__.TestDistBackendWithSpawn)

  #137761 closed Oct 25, 2024

• DISABLED test_ddp_apply_optim_in_backward_grad_as_bucket_view_false (__main__.TestDistBackendWithSpawn)

  #137766 closed Oct 25, 2024

• DISABLED test_metadata_consistency_check (__main__.DTensorMeshTest)

  #131598 closed Oct 25, 2024

• DISABLED test_scatter_uneven (__main__.DeviceMeshCollectiveTest)

  #96282 closed Oct 25, 2024

• DISABLED test_non_blocking_init (__main__.ProcessGroupNCCLGroupTest)

  #131203 closed Oct 25, 2024

• test/dynamo/test_minifier.py failed when the torch._inductor.config.post_grad_custom_post_pass parameter is configured

  #138892 closed Oct 25, 2024

• Illegal instruction (core dumped) in pytorch 1.10 version

  #136973 closed Oct 25, 2024

• DISABLED test_ddp_control_flow_different_across_ranks (__main__.TestDistBackendWithSpawn)

  #77992 closed Oct 25, 2024

• The dynamic library obtained by torch._export.aot_compile will cause cuda illegal memory access problem when running in parallel with multiple threads.

  #138462 closed Oct 25, 2024

• In a multithreading scenario, the dynamic library obtained from aot_compile encounters a CUDA illegal memory access error during inference.

  #132475 closed Oct 25, 2024

• weights_only=True allow set as well as dict

  #138851 closed Oct 25, 2024

• DISABLED test_all_gather_object_subgroup (__main__.TestDistBackendWithSpawn)

  #98508 closed Oct 25, 2024

• DISABLED test_barrier_group_cuda (__main__.TestDistBackendWithSpawn)

  #137751 closed Oct 25, 2024

• torch.compile + triton.jit: NameError: name 'constexpr' is not defined

  #138509 closed Oct 25, 2024

• Compiling flex attention with batch dependent block mask and dynamic shapes

  #136196 closed Oct 25, 2024

• [RFC] Allow lazy global init + eager subgroup init

  #137018 closed Oct 24, 2024

• torch.special.gammaln (libdevice.lgamma) with torch.compile produces AttributeError(CUDAOptions object has no attribute backend_name), torch.bfloat16

  #138809 closed Oct 24, 2024

• ObserverTest.TestMultipleNetBase intermittently segfaults

  #9137 closed Oct 24, 2024

• Caffe2 has undeclared dependencies

  #54497 closed Oct 24, 2024

• caffe2 load onnx model error:IndexError: Input 475 is undefined!

  #52666 closed Oct 24, 2024

• [ONNX] BUG for Upsample operator export re-used by caffe2

  #22070 closed Oct 24, 2024

• Export torch.cat to ONNX with Dynamic shape does not work on GPU

  #24899 closed Oct 24, 2024

• [Caffe2/ONNX] ONNX LSTM Loading

  #21843 closed Oct 24, 2024

• convert Onnx Slice operator to caffe2 failed

  #20858 closed Oct 24, 2024

• convert the model from pytorch to onnx to caffe2, but get a lower accuracy than before

  #20154 closed Oct 24, 2024

• The latest version of onnx-caffe2 does not support “pow” ？

  #18475 closed Oct 24, 2024

• Error in converting pytorch model to caffe2 using onnx framework

  #17154 closed Oct 24, 2024

• [Caffe2] Dropout modules exported from PyTorch with ONNX: BlobIsTensorType(*blob, CPU). Blob is not a CPU Tensor: [BlobNumber]

  #16991 closed Oct 24, 2024

• Undefined GPU Reference when importing torch with caffe2.onnx.backend for cpu-only pytorch

  #16500 closed Oct 24, 2024

• No Schema registered for ConstantOfShape with domain_version of 9

  #16363 closed Oct 24, 2024

• caffe2::onnx::OnnxExporter::Caffe2OpToOnnxNodes failure

  #16298 closed Oct 24, 2024

• caffe2::onnx::OnnxExporter::CreateGemmNodes fails on bvlc_alexnet

  #16296 closed Oct 24, 2024

• Assert In function importUnsqueeze

  #16248 closed Oct 24, 2024

• [ONNX CI] TestCaffe2End2End.test_squeezenet occasional error

  #14670 closed Oct 24, 2024

• [onnx][caffe2] Is there schedule to support onnxwhile in onnx exporter?

  #13122 closed Oct 24, 2024

• onnx_graph_to_caffe2_net takes a model, not a graph

  #13726 closed Oct 24, 2024

• [Caffe2] Failed to load ONNX model with python 3.6

  #15998 closed Oct 24, 2024

• How to run a pytorch-onnx-caffe2 model on GPU?

  #12702 closed Oct 24, 2024

• Error occuring while converting mnist or cifar model from caffe2 to onnx

  #12704 closed Oct 24, 2024

• [Caffe2] Error importing ConvTranspose2d to Caffe2 with ONNX

  #10667 closed Oct 24, 2024

• pytorch does not compatible with caffe2

  #10249 closed Oct 24, 2024

• Does this mean onnx do not support spatialBN operator?

  #10062 closed Oct 24, 2024

• [onnx-caffe2] Caffe2 only supports padding 2D Tensor

  #9411 closed Oct 24, 2024

• Meet “No module named'tools. setup_helpers‘“ when installing caffe2

  #134640 closed Oct 24, 2024

• LINK : fatal error LNK1102: out of memory ninja: build stopped: subcommand failed. "Caffe2 building failed"

  #131562 closed Oct 24, 2024

• Caffe2 usage of cuDNN RNNv6 API blocks upgrade to cuDNN v9+

  #124790 closed Oct 24, 2024

• can not find the caffe2::Threads

  #124038 closed Oct 24, 2024

• No module named 'caffe2' when using `add_scalar` with string

  #119195 closed Oct 24, 2024

• The derivation of swish activation function is wrong.

  #110815 closed Oct 24, 2024

• Pytorch - cpu only & caffe2 build failing

  #105655 closed Oct 24, 2024

• Issues building with caffe2 enabled

  #100960 closed Oct 24, 2024

• [Bug] Circular Import

  #83710 closed Oct 24, 2024

• [caffee2] Windows build / 'metanet_pb2' (a circular import) Anaconda

  #83379 closed Oct 24, 2024

• Error occurred , when compile source code setting BUILD_CAFFE2=ON

  #78034 closed Oct 24, 2024

• Failed to build `convert_and_benchmark.cc` due to missing `net_observer_reporter_print.h`.

  #75186 closed Oct 24, 2024

• Multiple new caffe2-related build failures.

  #73074 closed Oct 24, 2024

• Caffe2 uses FFMPEG functions that are deprecated in FFMPEG 4.0 and gone in 5.0

  #72254 closed Oct 24, 2024

• Remove Caffe2

  #72536 closed Oct 24, 2024

• gloo_test test_close_connection not working as intended due to unwanted comma

  #70609 closed Oct 24, 2024

• model trace error

  #43196 closed Oct 24, 2024

• [Caffe2] Missing CMAKE_CUDA_COMPILE_WHOLE_COMPILATION

  #18524 closed Oct 24, 2024

• cannot build with tensorrt

  #60228 closed Oct 24, 2024

• Caffe2 building failure when turning off USE_OPENMP

  #17853 closed Oct 24, 2024

• Unicode support for the MS Windows platform

  #13565 closed Oct 24, 2024

• "No module named 'tools.setup_helpers'" when use pip install caffe2

  #10155 closed Oct 24, 2024

• [caffe2] Run resnet50_trainer.py error between 2 machines using GLOO/Redis and ibverbs

  #6422 closed Oct 24, 2024

• [Caffe2] [feature request] How to freeze a layer ? (Selective Backward Propagation)

  #6581 closed Oct 24, 2024

• DISABLED test_sink_waits (__main__.TestComputeCommReorderingMultiProc)

  #121236 closed Oct 24, 2024

• DISABLED test_sink_waits_raise_comms (__main__.TestComputeCommReorderingMultiProc)

  #121316 closed Oct 24, 2024

• DISABLED test_raise_comms (__main__.TestComputeCommReorderingMultiProc)

  #120946 closed Oct 24, 2024

• [discussion] Have PyTorch functions support python scalars (like NumPy) + introduce convenience constants like `torch.pi` and `torch.e` and maybe analogue of `scipy.constants` namespace

  #110636 closed Oct 24, 2024

• sympy.ccode doesn't work with FloorDiv in torch.utils._sympy.functions

  #138523 closed Oct 24, 2024

• [PRE-EMPTIVE] Experimenting with new runners `linux.aws.a100` on `inductor-perf-compare.yml`

  #138708 closed Oct 24, 2024

• not working in site

  #138808 closed Oct 24, 2024

• [Break XPU] C10_UNUSED change and newly add Inductor UTs break XPU CI.

  #138577 closed Oct 24, 2024

• Flexattention: CUDA error: an illegal memory access was encountered

  #134852 closed Oct 24, 2024

• `slow` workflow has been broken for 4+ weeks

  #136694 closed Oct 24, 2024

• flex attention backward pass vmap error only when tensordict imported?

  #134004 closed Oct 24, 2024

• [FlexAttention] Compiled `flex_attention' crashes when training with mixed precision on RTX 3090

  #135723 closed Oct 24, 2024

• Flexattention: compilation fails when using block mask

  #135206 closed Oct 24, 2024

• [Flex attention] RuntimeError with vmap when using torch.compile in create_mask

  #136427 closed Oct 24, 2024

• FlexAttention errors with dynamic shapes when closing over a symbolic shape

  #136914 closed Oct 24, 2024

• [dynamo] Dynamo triggering on list comprehension <listcomp> frames and graph breaking

  #138654 closed Oct 24, 2024

• DISABLED test_save_with_without_initializer_include_initializer_no_fake_mode_no_exported_program (__main__.TestFxToOnnx)

  #131423 closed Oct 24, 2024

• _functional_collectives.all_gather_into_tensor cannot compile in aot_module_simplified

  #112009 closed Oct 24, 2024

• The calculation process of key_padding_mask does not match the document description

  #137887 closed Oct 24, 2024

• [cuSPARSELt] `test_sp24_compile` appears broken on sm80/sm90

  #138769 closed Oct 23, 2024

• DISABLED test_reduce_scatter_v_cuda (__main__.TestDistBackendWithSpawn)

  #137325 closed Oct 23, 2024

• The calculation process of key_padding_mask does not match the document description

  #138566 closed Oct 23, 2024

140 Issues opened by 99 people

• Plan to support “discrete” dynamic dimension on torch.export

  #139307 opened Oct 30, 2024

• torch.export torchaudio kaldi module

  #139305 opened Oct 30, 2024

• [RFC][Pipelining] RNG state communication, avoid seed checkpoint

  #139304 opened Oct 30, 2024

• [ONNX][RFC] Migrate torchlib from onnxscript (torch 2.6/2.7)

  #139301 opened Oct 30, 2024

• CUDNN sdp attention causes loss explosion

  #139298 opened Oct 30, 2024

• `torch.package` warning -- `TypedStorage` is deprecated

  #139297 opened Oct 30, 2024

• Using xpu module in the cuda version Pytorch

  #139290 opened Oct 30, 2024

• No macOS 12 wheels available, cannot use torch binary wheels on GitHub Actions

  #139288 opened Oct 30, 2024

• FSDP1 SHARD_GRAD_OP broken after torch upgrade to 2.4.1 and flash_attn upgrade

  #139287 opened Oct 30, 2024

• test/export/test_retraceability.py fails locally, likely flaky

  #139284 opened Oct 30, 2024

• Draft-mode export: ep.run_decompositions() doesn't run with real tensor prop

  #139283 opened Oct 30, 2024

• Calculate speed

  #139282 opened Oct 30, 2024

• `torch._numpy.ndarray.astype()` does not accept Numpy Dtypes correctly

  #139281 opened Oct 30, 2024

• Possible bug of tools::flight_recorder

  #139280 opened Oct 30, 2024

• Dynamo ignores non-inductor backend graph freezing option

  #139272 opened Oct 30, 2024

• TensorBoard images loading error

  #139269 opened Oct 30, 2024

• Torch Inductor should have a way for new backend to provide build options

  #139268 opened Oct 30, 2024

• Error compiling the torch.library.custom_op with input mutations with set_

  #139257 opened Oct 30, 2024

• DISABLED test_not_implemented_error (__main__.ExcTests)

  #139255 opened Oct 30, 2024

• DISABLED test_dynamo_error (__main__.LoggingTests)

  #139254 opened Oct 30, 2024

• DISABLED test_inductor_error (__main__.LoggingTests)

  #139253 opened Oct 30, 2024

• Ban relative imports in test/

  #139252 opened Oct 30, 2024

• export onnx error with sfft

  #139246 opened Oct 30, 2024

• ncclInternalError: Internal check failed

  #139245 opened Oct 30, 2024

• Tensor.index_reduce produces incorrect result

  #139242 opened Oct 30, 2024

• Please implement batching rule for torch.nn.functional.multi_margin_loss

  #139241 opened Oct 30, 2024

• TORCH_COMPILE_CPROFILE=1 doesn't work on python 3.12

  #139232 opened Oct 29, 2024

• torch.compile'ing individual linears for torchtitan debug model + FSDP2 leads to errors

  #139222 opened Oct 29, 2024

• [Runtime Error] Build PyTorch with cuda12.2 on Jetson AGX Orin with jetpack 5.1.4

  #139210 opened Oct 29, 2024

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

  #139207 opened Oct 29, 2024

• fail_on_recompile stance is unusable because it doesn't handle skipped frames

  #139202 opened Oct 29, 2024

• First run lint on just the changes in the PR before running it over the entire PR

  #139201 opened Oct 29, 2024

• find_spec does something weird to Python Path when loading modules

  #139200 opened Oct 29, 2024

• When I use register_full_backward_hook and register_forward_hook in nn.MultiheadAttention, The MultiheadAttention executed wrong conditional branch and get wrong backward grads.

  #139191 opened Oct 29, 2024

• PCH build fail with sccache-v0.8.2

  #139188 opened Oct 29, 2024

• Operators being traced as method calls in torch.fx

  #139185 opened Oct 29, 2024

• compile fails when split is called within a device cm

  #139183 opened Oct 29, 2024

• [ROCm] [Upstream Triton] num_stages=0 deprecation with stream pipeliner v2

  #139182 opened Oct 29, 2024

• the mpi backend not support to transfer float16/bfloat16 tensor and would post "IndexError: map::at" when transferring float16/bfloat16 tensor

  #139178 opened Oct 29, 2024

• [v2.6] Inductor issue tracker for Triton release/3.2

  #139175 opened Oct 29, 2024

• `pkg_resources` module is deprecated by setuptools

  #139170 opened Oct 29, 2024

• Deepwise convolution with `torch.compile` has inconsistent numerical precision with eager mode

  #139169 opened Oct 29, 2024

• [Dynamo] TypeError: `list` object is not callable

  #139167 opened Oct 29, 2024

• SubgraphMatcher may fail to match when the matching pattern having call_module IR

  #139163 opened Oct 29, 2024

• [ERROR]: fullgraph compilation failed: 'FakeRootModule' object has no attribute 'self___attr_list_0'

  #139160 opened Oct 29, 2024

• Dynamo capture of tensor.data assignment doesn't identical to eager call of tensor.data assignment

  #139152 opened Oct 29, 2024

• DISABLED TCPStoreTest.testMultiTenantStoresUV (__main__.TCPStoreTest)

  #139150 opened Oct 29, 2024

• DISABLED test_like_channels_last_cpu (__main__.CpuTritonTests)

  #139149 opened Oct 29, 2024

• Some files in sccache are owned by `hostmaster+pytorch`

  #139143 opened Oct 29, 2024

• torch.nn.InstanceNorm2d throws "mixed dtype" error with track_running_stats set to True

  #139140 opened Oct 29, 2024

• DISABLED test_manual_with_data_parallel_dp_type_DDP_ScheduleClass0_use_new_runtime_False (__main__.ComposabilityTest)

  #139139 opened Oct 29, 2024

• [FlexAttention] Update the way we generate kernel options

  #139131 opened Oct 28, 2024

• [pgnccl] unstable restarting of PGs

  #139129 opened Oct 28, 2024

• shift operators not supporting integral types

  #139124 opened Oct 28, 2024

• [Profiler/Kineto] Add group_trace_id and trace_id support to auto-trace python frontend

  #139118 opened Oct 28, 2024

• `_adjust_num_blocks_and_indices` gives wrong adjusted block mask

  #139117 opened Oct 28, 2024

• [ROCm] PyTorch TunableOps results in Memory Access Fault

  #139116 opened Oct 28, 2024

• torch.nn.AvgPool2d works with long on cpu but not gpu

  #139115 opened Oct 28, 2024

• `torch.compile` errors when inputs memory overlap.

  #139111 opened Oct 28, 2024

• torch.compile + FSDP1 CPU offloading + PT lightning validation loop throws an error

  #139110 opened Oct 28, 2024

• LibTorch build error on Windows for CUDA version (debug/release)

  #139108 opened Oct 28, 2024

• test/functorch/test_ops.py:test_extremal_numerics_l1_loss_cpu segfaults when wrapped with Dynamo

  #139103 opened Oct 28, 2024

• [inductor][rocm] Cooperative reductions on AMD GPUs

  #139099 opened Oct 28, 2024

• `tensor` not a `FakeTensor` under `FakeTensorMode` and `device('meta')`

  #139092 opened Oct 28, 2024

• [dynamo] `dynamo_expected_failures` is silent on certain tests that ended up passing

  #139080 opened Oct 28, 2024

• Add support for optional tensor kwargs in inductor

  #139077 opened Oct 28, 2024

• DISABLED test_aot_export_scan_simple_cuda_float32 (__main__.TestHOPCUDA)

  #139075 opened Oct 28, 2024

• DISABLED test_retrace_export_scan_simple_cuda_float32 (__main__.TestHOPCUDA)

  #139074 opened Oct 28, 2024

• DISABLED test_pre_dispatch_export_scan_simple_cuda_float32 (__main__.TestHOPCUDA)

  #139072 opened Oct 28, 2024

• DISABLED test_serialize_export_scan_simple_cuda_float32 (__main__.TestHOPCUDA)

  #139073 opened Oct 28, 2024

• Covert Conv3d to Conv2d, But output different

  #139066 opened Oct 28, 2024

• Cannot compile with latest LLVM-19

  #139065 opened Oct 28, 2024

• [Flex Attention] Cannot determine truth value of Relational

  #139064 opened Oct 28, 2024

• Quantization API examples in the docs are outdated

  #139063 opened Oct 28, 2024

• Triton pre_hooks are ignored by torch.compile

  #139059 opened Oct 28, 2024

• DISABLED test_where_broadcast_cpu (__main__.CpuTests)

  #139057 opened Oct 28, 2024

• DISABLED test_wrapper_subclass_aliasing_conv2d_cpu (__main__.TestWrapperSubclassAliasingCPU)

  #139056 opened Oct 28, 2024

• CI Failure for Instruction Count doesn't say what file to update

  #139034 opened Oct 28, 2024

• Stabilizing the Kumaraswamy Distribution

  #139019 opened Oct 27, 2024

• DISABLED test_extra_cuda_context (__main__.ProcessGroupNCCLGroupTest)

  #139011 opened Oct 27, 2024

• Dependency management corrupted in pytorch 2.5.0

  #139005 opened Oct 27, 2024

• fatal error C1083: 无法打开包括文件: “cstddef”: No such file or directory error on windows

  #139002 opened Oct 27, 2024

• Build failed when use shared MKL library

  #138994 opened Oct 26, 2024

• Support `attn_mask` in `jagged_scaled_dot_product_attention`

  #138993 opened Oct 26, 2024

• Error loading ".venv\Lib\site-packages\torch\lib\c10_xpu.dll" or one of its dependencies

  #138986 opened Oct 26, 2024

• The error compiling problem of Qt and libtorch(my solution to eliminate the bug)

  #138981 opened Oct 26, 2024

• compile_fx inplace modifly the input graph

  #138980 opened Oct 26, 2024

• When using torch.jit.trace for inference, the element of the shape attribute of tensor and the return value of numel() are tensor

  #138973 opened Oct 26, 2024

• What version or package name will be used in aarch64 release for 2.6 on pypi?

  #138971 opened Oct 26, 2024

• CompiledFxGraph.current_callable is not thread-safe

  #138961 opened Oct 26, 2024

• Write a script to convert PT2 compiler benchmark results to the standardize format for OSS benchmark

  #138960 opened Oct 26, 2024

• [Inductor][AOTInductor] `test_constant_folding_abi_compatible_cpu` surfaces `CUDA error: invalid argument` on H100

  #138958 opened Oct 25, 2024

• `DataLoader` probably shouldn't use `fork` by default

  #138957 opened Oct 25, 2024

• DISABLED test_custom_op_all_inputs_cpu_with_stack_allocation (__main__.AOTInductorTestABICompatibleCpuWithStackAllocation)

  #138955 opened Oct 25, 2024

• [Feature request] Flag to prevent certain parameters from being typecast by nn.module.to()

  #138952 opened Oct 25, 2024

• cudagraphed compiled module slows down when called recursively

  #138949 opened Oct 25, 2024

• DISABLED test_dtype_aware_codegen_load_upcast_to_fp32_False_bfloat16 (__main__.TritonCodeGenTests)

  #138944 opened Oct 25, 2024

• DISABLED test_dtype_aware_codegen_load_upcast_to_fp32_False_float16 (__main__.TritonCodeGenTests)

  #138943 opened Oct 25, 2024

• Libtorch torch::Tensor default constructor does not initialize on CPU like in python

  #138940 opened Oct 25, 2024

• torch.cond complains two tensors have different metadata

  #138926 opened Oct 25, 2024

• compiled autograd can specialize symints from backward graph without guarding

  #138920 opened Oct 25, 2024

• The "register_module" function got a "key_description=<error>" and reported that "Exception 0xc0000005 encountered at address 0x7ff7fb3dfbf3: Access violation reading location 0xffffffffffffffff"

  #138916 opened Oct 25, 2024

• DISABLED test_byte_mask (__main__.TestAdvancedIndexing)

  #138911 opened Oct 25, 2024

• Replace magma-cuda dependency on `pytorch/conda-builder` Docker image

  #138909 opened Oct 25, 2024

• Output mismatch between torch.compile and eager mode

  #138908 opened Oct 25, 2024

• Select SDPA backend smartly by `sdp_params` and benchmark results

  #138907 opened Oct 25, 2024

• RuntimeError: "addmm_cuda" not implemented for 'ComplexHalf'

  #138906 opened Oct 25, 2024

• DISABLED test_sdpa_mask_fp16_L6_S17_NH23_HS121 (__main__.TestSDPA)

  #138905 opened Oct 25, 2024

• Runtime error: retains_grad_hooks not implemented for compiled autograd

  #138901 opened Oct 25, 2024

• cpu faster than mps for both GRUCell and LSTMCell on apple silicon M3 max: MPS Metal

  #138898 opened Oct 25, 2024

• Inductor needs to be updated to work with upstream triton's AttrsDescriptor

  #138895 opened Oct 25, 2024

• How to Implement multi-card parallel Inference by torchrun?

  #138888 opened Oct 25, 2024

• DISABLED test_file_reader_no_memory_leak (__main__.TestScript)

  #138885 opened Oct 25, 2024

• DISABLED test_slice_with_floordiv_serdes_non_strict (__main__.SerDesExportNonStrictTestExport)

  #138884 opened Oct 25, 2024

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

  #138864 opened Oct 24, 2024

• Using int(shape) in export would result in silent specialization

  #138853 opened Oct 24, 2024

• [torch.export] round is not an allowed operator type

  #138852 opened Oct 24, 2024

• Using shape dependent conditionals with export resulted in weird KeyError

  #138847 opened Oct 24, 2024

• `torch.distributed._state_dict_utils._broadcast_tensors` does not properly support CPU tensors.

  #138842 opened Oct 24, 2024

• [export] `run_decompositions` fails with pytree error on `Llama-3.2-vision`

  #138839 opened Oct 24, 2024

• [PTD][RPC] Verify RPC Tutorials contents and scripts

  #138832 opened Oct 24, 2024

• Add mean and var operation for Nested Tensors

  #138831 opened Oct 24, 2024

• [ONNX] Support for `aten.istft`

  #138830 opened Oct 24, 2024

• torch.dot has inconsistent support for int64 (long)

  #138829 opened Oct 24, 2024

• [inductor][cpu]llama amp single thread inductor_max_autotune performance regression in 2024-10-16 nightly release

  #138823 opened Oct 24, 2024

• Failed to load image Python extension: libc10_cuda.so: cannot open shared object file: No such file or directory warn(f"Failed to load image Python extension: {e}")

  #138821 opened Oct 24, 2024

• RuntimeError: HalfTensor is not supported

  #138820 opened Oct 24, 2024

• autograd.Function HOP (maybe others) have name conflict when lifting freevars

  #138817 opened Oct 24, 2024

• Strange recompilations on torch 2.5 + FSDP + UNet

  #138813 opened Oct 24, 2024

• same data all reduce on H20, but results are different

  #138811 opened Oct 24, 2024

• Batching rule not defined for `aten::_make_dual`.

  #138800 opened Oct 24, 2024

• compile + allgather with group will fail for stack-style allgather

  #138795 opened Oct 24, 2024

• Support nn.Module arguments for the function to torch.compiler.allow_in_graph

  #138786 opened Oct 24, 2024

• torch.cond works, but would still print a misleading warning

  #138782 opened Oct 24, 2024

• [Inductor] Regression in test_comprehensive_nn_functional_max_pool2d_cuda from triton

  #138775 opened Oct 24, 2024

• [funcol] functional collectives are 67% slower than torch.distributed collectives

  #138773 opened Oct 24, 2024

• Accuracy issues (NANs) in torch.sdpa backward on ROCm

  #138764 opened Oct 23, 2024

• CI/CD: Figure out what to do with split build

  #138750 opened Oct 23, 2024

• Speed up gaussian_nll_loss when variance is always the same scalar value

  #138747 opened Oct 23, 2024

• Cleanup the scaling logic in runtime.triton_heuristics.triton_config

  #138743 opened Oct 23, 2024

446 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.

• Consolidate Triton cache into Inductor cache

  #138239 commented on Oct 30, 2024 • 52 new comments

• [pt2, docs] Add new PT2 troubleshooting doc

  #138620 commented on Oct 30, 2024 • 43 new comments

• Generalization of distributed test cases for non-CUDA devices

  #138216 commented on Oct 28, 2024 • 29 new comments

• Relax mutation to aten.to in export

  #138606 commented on Oct 30, 2024 • 28 new comments

• [hop free symbols] lift free symbols in example_value when create_graph_input

  #138363 commented on Oct 30, 2024 • 20 new comments

• [Inductor][CPP] Cache weight tiles in L1D for AMX int8 WoQ GEMM

  #136688 commented on Oct 30, 2024 • 20 new comments

• [Intel GPU] qconv at XPU backend

  #133080 commented on Oct 30, 2024 • 16 new comments

• [hierarchical-compilation][invoke-subgraph] Add fake tensor caching

  #137808 commented on Oct 29, 2024 • 11 new comments

• make it clearer (in docs) one can double decorate with torch.library.impl_* APIs

  #137608 commented on Oct 30, 2024 • 11 new comments

• Extending SVE VEC Backend Support in PyTorch to SVE128 and SVE512.

  #138388 commented on Oct 30, 2024 • 10 new comments

• [aotd] Unwrap unseen AsyncCollectiveTensor tangents

  #138731 commented on Oct 28, 2024 • 8 new comments

• Add TORCHDYNAMO_EXTENDED_ADVICE (#137159)

  #137196 commented on Oct 29, 2024 • 7 new comments

• Make c10::string_view an alias of std::string_view

  #130417 commented on Oct 30, 2024 • 6 new comments

• Support tensor betas in Adam and AdamW

  #134171 commented on Oct 29, 2024 • 6 new comments

• Allow inplacing buffer when other users are inconsequential

  #138383 commented on Oct 30, 2024 • 6 new comments

• [Dynamo] Allow `filter()` to handle infinite iterator

  #138305 commented on Oct 29, 2024 • 5 new comments

• [aotd] capture rrelu_with_noise noise mutation

  #138503 commented on Oct 30, 2024 • 5 new comments

• Add deterministic path for CUDA `cumsum`

  #136224 commented on Oct 30, 2024 • 5 new comments

• [AOTI] Introduce an extensibility mechanism for the c shim codegen to make it easy to produce c shims for out-of-tree OP kernels as well. Add c_shim for XPU.

  #136742 commented on Oct 30, 2024 • 5 new comments

• Add UTs for accelerator device-agnostic runtime APIs

  #133572 commented on Oct 30, 2024 • 4 new comments

• Initial work on "cuda graphs with arguments".

  #137318 commented on Oct 27, 2024 • 4 new comments

• [inductor] fix the unligned variable ranges issue in fuse node

  #138568 commented on Oct 29, 2024 • 4 new comments

• [PyTorch] Hook up fp16_gemv_trans to x86 fp16 GEMM

  #137918 commented on Oct 30, 2024 • 4 new comments

• init kineto after torch module initialized

  #131448 commented on Oct 30, 2024 • 4 new comments

• [hop free symbols][refactor] make create_graph_input always take example_value

  #138428 commented on Oct 29, 2024 • 4 new comments

• [fx graph cache] Support freezing with FX graph caching

  #136505 commented on Oct 29, 2024 • 4 new comments

• [cpu] add int8 sdpa api

  #138688 commented on Oct 25, 2024 • 4 new comments

• make equation behind torch.isclose element-wise

  #138459 commented on Oct 29, 2024 • 3 new comments

• Add BRGEMM API versioning to be compatible with different oneDNN versions

  #138184 commented on Oct 30, 2024 • 3 new comments

• FlopCounterMode: Decompose ops for inference mode

  #138508 commented on Oct 28, 2024 • 3 new comments

• typing ir.py - part 2

  #131846 commented on Oct 30, 2024 • 3 new comments

• [Environment Variable][5/N] Use thread-safe getenv functions

  #138438 commented on Oct 24, 2024 • 3 new comments

• [AOTI] add C shim for QConvPointWise

  #138540 commented on Oct 29, 2024 • 3 new comments

• [Inductor][CPP] Add oneDNN BRGEMM config for Half cpp gemm template

  #136255 commented on Oct 30, 2024 • 3 new comments

• [Pipelining] Runtime support and optimization for separate dI / dW

  #131762 commented on Oct 25, 2024 • 3 new comments

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

  #134817 commented on Oct 24, 2024 • 3 new comments

• Update pin memory related APIs to not pass 'device' argument

  #131858 commented on Oct 30, 2024 • 2 new comments

• Add clamp function for bool tensor

  #138618 commented on Oct 29, 2024 • 2 new comments

• [MPS] Compile kernels into Metallib

  #138636 commented on Oct 27, 2024 • 2 new comments

• [hop free symbols][refactor] make bound_symbols a dictionary

  #138345 commented on Oct 29, 2024 • 2 new comments

• [Quant][Inductor] support QConv/QLinear + broadcast add fusion

  #138185 commented on Oct 30, 2024 • 2 new comments

• softshrink nan fixes

  #138421 commented on Oct 28, 2024 • 2 new comments

• [Partitioner] Enumerate partitions by iterating partition ids

  #136598 commented on Oct 26, 2024 • 2 new comments

• Add float-to-int conversion errror checks

  #138592 commented on Oct 25, 2024 • 2 new comments

• [inductor] make requires_stride_order more unbacked-symint-aware

  #137063 commented on Oct 30, 2024 • 2 new comments

• unravel_index check if all elements in `indices` are within the acceptable range

  #138223 commented on Oct 29, 2024 • 2 new comments

• [FSDP] Add support for `FSDPCommContext` on PrivateUse1 backend

  #138687 commented on Oct 29, 2024 • 2 new comments

• fix dynamo tracking numpy 2 ops

  #138686 commented on Oct 30, 2024 • 2 new comments

• [Partitioner] Speed up the update of partition map

  #136616 commented on Oct 25, 2024 • 2 new comments

• Build transformed distributions within compiled code

  #135001 commented on Oct 25, 2024 • 2 new comments

• Make IPC features extendable on third-party devices

  #133222 commented on Oct 30, 2024 • 2 new comments

• [hop free symbols][refactor] lift freevar to parent graph before lifting to subgraph

  #138559 commented on Oct 30, 2024 • 2 new comments

• `NO_DEPRECATED` cmake switch & C++ macro prevents compilation of deprecated code

  #138405 commented on Oct 24, 2024 • 2 new comments

• Avoid args being parsed when common_utils imported

  #134592 commented on Oct 28, 2024 • 2 new comments

• Link with MKL::MKL instead of MKL_LIBRARIES

  #128195 commented on Oct 24, 2024 • 1 new comment

• fix sampler - force cpu device for .tolist tensors

  #135990 commented on Oct 30, 2024 • 1 new comment

• Fix lazy_property compatibility with dynamo

  #138080 commented on Oct 26, 2024 • 1 new comment

• [PT2D][DDP][WIP] Automatically detect whether to turn on CompiledDDP

  #138560 commented on Oct 26, 2024 • 1 new comment

• check fake/real mismatches during real tensor prop

  #137747 commented on Oct 28, 2024 • 1 new comment

• Use Manylinux2_28 for wheel builds

  #138732 commented on Oct 24, 2024 • 0 new comments

• [WIP][Inductor] auto-chunker

  #136702 commented on Oct 30, 2024 • 0 new comments

• [Partitioner] Reduce time consuming of partitions merger

  #136614 commented on Oct 25, 2024 • 0 new comments

• [Partitioner] Remove unnecessary upstream nodes in dependency viewer

  #136608 commented on Oct 23, 2024 • 0 new comments

• Make Context to be Device-agnostic Step by Step (2/N)

  #136526 commented on Oct 26, 2024 • 0 new comments

• [Reland] Remove outdated Caffe2 code from build scripts

  #135104 commented on Oct 28, 2024 • 0 new comments

• [experiment] xdist

  #137533 commented on Oct 29, 2024 • 0 new comments

• Clean up accelerator APIs

  #137493 commented on Oct 30, 2024 • 0 new comments

• [FlexAttention] add support for learnable biases in Inductor

  #137452 commented on Oct 30, 2024 • 0 new comments

• [sparse] add search for optimal alg_id to torch.compile

  #137427 commented on Oct 29, 2024 • 0 new comments

• [WIP] Added swizzle searching, disabled fp16 accum, and enabled ping-ong for cutlass

  #137410 commented on Oct 25, 2024 • 0 new comments

• Add XuehaiPan to CODEOWNERS for C++ PyTree utilities

  #137408 commented on Oct 30, 2024 • 0 new comments

• [pytree][4/N] make `torch.utils.pytree` as public API

  #137400 commented on Oct 30, 2024 • 0 new comments

• [dynamo][pytree][3/N] make CXX pytree traceable: `tree_map` / `tree_map_`

  #137399 commented on Oct 30, 2024 • 0 new comments

• [dynamo][pytree][2/N] make CXX pytree traceable: `tree_flatten` / `tree_unflatten` / `tree_structure`

  #137398 commented on Oct 30, 2024 • 0 new comments

• [dynamo][pytree][1/N] make CXX pytree traceable: `tree_iter` / `tree_leaves`

  #137397 commented on Oct 30, 2024 • 0 new comments

• [wip][compiled autograd] always lift python autograd function bwd

  #137359 commented on Oct 29, 2024 • 0 new comments

• Implement HPUHooksInterface

  #137338 commented on Oct 30, 2024 • 0 new comments

• [wip][compiled autograd] automatic compiled autograd

  #137326 commented on Oct 29, 2024 • 0 new comments

• [ROCm][Draft] Enable finding HIP and ROCm libraries on Windows

  #137279 commented on Oct 29, 2024 • 0 new comments

• [Inductor] Support tiling reduction dimensions

  #137243 commented on Oct 30, 2024 • 0 new comments

• [WIP] Fix test_out when run on CPU with CUDA available

  #137140 commented on Oct 29, 2024 • 0 new comments

• [reland] Flip triton kernel default layout constraint to "needs_fixed_stride_order"

  #137064 commented on Oct 30, 2024 • 0 new comments

• Update FlexAttention benchmarks

  #137036 commented on Oct 30, 2024 • 0 new comments

• Improvements for associative_scan - Autograd

  #136966 commented on Oct 23, 2024 • 0 new comments

• Avoid reorder in mkldnn_to_dense when output is already in a public format

  #136859 commented on Oct 24, 2024 • 0 new comments

• Enable XNNPACK for quantized add

  #136850 commented on Oct 25, 2024 • 0 new comments

• Add generator parameter to rand*_like functions

  #136780 commented on Oct 28, 2024 • 0 new comments

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

  #136753 commented on Oct 30, 2024 • 0 new comments

• Wrap torch_python with torch_compile_options

  #136743 commented on Oct 28, 2024 • 0 new comments

• [dont-review][dynamo][dicts] Prototype - Guard on dict keys lazily

  #134887 commented on Oct 30, 2024 • 0 new comments

• Adds Error Handling Bindings to Cudart

  #134869 commented on Oct 29, 2024 • 0 new comments

• dumy test : add dumy ci test

  #134814 commented on Oct 29, 2024 • 0 new comments

• Fix the bug to pass correct sequence id into PyTorch exection trace

  #134753 commented on Oct 28, 2024 • 0 new comments

• Bisect PR-134552 changes breaking pytorch github unit test [Not for land] 1

  #134751 commented on Oct 28, 2024 • 0 new comments

• Bisect PR-134552 changes breaking pytorch github unit test [Not for land] 2

  #134750 commented on Oct 28, 2024 • 0 new comments

• [Inductor][Optimus] Fix group fusion stride layout

  #134696 commented on Oct 28, 2024 • 0 new comments

• Make dot and vdot structured ops (#64)

  #134671 commented on Oct 29, 2024 • 0 new comments

• Add TORCHINDUCTOR_VEC_ISA_OK env var for vec_isa_ok

  #134667 commented on Oct 29, 2024 • 0 new comments

• Fix unused Python variables in test/[a-d]*

  #134665 commented on Oct 29, 2024 • 0 new comments

• [wip] TORCH_FAKE_PROCESS_GROUP to run distributed on single gpu

  #134634 commented on Oct 27, 2024 • 0 new comments

• Switch XNNPack tests to use _export_for_training (#4867)

  #134615 commented on Oct 28, 2024 • 0 new comments

• Add functional checkpoint API to torch.func.checkpoint

  #134584 commented on Oct 26, 2024 • 0 new comments

• Disable AMP when propagating fake tensors

  #134583 commented on Oct 29, 2024 • 0 new comments

• [Release only] turn off failing CI tests

  #134575 commented on Oct 26, 2024 • 0 new comments

• implement `torch._foreach_rsqrt`

  #134574 commented on Oct 30, 2024 • 0 new comments

• [ONNX] Allow opset 21 for ONNXRuntime==1.18 compatibility (#127167)

  #134571 commented on Oct 28, 2024 • 0 new comments

• [Inductor][FlexAttention] Temporarily disable non-divisible support until fixing all issues.

  #134505 commented on Oct 26, 2024 • 0 new comments

• [CMake] Turn libshm into an object lib

  #134437 commented on Oct 30, 2024 • 0 new comments

• [CI] Enable CI UT on windows xpu

  #134382 commented on Oct 27, 2024 • 0 new comments

• [aot_inductor] add commit info for triton cache file

  #134340 commented on Oct 30, 2024 • 0 new comments

• Support narrow() on batch dim for NJT

  #136444 commented on Oct 29, 2024 • 0 new comments

• [inductor] modify the heuristic for disabling vectorization

  #136422 commented on Oct 30, 2024 • 0 new comments

• Add `truediv` support in export serializer

  #136364 commented on Oct 24, 2024 • 0 new comments

• Fix unused Python variables outside torch/ and test/

  #136359 commented on Oct 30, 2024 • 0 new comments

• Add a new distributed backend (XCCL) for Intel GPUs

  #136343 commented on Oct 30, 2024 • 0 new comments

• Pass rounding_mode for div reference inputs through kwargs

  #136308 commented on Oct 23, 2024 • 0 new comments

• Set output num_float_feature to have dynamic dimension

  #136268 commented on Oct 29, 2024 • 0 new comments

• Add determinmistic kernel for reflection2d

  #136241 commented on Oct 29, 2024 • 0 new comments

• Add support for `@contextmanager` in Dynamo

  #136033 commented on Oct 30, 2024 • 0 new comments

• Bump `nn.functional.conv3d` tolerances for `test_comprehensive`

  #135719 commented on Oct 24, 2024 • 0 new comments

• [scan] support jit inductor

  #135603 commented on Oct 24, 2024 • 0 new comments

• vec: support RVV

  #135570 commented on Oct 29, 2024 • 0 new comments

• [Reland] Fix tensor.data_ptr() representation overflow

  #135567 commented on Oct 30, 2024 • 0 new comments

• Fix the use of fsspec transactions

  #135541 commented on Oct 26, 2024 • 0 new comments

• Add BFloat16 support for BRGEMM flash attention forward kernel

  #135473 commented on Oct 24, 2024 • 0 new comments

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

  #135465 commented on Oct 30, 2024 • 0 new comments

• [ONNX] New registration API

  #135403 commented on Oct 25, 2024 • 0 new comments

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

  #135338 commented on Oct 30, 2024 • 0 new comments

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

  #135337 commented on Oct 30, 2024 • 0 new comments

• [AOTI XPU] Rename test_cuda_cpp_wrapper.py to test_gpu_cpp_wrapper.py,

  #135320 commented on Oct 30, 2024 • 0 new comments

• [WIP][AOTI XPU] Enable Cpp wraper for Intel GPU.

  #135318 commented on Oct 30, 2024 • 0 new comments

• [Intel GPU] qconv_pointwise.binary XPU support

  #135189 commented on Oct 30, 2024 • 0 new comments

• [hop free symbols][refactor] make map's save_for_backward to handle int

  #138558 commented on Oct 29, 2024 • 0 new comments

• [Draft][CUDA][CI][cusparselt] Only CUDA 11.8 ships the libcusparseLt.so.0, CUDA 12 would use PYPI libcusparselt

  #138547 commented on Oct 30, 2024 • 0 new comments

• [Prototype] Adding lowering to persistent-tma device kernel for _scaled_mm

  #138536 commented on Oct 30, 2024 • 0 new comments

• `addmm`: error on output dtype mismatch.

  #138520 commented on Oct 30, 2024 • 0 new comments

• [Inductor] introduce comm buffer planning

  #138519 commented on Oct 26, 2024 • 0 new comments

• [c10d] user to explicitly specify whether split semantics should be used in new_group

  #138518 commented on Oct 26, 2024 • 0 new comments

• Add test for consistency between meta and CPU devices.

  #138515 commented on Oct 30, 2024 • 0 new comments

• Always produce XML

  #138513 commented on Oct 30, 2024 • 0 new comments

• torch/nn/modules/linear.py: docs: improvements

  #138484 commented on Oct 26, 2024 • 0 new comments

• [inductor] Remove an unused local function in select_algorithm.py

  #138475 commented on Oct 26, 2024 • 0 new comments

• [inductor] Remove an unused variable in ir.py

  #138474 commented on Oct 25, 2024 • 0 new comments

• [inductor] Replace set by OrderedSet

  #138466 commented on Oct 29, 2024 • 0 new comments

• Update test_function_base.py for Numpy 2.0 +

  #138463 commented on Oct 28, 2024 • 0 new comments

• [draft] set_linter finds and replaces builtin set in Python code

  #138454 commented on Oct 29, 2024 • 0 new comments

• Openreg: Add RNG Generator

  #138449 commented on Oct 29, 2024 • 0 new comments

• Add docs page for `torch.inf` and `torch.nan`

  #138430 commented on Oct 30, 2024 • 0 new comments

• [eazy][inductor] report peak GPU memory usage in the wrapper

  #138429 commented on Oct 30, 2024 • 0 new comments

• NJT OpInfo tests v2

  #138370 commented on Oct 29, 2024 • 0 new comments

• [cuDNN][SDPA] Match `query`'s memory layout ordering for `output` in cuDNN SDPA

  #138354 commented on Oct 25, 2024 • 0 new comments

• [logging] Use aot_graph_name for the metrics key for backward compilation.

  #138339 commented on Oct 23, 2024 • 0 new comments

• No-gil cherry picks for 2.5.X

  #138335 commented on Oct 25, 2024 • 0 new comments

• Add missing operator and corresponding unittest

  #138309 commented on Oct 29, 2024 • 0 new comments

• [Distributed] Fix todo relative with docs optimization

  #138302 commented on Oct 29, 2024 • 0 new comments

• Add documentation for torch.default_generator

  #138295 commented on Oct 29, 2024 • 0 new comments

• [Windows XPU] SYCL version compatibility

  #138728 commented on Oct 28, 2024 • 0 new comments

• [Windows XPU] Fix MSVC ambiguous symbol error

  #138727 commented on Oct 30, 2024 • 0 new comments

• cudagraph explicit sync on ROCm only after capture_begin()

  #138722 commented on Oct 23, 2024 • 0 new comments

• [wip] fix meta_registrations + alpha

  #138703 commented on Oct 25, 2024 • 0 new comments

• [ROCm] Enable e5m2 x e5m2 unit test in test_float8_basics

  #138699 commented on Oct 23, 2024 • 0 new comments

• Add overflow check for integer division

  #138684 commented on Oct 29, 2024 • 0 new comments

• [fx graph cache] Refactor FxGraphCachePickler, step 2

  #138683 commented on Oct 29, 2024 • 0 new comments

• [fx graph cache] Refactor FxGraphCachePickler

  #138682 commented on Oct 29, 2024 • 0 new comments

• Add device-agnostic runtime Device/Stream C++ API

  #138677 commented on Oct 29, 2024 • 0 new comments

• Move reduce to template parameter in vectorized_reduction

  #138672 commented on Oct 30, 2024 • 0 new comments

• only unspecialize float during aot_eager

  #138666 commented on Oct 26, 2024 • 0 new comments

• Make test_torchbind.py training IR compatible

  #138658 commented on Oct 29, 2024 • 0 new comments

• Solve issue 138464

  #138653 commented on Oct 25, 2024 • 0 new comments

• [Inductor][ROCm][CK] Enable lowering conv2d instances in CK Inductor backend

  #138643 commented on Oct 29, 2024 • 0 new comments

• Add COW support for MPS backend.

  #138640 commented on Oct 23, 2024 • 0 new comments

• ILP for auto FSDP wrapping

  #138635 commented on Oct 25, 2024 • 0 new comments

• Upgrade to fbscribelogger 0.1.7

  #138634 commented on Oct 24, 2024 • 0 new comments

• torch/nn/utils/rnn.py: docs: improvements

  #138628 commented on Oct 26, 2024 • 0 new comments

• Fix global namespace pollution in ATen/Dispatch.h (#138622)

  #138626 commented on Oct 30, 2024 • 0 new comments

• (wip)[CUTLASS] Support re-running with different shapes

  #138611 commented on Oct 24, 2024 • 0 new comments

• [ONNX] Bypass mark_static_address when converting a model to onnx

  #138580 commented on Oct 23, 2024 • 0 new comments

• [1/N] Apply py39 ruff fixes

  #138578 commented on Oct 29, 2024 • 0 new comments

• [1/N] Use isinstance in python code

  #138573 commented on Oct 29, 2024 • 0 new comments

• add CUDA 12.6 to docker for sbsa wheel

  #138562 commented on Oct 28, 2024 • 0 new comments

• Add getDeviceAllocator device agnostic API in accelerator hooks

  #137972 commented on Oct 24, 2024 • 0 new comments

• [WIP] avoid atomic add for XPU device in satter_add by deterministic mode

  #137966 commented on Oct 24, 2024 • 0 new comments

• [FlexAttention] Add sm86 config

  #137959 commented on Oct 29, 2024 • 0 new comments

• [Dynamo] allow dynamic callables on tensor variables

  #137940 commented on Oct 29, 2024 • 0 new comments

• Add more DSA

  #137934 commented on Oct 24, 2024 • 0 new comments

• update _unsafe_set_version_counter to accept lists of tensors

  #137921 commented on Oct 29, 2024 • 0 new comments

• Dont decompose aten.baddmm in inductor

  #137904 commented on Oct 30, 2024 • 0 new comments

• [WIP] Update triton xpu commit pin

  #137886 commented on Oct 30, 2024 • 0 new comments

• [POC][functorch] use public PyTree API in `torch.func`

  #137884 commented on Oct 30, 2024 • 0 new comments

• [BE]: Try turning on LTO in CMake in CI

  #137866 commented on Oct 30, 2024 • 0 new comments

• Turned Views, Layouts, and Loops into proper dataclasses

  #137861 commented on Oct 28, 2024 • 0 new comments

• Add Intel GPU info collection to the collect env script

  #137846 commented on Oct 29, 2024 • 0 new comments

• Fix for gcc10 torch.compile compiler error when march=aarch64+sve

  #137795 commented on Oct 30, 2024 • 0 new comments

• Refactor FxGraphDrawer to use HTML-like labels

  #137726 commented on Oct 25, 2024 • 0 new comments

• Pass all arguments when quantizing embedding bag from float

  #137697 commented on Oct 29, 2024 • 0 new comments

• [FlexAttention] Skip Calculating KV Gradient if Possible

  #137658 commented on Oct 25, 2024 • 0 new comments

• Globally enable Python dispatcher for all of Inductor compilation

  #137621 commented on Oct 30, 2024 • 0 new comments

• Flip default on weights_only

  #137602 commented on Oct 30, 2024 • 0 new comments

• Make Context to be Device-agnostic Step by Step (4/N)

  #137580 commented on Oct 25, 2024 • 0 new comments

• Make Context to be Device-agnostic Step by Step (3/N)

  #137578 commented on Oct 25, 2024 • 0 new comments

• [Intel GPU] allow_tf32 context at XPU backend

  #137570 commented on Oct 29, 2024 • 0 new comments

• Remove depracated alias macro(3/3)

  #137562 commented on Oct 26, 2024 • 0 new comments

• Remove depracated alias macro(2/3)

  #137559 commented on Oct 26, 2024 • 0 new comments

• [inductor] modify the heuristic for loop split optimization

  #137550 commented on Oct 30, 2024 • 0 new comments

• [CUDAGraph] Add multithreading support for cudagraph trees

  #138282 commented on Oct 27, 2024 • 0 new comments

• Small fix to Python rendering in documentation.

  #138281 commented on Oct 25, 2024 • 0 new comments

• [PyTorch] Support non-zero beta in fp16_gemv_trans

  #138275 commented on Oct 29, 2024 • 0 new comments

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

  #138214 commented on Oct 30, 2024 • 0 new comments

• [FX][export][dynamo] use `tuple` instead of `list` exported code signature

  #138213 commented on Oct 25, 2024 • 0 new comments

• Rewrite _reparametrize_module to use `@contextmanager`

  #138203 commented on Oct 30, 2024 • 0 new comments

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

  #138202 commented on Oct 30, 2024 • 0 new comments

• Fuse partition lists reorder

  #138176 commented on Oct 29, 2024 • 0 new comments

• Update cusparselt to 0.6.3.2 version (0.6.3 for short).

  #138175 commented on Oct 24, 2024 • 0 new comments

• fix trace nn.parameters()

  #138149 commented on Oct 28, 2024 • 0 new comments

• [Pipelining] add schedule simulator and chrometrace dump

  #138134 commented on Oct 25, 2024 • 0 new comments

• fix test_float_to_int_conversion_nonfinite for NumPy 2

  #138131 commented on Oct 30, 2024 • 0 new comments

• [Pipelining] Support V-schedules in IR and Runtime

  #138125 commented on Oct 25, 2024 • 0 new comments

• [wip] "Python compiled autograd II"

  #138101 commented on Oct 29, 2024 • 0 new comments

• Propagate NJT lengths through op calls

  #138098 commented on Oct 25, 2024 • 0 new comments

• [DRAFT] do not look--testing cpp_extension no python

  #138088 commented on Oct 24, 2024 • 0 new comments

• [BE]: Improve typing storage

  #138084 commented on Oct 29, 2024 • 0 new comments

• handle more devices in method_type method of TensorVariable

  #138078 commented on Oct 30, 2024 • 0 new comments

• Use clang18 asan

  #138066 commented on Oct 24, 2024 • 0 new comments

• [POC][pytree] test C++ pytree in `torch.utils.pytree` by default

  #138056 commented on Oct 30, 2024 • 0 new comments

• [PyTorch] Hook up fp16_gemv_trans to gemv fast path for non-aarch64 architectures

  #138005 commented on Oct 29, 2024 • 0 new comments

• [CI] Move mkl install logic to `requirements-ci.txt`

  #137995 commented on Oct 29, 2024 • 0 new comments

• [BE]: Update CUDNN for Unix OSS to 9.5.1.17

  #137978 commented on Oct 27, 2024 • 0 new comments

• Fix layout for SetSourceTensorKernel

  #137973 commented on Oct 30, 2024 • 0 new comments

• [ONNX] Export Phi3.5 onnx graph multiple slice nodes missing starts or ends attribute

  #138637 commented on Oct 29, 2024 • 0 new comments

• FakeMode should not fakify non persistent buffer

  #107879 commented on Oct 28, 2024 • 0 new comments

• General MPS op coverage tracking issue

  #77764 commented on Oct 28, 2024 • 0 new comments

• CVE-2024-5480 reported by security analyzers

  #129228 commented on Oct 28, 2024 • 0 new comments

• Flex attention with mask depending on queries and keys lengths (or how to implement `causal_lower_right` masking)

  #137779 commented on Oct 28, 2024 • 0 new comments

• Problem with building PyTorch from source for ROCm on Ubuntu 24.04

  #137858 commented on Oct 28, 2024 • 0 new comments

• [ROCm] PyTorch Profiler Seg Fault on PyTorch Nightly

  #138360 commented on Oct 28, 2024 • 0 new comments

• FlopCounterMode doesn't support HOP

  #134385 commented on Oct 28, 2024 • 0 new comments

• Distributed Weighted Sampler.

  #77154 commented on Oct 28, 2024 • 0 new comments

• [ONNX] dynamo: support conditional op `cond`

  #117655 commented on Oct 28, 2024 • 0 new comments

• Add support for Flash Attention for AMD/ROCm

  #112997 commented on Oct 28, 2024 • 0 new comments

• Torch Dynamo support for Flux Transformer model

  #138195 commented on Oct 28, 2024 • 0 new comments

• DISABLED test_aot_export_cond_simple_cuda_float32 (__main__.TestHOPCUDA)

  #123096 commented on Oct 28, 2024 • 0 new comments

• [v2.5.1] Release Tracker

  #138613 commented on Oct 28, 2024 • 0 new comments

• [FSDP2] Eager-Mode Execution Tracker

  #120003 commented on Oct 28, 2024 • 0 new comments

• DISABLED test_save_with_without_initializer_dont_include_initializer_no_fake_mode_no_exported_program (__main__.TestFxToOnnx)

  #125020 commented on Oct 28, 2024 • 0 new comments

• AOT compilation fails for nested models using torchvision V2 API

  #137743 commented on Oct 28, 2024 • 0 new comments

• Compiling a module leads to `AssertionError: expected size 64==64, stride 1==49 at dim=1`

  #136837 commented on Oct 28, 2024 • 0 new comments

• torch._inductor.exc.LoweringException in `torch._export.aot_compile`

  #138254 commented on Oct 28, 2024 • 0 new comments

• Build errors while building PyTorch with BLIS

  #134399 commented on Oct 28, 2024 • 0 new comments

• Infinite cycle on BSC tensor: base -> values -> base

  #122089 commented on Oct 28, 2024 • 0 new comments

• [Dynamo] Eager fallback casued by graph breaks in module hooks

  #135410 commented on Oct 28, 2024 • 0 new comments

• [PTD BE DAY]Burn Down Distributed Disabled Tests!!

  #132845 commented on Oct 28, 2024 • 0 new comments

• The latest version of pytorch cannot export onnx with loop operator

  #122588 commented on Oct 27, 2024 • 0 new comments

• Obscure error: Expected a value of type 'List[int]' for argument 'sizes' but instead found type 'immutable_list'

  #122129 commented on Oct 27, 2024 • 0 new comments

• _tensors_definitely_do_not_overlap guard explosion

  #118214 commented on Oct 27, 2024 • 0 new comments

• [torch.export] Detect internal constrains

  #136216 commented on Oct 27, 2024 • 0 new comments

• Pruning step for users variable in Scheduler

  #138721 commented on Oct 27, 2024 • 0 new comments

• [RFC] Cuda support matrix for Release 2.6

  #138609 commented on Oct 29, 2024 • 0 new comments

• [torch.compile] Integers stored on nn.Modules as dynamic causing errors

  #133166 commented on Oct 29, 2024 • 0 new comments

• vmap, jacrev, jacfwd, hessian, etc., in libTorch

  #106455 commented on Oct 29, 2024 • 0 new comments

• torch._dynamo.exc.Unsupported: Unexpected type in sourceless builder torch.Tensor when running Mamba models in vLLM

  #136497 commented on Oct 29, 2024 • 0 new comments

• `torch.onnx.export` doesn't correctly constfold constants and DequantLinear

  #123628 commented on Oct 29, 2024 • 0 new comments

• [Intel GPU] Lower FLOPs and Bandwidth on Arc 770

  #136342 commented on Oct 29, 2024 • 0 new comments

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

  #24422 commented on Oct 29, 2024 • 0 new comments

• Using AOTInductor in C++ crashes if the model uses torch.linalg.eigh with CUDA

  #138601 commented on Oct 29, 2024 • 0 new comments

• [Inductor] unaligned variable ranges during node fusion

  #138550 commented on Oct 29, 2024 • 0 new comments

• [RFC] Per-Parameter-Sharding FSDP

  #114299 commented on Oct 29, 2024 • 0 new comments

• `maybe_mark_dynamic` causes max recursion error when used with compile during tensordict consolidation

  #138729 commented on Oct 29, 2024 • 0 new comments

• dynamo(eval_frame.py) failed on Windows UT.

  #132561 commented on Oct 29, 2024 • 0 new comments

• dynamo IndexError: list index out of range on Windows UT

  #132569 commented on Oct 29, 2024 • 0 new comments

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

  #130829 commented on Oct 29, 2024 • 0 new comments

• Microsoft Visual C++ Redistributable is not installed, this may lead to the DLL load failure.

  #126507 commented on Oct 29, 2024 • 0 new comments

• torch.onnx.dynamo_export fails to convert torchaudio.transforms.MFCC to onnx

  #125375 commented on Oct 29, 2024 • 0 new comments

• Missing librt.so.1, libdl.so.2, libz.so.1 and libpthread.so.0 in PyTorch Source Build (aarch64)

  #137078 commented on Oct 29, 2024 • 0 new comments

• Direct Implementation of K-Nearest neighbor (KNN) in pytorch

  #71386 commented on Oct 29, 2024 • 0 new comments

• Add support for immutable tensors in torch.export

  #136642 commented on Oct 29, 2024 • 0 new comments

• `_amp_foreach_non_finite_check_and_unscale_` can be torch.compiled inside torch.amp, but not in identical code outside it

  #138412 commented on Oct 29, 2024 • 0 new comments

• ROCm: /usr/bin/ld: failed to convert GOTPCREL relocation

  #138427 commented on Oct 29, 2024 • 0 new comments

• DISABLED test_constant_insertion (__main__.TestJit)

  #96894 commented on Oct 29, 2024 • 0 new comments

• Support for `uint16`, `uint32`, and `uint64`

  #58734 commented on Oct 28, 2024 • 0 new comments

• torch.compile(create_block_mask) errors in certain cases and hangs in others

  #138514 commented on Oct 28, 2024 • 0 new comments

• getting "handle_foreach_pow_scalar() got multiple values for argument 'self'" since PT2.5

  #138698 commented on Oct 28, 2024 • 0 new comments

• TorchInductor CPU Performance Dashboard

  #93531 commented on Oct 28, 2024 • 0 new comments

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

  #127167 commented on Oct 28, 2024 • 0 new comments

• DISABLED test_serialize_export_auto_functionalize_simple_cuda_float32 (__main__.TestHOPCUDA)

  #123563 commented on Oct 24, 2024 • 0 new comments

• Investigate potential cost savings for inductor workflows

  #138476 commented on Oct 24, 2024 • 0 new comments

• [export] Cannot view a tensor with shape torch.Size([1, 512, 32, 128]) and strides (2097152, 128, 65536, 1) as a tensor with shape (1, 512, 4096)

  #136543 commented on Oct 24, 2024 • 0 new comments

• RFC: Integration of KleidiAI 4-Bit MatMul Kernels into PyTorch

  #137830 commented on Oct 24, 2024 • 0 new comments

• torch.compile+cudagraphs asserts in multithreaded context

  #123177 commented on Oct 24, 2024 • 0 new comments

• More comprehensive debug logging of intermediate FX graphs for Inductor passes

  #118123 commented on Oct 24, 2024 • 0 new comments

• inductor error with PT Lightning + FSDP + torchao.float8 + torch.compile

  #138715 commented on Oct 24, 2024 • 0 new comments

• with torch.cuda.amp.autocast() get out of memory error when using with torch.no_grad() during validation

  #45910 commented on Oct 24, 2024 • 0 new comments

• Performance regression in torch.compile

  #136254 commented on Oct 24, 2024 • 0 new comments

• `torch.bernoulli` and `torch.rand` to support `dtype=torch.bool` kwarg (and sypport operation of `tensor.uniform_` on bool tensors)

  #35527 commented on Oct 24, 2024 • 0 new comments

• DataLoader num_workers > 0 causes CPU memory from parent process to be replicated in all worker processes

  #13246 commented on Oct 24, 2024 • 0 new comments

• [CI] Continuous spam of PyTorch CI build status notifications in fork

  #138564 commented on Oct 24, 2024 • 0 new comments

• [inductor] cpp gemm autotune doesn't work on AMD EPYC

  #138718 commented on Oct 24, 2024 • 0 new comments

• `export()` fails for `full((n,), v)` but succeeds for `ones((n,)) * v` where `v` is dynamic

  #138073 commented on Oct 24, 2024 • 0 new comments

• DISABLED test_dropout_deterministic_cuda (__main__.GPUTests)

  #133025 commented on Oct 24, 2024 • 0 new comments

• NCCL watchdog thread terminated with exception

  #113128 commented on Oct 24, 2024 • 0 new comments

• dynamic control flow failed

  #138689 commented on Oct 24, 2024 • 0 new comments

• Less restrictive guards when using int(log2(sz))

  #137220 commented on Oct 23, 2024 • 0 new comments

• [ONNX]: Fail to export onnx when GroupNorm input shape rank=2

  #130108 commented on Oct 23, 2024 • 0 new comments

• dead code in test that can be removed

  #138673 commented on Oct 23, 2024 • 0 new comments

• The default setting of PyTorch should show warning always.

  #138552 commented on Oct 23, 2024 • 0 new comments

• [CI/CD] Deprecating PyTorch’s official Anaconda channel

  #138696 commented on Oct 23, 2024 • 0 new comments

• Nightly job `linux-binary-manywheel / manywheel-py3_10-xpu-build / build` is currently broken

  #138695 commented on Oct 23, 2024 • 0 new comments

• Flex attention underperforms SDPA (cuDNN), constructing T5 attention bias via embedding weights

  #138493 commented on Oct 23, 2024 • 0 new comments

• torch.compile silently treats None argument passed to Triton kernel as a `*i8`

  #138546 commented on Oct 23, 2024 • 0 new comments

• [Validation] pypi binaries with slimmed dependencies are usable in standard AWS containers (amazonlinux:2023 regression in 1.13 and 2.5)

  #138482 commented on Oct 23, 2024 • 0 new comments

• Fusion causes peak memory increase

  #138685 commented on Oct 23, 2024 • 0 new comments

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

  #135704 commented on Oct 27, 2024 • 0 new comments

• load_state_dict unexpectedly does not load Tensor to buffers that currently have None value

  #8104 commented on Oct 27, 2024 • 0 new comments

• Tensor loses `bool` method during scripting

  #70544 commented on Oct 27, 2024 • 0 new comments

• Investigate torch.compile Windows support.

  #122094 commented on Oct 26, 2024 • 0 new comments

• [Performance] [CuDNN-Attention] CuDNN backend should return the output in the same stride order as input Query

  #138340 commented on Oct 26, 2024 • 0 new comments

• Cleanup stale Dynamo feature flags

  #136862 commented on Oct 26, 2024 • 0 new comments

• The doc of `linalg.vector_norm()` should not say `ord` parameter accepts the `str` value `fro` or `nuc`

  #136563 commented on Oct 26, 2024 • 0 new comments

• torch 2.5 slower than 2.4.1 ?

  #138386 commented on Oct 26, 2024 • 0 new comments

• DISABLED test_save_load_checkpoint (__main__.DistributedDataParallelTest)

  #137771 commented on Oct 25, 2024 • 0 new comments

• Enable CUDA 12.6.2

  #138440 commented on Oct 25, 2024 • 0 new comments

• DISABLED test_device_mode_ops_sparse_mm_reduce_cpu_bfloat16 (__main__.TestDeviceUtilsCPU)

  #132494 commented on Oct 25, 2024 • 0 new comments

• [RFC] Add new CPP builder for inductor on pytorch Windows

  #124245 commented on Oct 25, 2024 • 0 new comments

• SDPA 2.5 Issue tracking

  #138649 commented on Oct 25, 2024 • 0 new comments

• DISABLED test_is_isnot (__main__.TestScript)

  #120694 commented on Oct 25, 2024 • 0 new comments

• [dynamo] Restart mechanism for unboxing cell optimization can hide speculation log divergence

  #138491 commented on Oct 25, 2024 • 0 new comments

• Batching rule for aten::bincount.

  #105912 commented on Oct 25, 2024 • 0 new comments

• get_model_state_dict failed after FSDP_model.to(dtype)

  #138467 commented on Oct 25, 2024 • 0 new comments

• [RFC] Intel GPU Distributed Support in PyTorch

  #135791 commented on Oct 25, 2024 • 0 new comments

• Pytorch Error: ImportError: /home/dp/anaconda3/envs/ComfyUI/lib/python3.11/site-packages/torch/lib/../../nvidia/cusparse/lib/libcusparse.so.12: undefined symbol: __nvJitLinkComplete_12_4, version libnvJitLink.so.12

  #131312 commented on Oct 25, 2024 • 0 new comments

• Support writing tensorboard traces to AWS S3 (and other cloud storage services) in profiler

  #73131 commented on Oct 25, 2024 • 0 new comments

• Libtorch C++ register_module raise "read access violation error"

  #116568 commented on Oct 25, 2024 • 0 new comments

• Building project using libtorch results in "Failed to find nvToolsExt"

  #116242 commented on Oct 24, 2024 • 0 new comments

• AOT inductor can't handle dynamic shape 2*s1

  #137748 commented on Oct 24, 2024 • 0 new comments

• Implement missing torch.nan* operators

  #61474 commented on Oct 24, 2024 • 0 new comments

• [export] Torch custom class export issue

  #138344 commented on Oct 24, 2024 • 0 new comments

• DISABLED test_assigning_back_deleter_fns_to_tensor (__main__.TestBlockStateAbsorption)

  #134810 commented on Oct 24, 2024 • 0 new comments

• DISABLED test_retrace_export_auto_functionalize_simple_cuda_float32 (__main__.TestHOPCUDA)

  #123449 commented on Oct 24, 2024 • 0 new comments

• foreach global norm

  #134327 commented on Oct 29, 2024 • 0 new comments

• Temp disable MKL in DistributionKernels.cpp

  #132532 commented on Oct 26, 2024 • 0 new comments

• [WIP] Refactor distributed code via device-agnostic API

  #132371 commented on Oct 30, 2024 • 0 new comments

• Add pybind support for is_hpu property of Tensor object

  #132228 commented on Oct 24, 2024 • 0 new comments

• [torch.special] Adding betainc with backward operation

  #132135 commented on Oct 27, 2024 • 0 new comments

• Add Weighted Loss Functions to PyTorch : WMSE, WMAE, and Weighted Huber Loss

  #132049 commented on Oct 30, 2024 • 0 new comments

• [xla hash update] update the pinned xla hash

  #132021 commented on Oct 28, 2024 • 0 new comments

• Added missing `__all__` in `__init__` files.

  #131800 commented on Oct 25, 2024 • 0 new comments

• track number of cpp->python exceptions thrown in torch.compile benchmark suite

  #131481 commented on Oct 29, 2024 • 0 new comments

• BE: reset dynamo before each test in test_ops_gradients.py

  #131397 commented on Oct 28, 2024 • 0 new comments

• Test module importability in subprocess

  #131317 commented on Oct 25, 2024 • 0 new comments

• [CI] enable operator benchmark on CPU

  #131305 commented on Oct 25, 2024 • 0 new comments

• cmake looks for FP16_PATH environment variable

  #130865 commented on Oct 28, 2024 • 0 new comments

• [1/N] Fix sign comparison errors in CUDA code

  #130718 commented on Oct 27, 2024 • 0 new comments

• LOG(INFO) -> VLOG(2) in ProcessGroupNCCL

  #130696 commented on Oct 30, 2024 • 0 new comments

• [ROCm] Unskip functorch efficient attention tests

  #130481 commented on Oct 28, 2024 • 0 new comments

• [pytree] implement key path APIs for CXX pytree

  #130141 commented on Oct 30, 2024 • 0 new comments

• [pytree] preserve `dict` keys in insertion order in CXX pytree

  #130140 commented on Oct 30, 2024 • 0 new comments

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

  #129887 commented on Oct 24, 2024 • 0 new comments

• [BE][Easy] Fix `PYI034`: non-self-return-type in tensor method hints

  #129886 commented on Oct 24, 2024 • 0 new comments

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

  #129873 commented on Oct 24, 2024 • 0 new comments

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

  #129872 commented on Oct 24, 2024 • 0 new comments

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

  #129871 commented on Oct 24, 2024 • 0 new comments

• [PT2E Quantization] Fix RecursionError when prepare_pt2e graph with concat of the same node

  #129567 commented on Oct 30, 2024 • 0 new comments

• Use Generic TypeAlias (PEP 585) and Union Type (PEP 604) in generated `.pyi` stub files

  #129420 commented on Oct 25, 2024 • 0 new comments

• Use absolute path `path.resolve()` -> `path.absolute()`

  #129409 commented on Oct 25, 2024 • 0 new comments

• [BE][Easy] use `pathlib.Path` instead of `dirname` / `".."` / `pardir`

  #129374 commented on Oct 25, 2024 • 0 new comments

• Fix numerical instability for norm

  #129352 commented on Oct 24, 2024 • 0 new comments

• Fix unbind_copy and add its decomposition

  #134319 commented on Oct 29, 2024 • 0 new comments

• [UT] Add test about PrivateUse1, its module must be registered before calling device-related APIs

  #134318 commented on Oct 26, 2024 • 0 new comments

• [profiler][UT] instantiate profiler UTs for devices and enable UTs for xpu profiler

  #134316 commented on Oct 29, 2024 • 0 new comments

• [CI][dashboard][experiment][not for landing] Change aarch64 to measure fp16

  #134282 commented on Oct 25, 2024 • 0 new comments

• [export] enumerate unsupported sympy.Functions

  #134271 commented on Oct 26, 2024 • 0 new comments

• [Dynamo][autograd.Function] Support ctx.setup_context properly and compose with vmap

  #134256 commented on Oct 28, 2024 • 0 new comments

• [DO NOT LAND] DTensor + GradScaler: DTensor dispatch without changing amp

  #134185 commented on Oct 25, 2024 • 0 new comments

• Enable c10d ops for XPU device using MPI backend

  #134132 commented on Oct 26, 2024 • 0 new comments

• cpuinfo git branch fix (main instead of master)

  #134101 commented on Oct 26, 2024 • 0 new comments

• [export] don't set complex replacements on hybrid symints

  #133920 commented on Oct 23, 2024 • 0 new comments

• [dynamo] refactor `builtins.zip` to use polyfill

  #133895 commented on Oct 24, 2024 • 0 new comments

• [dynamo][itertools] refactor `itertools.count` to use polyfill

  #133875 commented on Oct 24, 2024 • 0 new comments

• [cpu] Modify inductor opt flag --- ffast-math

  #133842 commented on Oct 29, 2024 • 0 new comments

• [dynamo] support `functools.cmp_to_key`

  #133805 commented on Oct 30, 2024 • 0 new comments

• Ensure SWA boundary conditions w.r.t. definition

  #133773 commented on Oct 28, 2024 • 0 new comments

• Allow mp.start_processes to create processes in parallel

  #133707 commented on Oct 26, 2024 • 0 new comments

• Implements user buffer registration using MemPool

  #133603 commented on Oct 29, 2024 • 0 new comments

• [BE]: Update NCCL submodule to 2.22.3

  #133593 commented on Oct 29, 2024 • 0 new comments

• Bijection and Jacobian for lower Cholesky and positive definite transforms

  #133500 commented on Oct 27, 2024 • 0 new comments

• Remove unused Python variables in torch/[_-a]*

  #133492 commented on Oct 26, 2024 • 0 new comments

• [Inductor][Do NOT Review] Inplacing with Donated Buffer

  #133368 commented on Oct 29, 2024 • 0 new comments

• [Intel GPU] qlinear at XPU backend

  #133307 commented on Oct 30, 2024 • 0 new comments

• [BE] enable `ruff` rule series `PIE` from `flake8-pie`

  #133202 commented on Oct 25, 2024 • 0 new comments

• [BE][Easy] enable `ruff` rule `PIE808`: unnecessary `start` argument in `range(...)`

  #133201 commented on Oct 25, 2024 • 0 new comments

• autograd codegen: bump VC properly for mutable ops with no returns

  #133044 commented on Oct 29, 2024 • 0 new comments

• Improvements for associative_scan - vmap fixes

  #133013 commented on Oct 27, 2024 • 0 new comments

• S390x update builder image

  #132983 commented on Oct 29, 2024 • 0 new comments

• `(*bias): last dimension must be contiguous` when running compiled SDPA on length 1 tensors

  #138317 commented on Oct 30, 2024 • 0 new comments

• Contribution Proposal: ScatterND Implementation for PyTorch

  #138502 commented on Oct 30, 2024 • 0 new comments

• FlexAttention result deviates with torch.compile() and torch.set_float32_matmul_precision('high')

  #138556 commented on Oct 30, 2024 • 0 new comments

• [export] Failed to save the model using torch.export.save

  #136897 commented on Oct 30, 2024 • 0 new comments

• [export] Dynamic shape torch-trt models fail on torch.export.load

  #137365 commented on Oct 30, 2024 • 0 new comments

• DISABLED test_comprehensive__unsafe_masked_index_cuda_bool (__main__.TestInductorOpInfoCUDA)

  #131748 commented on Oct 30, 2024 • 0 new comments

• Easier way of seeing dynamic shape specializations in C++?

  #137657 commented on Oct 30, 2024 • 0 new comments

• tensor.triu_(1) not working properly with large matrix

  #136611 commented on Oct 30, 2024 • 0 new comments

• FSDP.forward() fails "_is_root should not have been set" error after saving a distributed checkpoint

  #113496 commented on Oct 30, 2024 • 0 new comments

• Sequence Ops usage when exporting embedding bag into onnx

  #138485 commented on Oct 30, 2024 • 0 new comments

• ninja: build stopped: subcommand failed.

  #108209 commented on Oct 30, 2024 • 0 new comments

• Increased memory usage with AMP

  #61173 commented on Oct 30, 2024 • 0 new comments

• DISABLED test_comprehensive_fft_ifft_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #127344 commented on Oct 30, 2024 • 0 new comments

• [MPS] Inconsistent performance issues

  #136003 commented on Oct 30, 2024 • 0 new comments

• torch._dynamo.exc.Unsupported: builtin: bool [<class 'torch._dynamo.variables.tensor.SymNodeVariable'>] False

  #136075 commented on Oct 30, 2024 • 0 new comments

• take_along_dim or gather unstable results on cpu with stride 1

  #129093 commented on Oct 30, 2024 • 0 new comments

• TORCH_COMPILE_CPROFILE does not work for python 3.12

  #137869 commented on Oct 30, 2024 • 0 new comments

• Python 3.13 support for PyTorch

  #130249 commented on Oct 29, 2024 • 0 new comments

• [Tracker] Nested tensor op coverage requests

  #118107 commented on Oct 29, 2024 • 0 new comments

• AOT eager accuracy regression in Segformer in 2.5.0 release

  #138652 commented on Oct 29, 2024 • 0 new comments

• Diffusers import change breaking newly built (non-cached) test-infra dockers.

  #138591 commented on Oct 29, 2024 • 0 new comments

• `Calculate Docker Image` step times out if run on LF runners

  #138042 commented on Oct 29, 2024 • 0 new comments

• vmap with torch.autograd.grad does not work on output of compiled function

  #138422 commented on Oct 29, 2024 • 0 new comments

• `Check label` failures are not very informative

  #137927 commented on Oct 29, 2024 • 0 new comments

• [CI][LF] Docker builds to new ECR fails

  #137361 commented on Oct 29, 2024 • 0 new comments

• There is a performance drop because we have not yet implemented the batching rule for aten::native_dropout_backward. Please file us an issue on GitHub so that we can prioritize its implementation

  #122432 commented on Oct 29, 2024 • 0 new comments

• torch.stft sometimes raises RuntimeError: cuFFT error: CUFFT_INTERNAL_ERROR on low free memory

  #119420 commented on Oct 29, 2024 • 0 new comments

• Provide a method to unregister privateuse1

  #129056 commented on Oct 29, 2024 • 0 new comments

• [inductor] refine loop split logic

  #128812 commented on Oct 25, 2024 • 0 new comments

• [Inductor] support masked vectorization for the tail_loop for integer datatypes and bool datatype

  #128802 commented on Oct 30, 2024 • 0 new comments

• Remove caffe2 namespace from c10/macros/Macros.h

  #128672 commented on Oct 29, 2024 • 0 new comments

• Add MaskedTensor support to *_like API

  #128637 commented on Oct 28, 2024 • 0 new comments

• Adds support for accelerated sorting with x86-simd-sort

  #127936 commented on Oct 23, 2024 • 0 new comments

• Reuse UT for Intel GPU backend [Part1]

  #127602 commented on Oct 29, 2024 • 0 new comments

• [Intel GPU]Enable fp64 double GEMM

  #127508 commented on Oct 26, 2024 • 0 new comments

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

  #127294 commented on Oct 30, 2024 • 0 new comments

• Enable Leak Sanitizer

  #127171 commented on Oct 30, 2024 • 0 new comments

• [inductor] online softmax

  #127011 commented on Oct 28, 2024 • 0 new comments

• Add NHWC support for group normalization

  #126635 commented on Oct 30, 2024 • 0 new comments

• [vision hash update] update the pinned vision hash

  #125806 commented on Oct 30, 2024 • 0 new comments

• Fix hardcoded rocm warp size

  #125433 commented on Oct 28, 2024 • 0 new comments

• S390x ci periodic tests

  #125401 commented on Oct 24, 2024 • 0 new comments

• [DO NOT MERGE] Test new ROCm CI nodes

  #124424 commented on Oct 25, 2024 • 0 new comments

• Improve decomposition for constant_pad_nd

  #123661 commented on Oct 30, 2024 • 0 new comments

• New improved Conv3D implementation for MPS and support for ConvTranspose3D

  #116580 commented on Oct 24, 2024 • 0 new comments

• Conversions between strided and jagged layouts for Nested Tensors

  #115749 commented on Oct 29, 2024 • 0 new comments

• [triton hash update] update the pinned triton hash

  #115529 commented on Oct 28, 2024 • 0 new comments

• Automated submodule update: FBGEMM

  #115316 commented on Oct 30, 2024 • 0 new comments

• [pytree] support PyStructSequence types for Python pytree

  #113258 commented on Oct 30, 2024 • 0 new comments

• [pytree] add APIs to determine a class is a namedtuple or PyStructSequence

  #113257 commented on Oct 30, 2024 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on Oct 29, 2024 • 0 new comments

• DISABLED test_device_mode_ops_sparse_mm_reduce_cpu_float64 (__main__.TestDeviceUtilsCPU)

  #132619 commented on Oct 30, 2024 • 0 new comments

• DISABLED test_input_mutation2_dynamic_shapes_cpu (__main__.DynamicShapesCpuTests)

  #135295 commented on Oct 30, 2024 • 0 new comments

• ROCm loses some supported GPUs by requiring hipblaslt

  #119081 commented on Oct 30, 2024 • 0 new comments