1 Release published by 1 person

• v2.3.0 PyTorch 2.3: User-Defined Triton Kernels in torch.compile, Tensor Parallelism in Distributed

  published Apr 24, 2024

2 Pull requests merged by 2 people

• [EZ] Get rid of utf-8 quotes

  #124932 merged Apr 25, 2024

• Specify the exact table we upload metrics to

  #124321 merged Apr 23, 2024

164 Pull requests opened by 105 people

• Add basic sanity checks for graph ops to cache key

  #124745 opened Apr 23, 2024

• Fix issue 112919

  #124746 opened Apr 23, 2024

• Update descriptor fields to resolve fft precision issue

  #124756 opened Apr 23, 2024

• Allow device tensors that use numpy for serialization to use weights_only unpickler

  #124763 opened Apr 23, 2024

• [FSDP] Errored on wrapping `ModuleList`/`ModuleDict`

  #124764 opened Apr 23, 2024

• [DCP] Adds storage metadata, and passes it during the save path

  #124772 opened Apr 23, 2024

• Add Sanity Testing to Pytorch Profiler

  #124773 opened Apr 23, 2024

• [sparse] Fix type-dispatch errors

  #124777 opened Apr 23, 2024

• Attach stack traces to plain callables when using aot_export_joint_simple

  #124792 opened Apr 23, 2024

• Meta kernel for _pack_padded_sequence

  #124794 opened Apr 23, 2024

• [rfc]: vendor in open-telemetry

  #124800 opened Apr 23, 2024

• [NT] Support NestedTensor in is_concrete_int

  #124803 opened Apr 23, 2024

• Include support for the scatter gather cuda kernels to allow for comp…

  #124809 opened Apr 24, 2024

• Upgrade nightly wheels to rocm6.1

  #124811 opened Apr 24, 2024

• [DONT MERGE][dynamo] Turn on inlining of nn modules

  #124815 opened Apr 24, 2024

• Prevent rendezvous shutdown on worker restarts

  #124819 opened Apr 24, 2024

• [inductor] add uint8 SDPA pattern

  #124832 opened Apr 24, 2024

• [privateuse1] _refs.masked_fill support privateuse1 when value.device.type is cpu

  #124835 opened Apr 24, 2024

• update xla pin

  #124839 opened Apr 24, 2024

• [RFC] Switch from black to ruff fmt

  #124845 opened Apr 24, 2024

• [not for land] check CI with symint set_() fixes

  #124867 opened Apr 24, 2024

• [dtensor] implement shard dim change with alltoall

  #124872 opened Apr 24, 2024

• [pipelining] Add util and debug facilities

  #124875 opened Apr 24, 2024

• [dtensor] delete the old unused mesh_alltoall

  #124879 opened Apr 24, 2024

• [dynamo][eval_frame] Create a dynamic wrapper fn to avoid cache collisions

  #124881 opened Apr 24, 2024

• Added FixedQParam per-tensor observer for weight tensors

  #124883 opened Apr 24, 2024

• Add Efficient Attention support on ROCM

  #124885 opened Apr 24, 2024

• [easy] remove some unecessary windows skips for mmap tests

  #124891 opened Apr 24, 2024

• [export] handle weight sharing in FQN mapping + unflattening

  #124892 opened Apr 24, 2024

• test_cuda.py test_grad_scaling_autocast_fused_optimizers migration to OptimizerInfo

  #124893 opened Apr 24, 2024

• [MPS] Remove in place views (causes too many crashes)

  #124895 opened Apr 25, 2024

• Implemented isin_Tensor_Tensor_out for MPS backend

  #124896 opened Apr 25, 2024

• [optim]fix ut and sgd kernel

  #124904 opened Apr 25, 2024

• [optim] add fused_adagrad support for CPU device

  #124905 opened Apr 25, 2024

• [onnx.export] Avoid linear look up in env for exist_in_env

  #124909 opened Apr 25, 2024

• [onnx.export] Cache SetGraphInputTypeReliable

  #124912 opened Apr 25, 2024

• test

  #124916 opened Apr 25, 2024

• Modify device check in capturable optimizer to support more devices

  #124919 opened Apr 25, 2024

• remove empty partition

  #124920 opened Apr 25, 2024

• Support aten operations with out tensor

  #124926 opened Apr 25, 2024

• [Inductor max autotune] Make autotune_select_algorithm more robust

  #124928 opened Apr 25, 2024

• [Inductor cutlass backend] Remove epilogue nodes from Kernel call

  #124929 opened Apr 25, 2024

• [Inductor cutlass backend] Fix cutlass_utils.get_max_alignment() for strided layouts.

  #124930 opened Apr 25, 2024

• [DCP] Provides default AsyncStager

  #124939 opened Apr 25, 2024

• [DCP] Move async logic into filesystem for better encapsulation

  #124944 opened Apr 25, 2024

• [export] disable_forced_specializations

  #124949 opened Apr 25, 2024

• [pipelining] Add stage backward function

  #124958 opened Apr 25, 2024

• Wrap the test func with try/except to always call destroy_process_group

  #124961 opened Apr 25, 2024

• [dtensor] Make distribute_tensor support distributing DTensors

  #124962 opened Apr 25, 2024

• [wip] add comm params in et

  #124963 opened Apr 25, 2024

• PT2 Inductor ComboKernels

  #124969 opened Apr 25, 2024

• Delete erroneous print

  #124972 opened Apr 25, 2024

• [dynamo] Add ID_MATCH guards on inlined functions to force compilation on monkeypatching

  #124975 opened Apr 25, 2024

• Fix bfloat16 serialization for ONNXProgram.save

  #124977 opened Apr 25, 2024

• [dynamo] support torchbind object input

  #124978 opened Apr 25, 2024

• WIP: [Inductor] log fusion failure due to loop orders

  #124986 opened Apr 26, 2024

• [Distributed] [7/N] Fix clang-tidy warnings in torch/csrc/distributed/c10d

  #124987 opened Apr 26, 2024

• add meta for segment_reduce_backward

  #124988 opened Apr 26, 2024

• Print flexibility for tensor output

  #124991 opened Apr 26, 2024

• Remove Inductor IRs for legacy functional collectives

  #124992 opened Apr 26, 2024

• Enable UFMT on `test/test_datapipe.py`

  #124994 opened Apr 26, 2024

• [BE] Remove JNI from libtorch builds

  #124995 opened Apr 26, 2024

• [ROCm] Enable int_mm_error tests for rocm 6.0+

  #124999 opened Apr 26, 2024

• [Torch] Add more mm kernel choices

  #125000 opened Apr 26, 2024

• Convert `ForeachFuncInfo` to `dataclass`

  #125001 opened Apr 26, 2024

• Fix process group initialize twice error in distributed launcher test

  #125006 opened Apr 26, 2024

• Adding Compare in torch.utils.benchmark documentation

  #125009 opened Apr 26, 2024

• [DataLoader] Select available CUDA or 3rd devices automatically to pin memory

  #125016 opened Apr 26, 2024

• [AMD] [Draft] New inductor gemm configs

  #125017 opened Apr 26, 2024

• [BE] Remove static files from libtorch builds

  #125027 opened Apr 26, 2024

• [BE]: Update ruff to v0.4.2

  #125031 opened Apr 26, 2024

• [WIP] Introduce bind_unbacked HOP

  #125034 opened Apr 26, 2024

• [Profiler] Add TSC Clock Callback to CUPTI

  #125036 opened Apr 26, 2024

• [Torch][Timer] Skip expired timer logging for empty expired timers

  #125039 opened Apr 26, 2024

• [unwind] replace LONG_LONG_MAX by the portable LLONG_MAX

  #125043 opened Apr 26, 2024

• Remove caffe2 image and video

  #125045 opened Apr 26, 2024

• [WIP] Added a function to calculate the deterministic version of MaxPool3D …

  #125048 opened Apr 26, 2024

• [TD] Enable td on cpu windows

  #125049 opened Apr 26, 2024

• [Draft] Warning message when user calls unscriptable component

  #125053 opened Apr 26, 2024

• [aoti] Add pt2 package save/load (python)

  #125054 opened Apr 26, 2024

• [BE] Make macos build libtorch with wheel

  #125060 opened Apr 26, 2024

• [dynamo, 3.12] xfail refleaking tests due to buggy getattr_static

  #125062 opened Apr 26, 2024

• Skip ONNX optimization when it fails

  #125063 opened Apr 26, 2024

• Force specialization of bool in scaled_dot_product_attention

  #125067 opened Apr 26, 2024

• [ROCm] Implement forward AD for miopen_batch_norm

  #125069 opened Apr 26, 2024

• Updating optims and combining torch functions

  #125071 opened Apr 26, 2024

• Forcing specialization of bool from symbool

  #125072 opened Apr 26, 2024

• [DTensor] allow numel 1 tensor operand to be implicitly replicate DTensor

  #125073 opened Apr 26, 2024

• Adding state_dicts test to adam_test

  #125076 opened Apr 26, 2024

• forward fix preferred blas backend and windows CI

  #125080 opened Apr 26, 2024

• add uuid in cudaDeviceProperties

  #125083 opened Apr 26, 2024

• [dnl] add NCCL/PT debug log for S413673

  #125085 opened Apr 27, 2024

• [WIP] Implement matrix_exp Batching Rule

  #125086 opened Apr 27, 2024

• [inductor][easy] add buffer layout to SchedulerNode.debug_str

  #125088 opened Apr 27, 2024

• [inductor] add triton code to SchedulerNode.debug_str

  #125089 opened Apr 27, 2024

• Remove caffe2 db

  #125092 opened Apr 27, 2024

• [Test][Distributed] Make more tests multi-threaded.

  #125095 opened Apr 27, 2024

• [pruning]add dropout to list of supported activation functions

  #125101 opened Apr 27, 2024

• Enable clang-tidy coverage on torch/csrc/distributed/c10d/*

  #125102 opened Apr 27, 2024

• [WIP] make torch.amp.autocast more generic

  #125103 opened Apr 27, 2024

• Add extra cuda_to_hip_mappings.py

  #125108 opened Apr 27, 2024

• Allow linalg.lstsq to use svd to compute the result for rank deficient matrices.

  #125110 opened Apr 28, 2024

• Enable UFMT on test_indexing&test_view_ops

  #125112 opened Apr 28, 2024

• Use BFloat16 in distributed quantization when supported by NCCL

  #125113 opened Apr 28, 2024

• Move autocast op list to autocast_mode.h to make sure other backends can reuse it.

  #125114 opened Apr 28, 2024

• Add propagate_real_tensors mode for unbacked

  #125115 opened Apr 28, 2024

• Enable UFMT on `test_decomp.py`, `test_expanded_weights.py` and some files

  #125117 opened Apr 28, 2024

• Refactor and Fix Some Prombles on Autocast

  #125118 opened Apr 28, 2024

• [inductor] Check if n is the input tensor of conv_pointwise

  #125119 opened Apr 28, 2024

• [Storage_ipc] Provides IPC extensions for 3rd devices.

  #125122 opened Apr 28, 2024

• Updated test_graph_optims and test_graph_scaling_fused_optimizers to use new OptimizerInfo infrastructure

  #125127 opened Apr 28, 2024

• Fix exception handling in torch._dynamo.utils.same and add corresponding test

  #125132 opened Apr 29, 2024

• Fix bug in graph partitioner and update graph signature after partitioning.

  #125133 opened Apr 29, 2024

• [PT2][Optimus] Read the patterns from the config instead of hard-code passes

  #125136 opened Apr 29, 2024

• Add templated attention BLOCK_M & BLOCK_N default size for different head_dim

  #125139 opened Apr 29, 2024

• Remove Caffe2 python

  #125143 opened Apr 29, 2024

• Fix AttributeError when doing mock patch for FileTimerServerTest.test_expired_timers

  #125144 opened Apr 29, 2024

• add avx512 specialization for vec_shuffle_down

  #125147 opened Apr 29, 2024

• save the reciprocal of weights for welford_reduce

  #125148 opened Apr 29, 2024

• [inductor][cpp] move some common cpp utils to cpp_utils.py

  #125152 opened Apr 29, 2024

• Merge the pyi files into py files of optimizer

  #125153 opened Apr 29, 2024

• Test benchmark suite with data dependent options on for dynamic shapes

  #125156 opened Apr 29, 2024

• WIP save to cache

  #125157 opened Apr 29, 2024

• [inductor] autotune benchmark support for cpu

  #125159 opened Apr 29, 2024

• Renable running before

  #125160 opened Apr 29, 2024

• [MPS] And naive int8 and int4 Linear

  #125163 opened Apr 29, 2024

• temp

  #125164 opened Apr 29, 2024

• ignore verify placeholders

  #125165 opened Apr 29, 2024

• fix more invalid inputs, remove fuse

  #125166 opened Apr 29, 2024

• Disable running before and update error message to be less verbose

  #125167 opened Apr 29, 2024

• Enable running before

  #125168 opened Apr 29, 2024

• Enable fuse

  #125169 opened Apr 29, 2024

• [ATen][CUDA][AMP] Fix dtype mismatch in linalg_vector_norm

  #125175 opened Apr 29, 2024

• TEST: add some random logging

  #125176 opened Apr 29, 2024

• Add a space on APPEND to CUDA flags

  #125178 opened Apr 29, 2024

• Refactored _remove_auto_functionalization_from_graph_helper

  #125180 opened Apr 29, 2024

• Add `write_record_metadata` to PyTorchFileWriter

  #125184 opened Apr 29, 2024

• [export] Don't create a new fake mode if dynamo tracing

  #125185 opened Apr 29, 2024

• Use torch._check for safety assert in _reshape_view_helper

  #125187 opened Apr 29, 2024

• Don't short circuit if shape is same

  #125188 opened Apr 29, 2024

• [export] Fix for unflattening modules with duplicate tensors

  #125192 opened Apr 29, 2024

• [ez][CI] Move test_modules and test_schema_check off CI_SERIAL_LIST

  #125193 opened Apr 29, 2024

• Fix bug in get_update_constraint

  #125194 opened Apr 29, 2024

• Fix mem size mismatch from split/chunk in const folding

  #125199 opened Apr 29, 2024

• [compiled autograd] compile fwd in inference mode

  #125201 opened Apr 29, 2024

• [DONT MERGE][FOR CI][dynamo] Turn on guard_nn_modules

  #125202 opened Apr 29, 2024

• [dynamo] support inactive context managers across graph breaks

  #125203 opened Apr 30, 2024

• FP8 rowwise scaling

  #125204 opened Apr 30, 2024

• Fix PT2E Dynamic Quant regression

  #125207 opened Apr 30, 2024

• [quant][pt2e] Fix conv-bn weight + bias per channel QAT

  #125208 opened Apr 30, 2024

• Export `torch.jit.interface` from `torch.jit` package

  #125209 opened Apr 30, 2024

• [Don't Merge] dump indoctor build command.

  #125210 opened Apr 30, 2024

• Enable AOTI shim v2 build and add into libtorch

  #125211 opened Apr 30, 2024

• Intel GPU: specify the tolerance for torchbench models

  #125213 opened Apr 30, 2024

• fix loading optimizer options from archive

  #125215 opened Apr 30, 2024

• Require nnz==0 in sparse meta tensors

  #125221 opened Apr 30, 2024

• [autocast] using new autocast api with device name.

  #125225 opened Apr 30, 2024

• fix: typo

  #125226 opened Apr 30, 2024

• Fix logic to find sbgemm in BLAS library

  #125227 opened Apr 30, 2024

• Fix random_mps_impl to accept non-contiguous tensors

  #125231 opened Apr 30, 2024

• Change to fix cuda python on corp based cluster (#117789)

  #125232 opened Apr 30, 2024

• [AOTI] Update C-shim codegen to handle rocm

  #125233 opened Apr 30, 2024

• [ncclx] Rename NCCL-EXP to NCCLX

  #125238 opened Apr 30, 2024

• [Inductor] Properly package target info for triton.compile

  #125241 opened Apr 30, 2024

116 Issues closed by 43 people

• Cudnn 9 is out!

  #119400 closed Apr 30, 2024

• Add option to `torch.load(mmap=True)` to do `MAP_SHARED` rather than `MAP_PRIVATE`

  #124528 closed Apr 30, 2024

• DISABLED test_comprehensive_nn_functional_conv1d_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #123874 closed Apr 30, 2024

• DISABLED test_super_resolution_cuda (__main__.TestModels)

  #105332 closed Apr 30, 2024

• DISABLED test_torchvision_models_efficientnet_v2_m (__main__.TestVisionTracing)

  #124152 closed Apr 30, 2024

• device_mesh / fsdp issue with _get_device_handle

  #124327 closed Apr 30, 2024

• Unable to use `torch.compile()` with triton's `TensorWrapper` in custom triton kernel

  #124601 closed Apr 30, 2024

• torch._dynamo.exc.Unsupported: comparison AutogradFunctionVariable() <built-in function is_not> ConstantVariable(NoneType)

  #125140 closed Apr 30, 2024

• ERROR: No matching distribution found for torchvision==0.6.0+cu121

  #124587 closed Apr 30, 2024

• [inductor][cpu]LoweringException: AttributeError: 'NoneType' object has no attribute 'get_origin_node' in 2024-04-20 nightly release

  #124844 closed Apr 30, 2024

• DISABLED test_torchvision_models_alexnet (__main__.TestVisionTracing)

  #123908 closed Apr 30, 2024

• problematic math backend for F.scaled_dot_product_attention in ROCm 6.0 when testing using vllm for generate

  #119389 closed Apr 30, 2024

• DISABLED test_open_device_registration (__main__.TestCppExtensionOpenRgistration)

  #100152 closed Apr 29, 2024

• How can i train my model with torchrun on multiple GPUs but without DDP?

  #125012 closed Apr 29, 2024

• [typing] `nn.Parameter` return type identified as `Tensor` by `pyright`

  #125105 closed Apr 29, 2024

• DISABLED test_max_autotune_remote_caching_dynamic_False (__main__.TestMaxAutotune)

  #121166 closed Apr 29, 2024

• DISABLED test_expand_cuda (__main__.TestUnbackedSymintsCUDA)

  #124074 closed Apr 29, 2024

• DISABLED test_binary_op_list_error_cases__foreach_add_cuda_bool (__main__.TestForeachCUDA)

  #122900 closed Apr 29, 2024

• DISABLED test_max_autotune_remote_caching_dynamic_True (__main__.TestMaxAutotune)

  #121194 closed Apr 29, 2024

• Dataloader codeowner

  #124473 closed Apr 29, 2024

• torch.export.export doesn't capture input argument names from the module's forward(...) function.

  #122842 closed Apr 29, 2024

• CI with >8G CUDA memory

  #18856 closed Apr 29, 2024

• No continuous integration coverage for Python 2 CUDA

  #21467 closed Apr 29, 2024

• Auto-applying labels if one-or-more Labels are applied?

  #117051 closed Apr 29, 2024

• [ignore this] Testing

  #125189 closed Apr 29, 2024

• Warning can not be disabled ?

  #123626 closed Apr 29, 2024

• Bug when indexing 2D tensors using an MPS device

  #125100 closed Apr 29, 2024

• Accessing the `device` attribute of bias terms of `TransformerEncoderLayer` initialized with `bias = False` causes Attribute error

  #125015 closed Apr 29, 2024

• [inductor] Get wrong results when supporting module buffer mutation

  #124583 closed Apr 29, 2024

• [dynamo] "step unsupported" graph break will make dynamo can't completely trace code after break

  #125138 closed Apr 29, 2024

• use bfloat16 on nvidia V100 GPU

  #124996 closed Apr 29, 2024

• doc link failure of `torch.compile`

  #125123 closed Apr 29, 2024

• DISABLED test_triton_kernel_extern_kernel_arg_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #118545 closed Apr 29, 2024

• DISABLED test_triton_kernel_reinterpret_view_mem_leak_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #118652 closed Apr 29, 2024

• DISABLED test_triton_kernel_multi_output_arg_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #118632 closed Apr 29, 2024

• DISABLED test_triton_kernel_reinterpret_view_mem_leak_abi_compatible_cuda (__main__.AOTInductorTestABICompatibleCuda)

  #118640 closed Apr 29, 2024

• DISABLED test_triton_kernel_extern_kernel_arg_abi_compatible_cuda (__main__.AOTInductorTestABICompatibleCuda)

  #118544 closed Apr 29, 2024

• DISABLED test_torchvision_models_regnet_y_8gf (__main__.TestVisionTracing)

  #123977 closed Apr 29, 2024

• DISABLED test_torchvision_models_regnet_y_16gf (__main__.TestVisionTracing)

  #123976 closed Apr 29, 2024

• GuardOnDataDependent error with differentiable output that has data dependent size

  #124766 closed Apr 28, 2024

• Can't fine-tune a MeloTTS model on my dataset in Google Colab

  #125128 closed Apr 28, 2024

• `torch.autocast` produces confusing error message when passing `torch.device`

  #124738 closed Apr 28, 2024

• Running speechbrain demo on aarch64 with pytorch 2.1.2 is much slower than pytorch 1.10.0

  #123143 closed Apr 28, 2024

• [inductor][cpu]basic_gnn_gcn AMP static/dynamic shape default/cpp wrapper single thread performance regression

  #123502 closed Apr 28, 2024

• TOR901 lint is too aggressive

  #125050 closed Apr 27, 2024

• [numpy] Add torch.newdim/torch.newaxis

  #65307 closed Apr 27, 2024

• Compiled model raises error "attn_bias is not correctly aligned" in pytorch 2.2

  #121943 closed Apr 27, 2024

• Unexpected instruction types specified for 'sub' on TIMM seresnext26d_32x4d model

  #118589 closed Apr 27, 2024

• UNSTABLE rocm / linux-focal-rocm6.0-py3.8 / test (default)

  #119908 closed Apr 27, 2024

• [MPS] `torch.nextafter` incorrect handling of negative inputs

  #124985 closed Apr 27, 2024

• [INDUCTOR] [CPU] [GPT-FAST-MOE] large perf regression with coordinate_descent_tuning disabled

  #124697 closed Apr 27, 2024

• DISABLED test_comprehensive_fft_fft_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #122715 closed Apr 26, 2024

• [ONNX] Memory leak

  #86518 closed Apr 26, 2024

• torch dynamo's ORT backend uses "ort" dispatch key instead of "maia"

  #124966 closed Apr 26, 2024

• DISABLED test_all_to_all_single_inductor (__main__.TestFunctionalAutograd)

  #123933 closed Apr 26, 2024

• DISABLED test_load_tensor_cuda (__main__.TestContentStoreCUDA)

  #123849 closed Apr 26, 2024

• DISABLED test_buffer_mutation_3_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #123321 closed Apr 26, 2024

• DISABLED test_equivalent_backed_unbacked_cuda (__main__.TestUnbackedSymintsCUDA)

  #123947 closed Apr 26, 2024

• "torch._dynamo.exc.Unsupported: torch.* op returned non-Tensor bool call_method is_complex" error

  #122692 closed Apr 26, 2024

• [v.2.3.0] Release Tracker

  #121760 closed Apr 26, 2024

• Add a GitHub actions workflow for Macos

  #63466 closed Apr 26, 2024

• a weird bug in torch.compile

  #124817 closed Apr 26, 2024

• h

  #124993 closed Apr 26, 2024

• DISABLED test_torchvision_models_efficientnet_b1 (__main__.TestVisionTracing)

  #123889 closed Apr 26, 2024

• DISABLED test_broadcast_tensors_cuda (__main__.TestUnbackedSymintsCUDA)

  #123862 closed Apr 26, 2024

• DISABLED test_autotuning_cuda (__main__.TestUnbackedSymintsCUDA)

  #123729 closed Apr 26, 2024

• DISABLED test_conv_unary_fusion_nnc (__main__.TestMkldnnFusion)

  #123905 closed Apr 26, 2024

• DISABLED test_output_misaligned_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #123818 closed Apr 26, 2024

• DISABLED test_basic_cuda (__main__.TestContentStoreCUDA)

  #100209 closed Apr 26, 2024

• DISABLED test_torchvision_models_maxvit_t (__main__.TestVisionTracing)

  #123918 closed Apr 26, 2024

• DISABLED test_comprehensive_ones_cuda_int64 (__main__.TestInductorOpInfoCUDA)

  #123837 closed Apr 26, 2024

• DISABLED test_torchvision_models_efficientnet_b3 (__main__.TestVisionTracing)

  #123891 closed Apr 26, 2024

• DISABLED test_torchvision_models_efficientnet_b7 (__main__.TestVisionTracing)

  #123890 closed Apr 26, 2024

• DISABLED test_torchvision_models_densenet169 (__main__.TestVisionTracing)

  #123907 closed Apr 26, 2024

• DISABLED test_batched_mm_bfloat16_bs_10_cuda_bfloat16 (__main__.TestDecompCUDA)

  #123728 closed Apr 26, 2024

• [2.3 dynamic shapes] backend='inductor' raised: LoweringException: AssertionError: indices must be int64, byte or bool. Got [torch.float32]

  #124006 closed Apr 26, 2024

• Support benchmark fusion for TemplateKernel

  #108716 closed Apr 26, 2024

• DISABLED test_comprehensive_amax_cuda_float16 (__main__.TestInductorOpInfoCUDA)

  #124640 closed Apr 26, 2024

• [onnx] support more combinations of args/kwargs as model inputs for pytorch-onnx converter

  #81478 closed Apr 26, 2024

• VecISA.__bool__ is very expensive (nearly a second) on startup

  #100378 closed Apr 25, 2024

• on MPS, torch.embedding, Linear and others raise: RuntimeError: Placeholder storage has not been allocated on MPS device!

  #123995 closed Apr 25, 2024

• Torch nightly wheels no longer include `torchgen` YAML files

  #124941 closed Apr 25, 2024

• Export serializes empty list as list of bools

  #123480 closed Apr 25, 2024

• [Dynamo] Unsupported: missing: DELETE_SUBSCR

  #123317 closed Apr 25, 2024

• XML test-reports get overwritten in case of retry

  #123882 closed Apr 25, 2024

• How to catch NCCL collective timeout in Python

  #124887 closed Apr 25, 2024

• Maybe consider vendoring opentelemetry by hand, rather than submodule

  #124612 closed Apr 25, 2024

• UNSTABLE rocm

  #124951 closed Apr 25, 2024

• Dataloader crashes with FileNotFoundError randomly when num_workers>0 on Ubuntu 22.04

  #124903 closed Apr 25, 2024

• torch.compile : RuntimeError: "foreach_tensor_copy" not implemented for 'Int'

  #124170 closed Apr 25, 2024

• pytorch Windows MKL cmake file don't support static link mkl.

  #124869 closed Apr 25, 2024

• Backwards with cat with data-dependent sizes doesn't work

  #124652 closed Apr 25, 2024

• Support CUDA 12.4

  #104417 closed Apr 25, 2024

• `Enum` used as a key of the input raises guards error

  #111603 closed Apr 25, 2024

• Find a bug from beta-released "scaled_dot_product_attention"

  #124464 closed Apr 25, 2024

• Interpolate nearest

  #121390 closed Apr 25, 2024

• [dynamo] Tracing triton kernel unexpectedly

  #122768 closed Apr 25, 2024

• Substitutions result in unbacked SymInts showing up before their definition sites

  #123854 closed Apr 25, 2024

• SDPA + torch.compile: (*bias): last dimension must be contiguous

  #124289 closed Apr 24, 2024

• Correctly handle `F.interpolate` upsample with amp

  #121072 closed Apr 24, 2024

• [ONNX] Discuss improvements to Diagnostic public API

  #103713 closed Apr 24, 2024

• topk, bmm, max ops - Multiple dispatch failed for 'torch.ops.aten.size'; all __torch_dispatch__ handlers returned NotImplemented:

  #122772 closed Apr 24, 2024

• Not loading optimizer state separately from checkpoint causes errors with FQNs

  #124546 closed Apr 24, 2024

• Switch `libopenblas` and `libopenblas-dev` to `libopenblas64` and `libopenblas64-dev`

  #123534 closed Apr 24, 2024

• Checkpointed function does not preserve `requires_grad` if output is a dataclass.

  #124725 closed Apr 24, 2024

• Release 2.3 manual validations

  #123736 closed Apr 24, 2024

• Validate cheerry-picks for release 2.3

  #123734 closed Apr 24, 2024

• `aminmax` will trigger INTERNAL ASSERT if input is empty on cuda

  #85439 closed Apr 24, 2024

• `test_triton_scaled_dot_product_attention_block_size_16_cuda_bfloat16` is broken on A100

  #124333 closed Apr 24, 2024

• [PREEMPTIVE] Migration for ARC runners - Possible disrruption of jobs or increased queue times

  #124831 closed Apr 24, 2024

• [Profiler] Maybe append a kernel to an unrelated event.

  #124388 closed Apr 24, 2024

• schedular

  #124367 closed Apr 24, 2024

• AOT Inductor / Support dict[string][tensor] as the argument of the C++ AOTIModelContainerRunnerCpu::run functions (same for cuda)

  #121785 closed Apr 23, 2024

• Clean way to distinguish python subclass NT vs. C++ NT

  #110543 closed Apr 23, 2024

• Verbose log: [__aot_joint_graph] could not reconstruct view by re-applying a ViewMeta sequence.

  #124499 closed Apr 23, 2024

• In `_scaled_mm`, `scale_result` not changing result tensor at all

  #119135 closed Apr 23, 2024

138 Issues opened by 87 people

• DISABLED test_bmm_multithreaded (__main__.TestTorch)

  #125240 opened Apr 30, 2024

• Improved strategy for dealing with deterministically flaky tests which are order sensitive

  #125239 opened Apr 30, 2024

• torch.no_grad() is not working for dynamo inductor backend

  #125236 opened Apr 30, 2024

• [Inductor] [Distributed] DDP torch.compile model hangs on exit (python 3.8/3.9)

  #125235 opened Apr 30, 2024

• torch.Library can easily cause segfault on loading/unloading

  #125234 opened Apr 30, 2024

• ROCm: `fatal error: aotriton/flash.h: No such file or directory` when building with `USE_ROCM=1`

  #125230 opened Apr 30, 2024

• DISABLED test_variable_traverse (__main__.TestAutogradWithCompiledAutograd)

  #125229 opened Apr 30, 2024

• DISABLED test_issue106555 (__main__.TestCompiledAutograd)

  #125228 opened Apr 30, 2024

• Strange behavior of randint using device=cuda

  #125224 opened Apr 30, 2024

• torch.uniform_() is single-threaded on CPU

  #125223 opened Apr 30, 2024

• DISABLED test_var_mean_differentiable (__main__.TestAutogradWithCompiledAutograd)

  #125220 opened Apr 30, 2024

• DISABLED test_inplace_grad_update (__main__.TestCompiledAutograd)

  #125219 opened Apr 30, 2024

• DISABLED test_perfect_match_on_sequence_and_bool_attributes (__main__.TestFxToOnnx)

  #125218 opened Apr 30, 2024

• MaxPool2D memory leakage on device MPS

  #125217 opened Apr 30, 2024

• torch.inference_mode documentation not availble

  #125216 opened Apr 30, 2024

• [NT] Implementing Multi-Head Attention with NestedTensors

  #125214 opened Apr 30, 2024

• DISABLED test_type_conversions (__main__.TestAutogradWithCompiledAutograd)

  #125206 opened Apr 30, 2024

• DISABLED test_unused_output (__main__.TestAutogradWithCompiledAutograd)

  #125205 opened Apr 30, 2024

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

  #125198 opened Apr 29, 2024

• Pytorch running on macosx-14.xlarge reports MPS is available when it is not.

  #125197 opened Apr 29, 2024

• DISABLED test_too_many_grads (__main__.TestAutogradWithCompiledAutograd)

  #125195 opened Apr 29, 2024

• multithreaded autograd backward doesn't respect autocast dtype context manager

  #125186 opened Apr 29, 2024

• torch/_refs/__init__.py is not autoreloadable in bento / jupyter notebook

  #125183 opened Apr 29, 2024

• QR Decomposition for Sparse Matrix

  #125182 opened Apr 29, 2024

• Mitigate pypi issue with space (short term)

  #125179 opened Apr 29, 2024

• Can't load on rank 0 only with `set_optimizer_state_dict`

  #125177 opened Apr 29, 2024

• [CUDA][AMP] Size-1 (scalar) norms are broken on CUDA + AMP following #122143

  #125174 opened Apr 29, 2024

• Flight Recorder Sequence IDs are insufficient

  #125173 opened Apr 29, 2024

• datettime.now() is not supported by Dynamo

  #125171 opened Apr 29, 2024

• Calling `get_model_state_dict/set_model_state_dict` requires forward pass for `_lazy_init`

  #125170 opened Apr 29, 2024

• `scale` parsed as `float` in ONNX `scaled_dot_product_attention` implementation

  #125158 opened Apr 29, 2024

• DISABLED test_comprehensive_special_bessel_y1_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #125151 opened Apr 29, 2024

• DISABLED test_slice_expanded_v (__main__.TestAutogradWithCompiledAutograd)

  #125149 opened Apr 29, 2024

• Race condition in FileTimerServerTest.test_expired_timers

  #125146 opened Apr 29, 2024

• The "step unsupported" graph break will make dynamo can't completely trace code after break

  #125141 opened Apr 29, 2024

• Tensor.abs() gives incorrect results on Complex64 when using MPS

  #125135 opened Apr 29, 2024

• binaries/dump_operator_names.cc missing iostream include

  #125134 opened Apr 29, 2024

• DISABLED test_sharded_grad (__main__.TestAutogradWithCompiledAutograd)

  #125130 opened Apr 29, 2024

• FP6 dtype!

  #125129 opened Apr 28, 2024

• torch.is_signed on new uint dtypes raises Unknown ScalarType

  #125124 opened Apr 28, 2024

• Does torch.nn.Linear check the weights shape before assignment?

  #125116 opened Apr 28, 2024

• 2.3.0 on Windows missing dependency?

  #125109 opened Apr 27, 2024

• Further explanation for `batch_isend_irecv`

  #125099 opened Apr 27, 2024

• RuntimeError: MPS: Unsupported Border padding mode

  #125098 opened Apr 27, 2024

• PyTorch Distributed Load Updates or Returns `state_dict`

  #125096 opened Apr 27, 2024

• Broken Docker Image on dockerhub

  #125094 opened Apr 27, 2024

• Triton installation not found.

  #125093 opened Apr 27, 2024

• [Distributed] P2P Operations on NCCL do not respect tag

  #125079 opened Apr 26, 2024

• `torch.compile` fails with `jacfwd` when multiplying/dividing float and tensor

  #125078 opened Apr 26, 2024

• [Inductor] Generate triton block pointers for discontiguous strided tensors

  #125077 opened Apr 26, 2024

• Inductor can not fuse cat with a pointwise

  #125075 opened Apr 26, 2024

• DISABLED test_setting_default_saved_variable_hooks_twice_should_use_inner (__main__.TestAutogradWithCompiledAutograd)

  #125074 opened Apr 26, 2024

• 2.2.0+ regresses SDPA performance on Windows

  #125070 opened Apr 26, 2024

• ValueError: weight_norm of 'weight' not found in ParametrizedConvTranspose1d

  #125064 opened Apr 26, 2024

• 404 on torch.inference_mode doc page

  #125059 opened Apr 26, 2024

• DISABLED test_sdpa_backwards_cuda_bfloat16 (__main__.TestNestedTensorSubclassCUDA)

  #125058 opened Apr 26, 2024

• DISABLED test_foreach_matches_forloop_RAdam_cpu_float64 (__main__.TestOptimRenewedCPU)

  #125057 opened Apr 26, 2024

• DISABLED test_select_expanded_v (__main__.TestAutogradWithCompiledAutograd)

  #125056 opened Apr 26, 2024

• DISABLED test_dynamic_shapes (__main__.TestCompiledAutograd)

  #125055 opened Apr 26, 2024

• MPS backend thinks that small floats are less than zero

  #125051 opened Apr 26, 2024

• Support auto_functionalized for None returns

  #125044 opened Apr 26, 2024

• [Feature Request] Support `dtype` arg in `torch._foreach_norm`

  #125040 opened Apr 26, 2024

• DISABLED test_binary_op_list_error_cases__foreach_clamp_max_cuda_float64 (__main__.TestForeachCUDA)

  #125035 opened Apr 26, 2024

• Tensorboard's SummaryWriter.add_graph() doesn't work with packedSequence

  #125033 opened Apr 26, 2024

• Allow operators with no returns to be "out=" operations

  #125030 opened Apr 26, 2024

• OpInfo testing is inadequate for `nextafter`

  #125028 opened Apr 26, 2024

• CUDA memory summary in FSDP all_gather causes excessive noise.

  #125025 opened Apr 26, 2024

• DISABLED test_custom_fn_output_metadata (__main__.TestCompiledAutograd)

  #125024 opened Apr 26, 2024

• DISABLED test_binary_op_float_inf_nan__foreach_clamp_min_cuda_float32 (__main__.TestForeachCUDA)

  #125022 opened Apr 26, 2024

• DISABLED test_saved_variable_packing_unpacking_saved_original_with_default_hooks (__main__.TestAutogradWithCompiledAutograd)

  #125023 opened Apr 26, 2024

• DISABLED test_allocation_id_uniqueness (__main__.TestTorchTidyProfiler)

  #125021 opened Apr 26, 2024

• DISABLED test_unbacked_cat_backwards_cuda (__main__.TestInductorDynamicCUDA)

  #125019 opened Apr 26, 2024

• DISABLED test_save_with_without_initializer_dont_include_initializer_no_fake_mode_no_exported_program (__main__.TestFxToOnnx)

  #125020 opened Apr 26, 2024

• libtorch C++ windows: The specified module could not be found. mkl_vml_def.1.dll

  #125013 opened Apr 26, 2024

• Tensor can not be accessed!

  #125010 opened Apr 26, 2024

• Support exporting the compute graph JiT compiled by inductor?

  #125007 opened Apr 26, 2024

• RuntimeError: Unsupported qscheme: per_channel_affine. When using quantise_fx(

  #125004 opened Apr 26, 2024

• bool as scalar in jit ir API not working as expected.

  #125003 opened Apr 26, 2024

• When I set the operator weights all to 1, the operator weights change as soon as an input is made to the operator

  #124998 opened Apr 26, 2024

• [DTensor] Sharding strategy not implemented error should be thrown earlier

  #124990 opened Apr 26, 2024

• [DTensor] Keep track of data-dependent ops and skip _propagate_tensor_meta to go thru fake tensor

  #124989 opened Apr 26, 2024

• [Releng] Triton version for minor releases starting with 2.4

  #124974 opened Apr 25, 2024

• Export llama v3 to ONNX

  #124973 opened Apr 25, 2024

• Give the possibly to get back normal `Tensor`s as `MaskedTensor` gradients

  #124964 opened Apr 25, 2024

• `ascii` error during recompilation after cache limit is reached (likely due to `einsum`)

  #124960 opened Apr 25, 2024

• torch.onnx.export generates an incorrect bias shape for certain conv transpose models

  #124956 opened Apr 25, 2024

• PRs with unbalanced quotes can not be merged

  #124953 opened Apr 25, 2024

• [distributed] First NCCL barrier does not respect timeout

  #124950 opened Apr 25, 2024

• torch.compile fails on hugging face Mistral7b

  #124946 opened Apr 25, 2024

• Support torch.distributed.checkpoint.state_dict.get_model_state_dict/set_model_state_dict when torch dist is not initialized

  #124942 opened Apr 25, 2024

• Conflict between bias=False and why_not_sparsity_fast_path in Transformer Module

  #124937 opened Apr 25, 2024

• Tensor's storage changes computation outcome for CPU tensors.

  #124934 opened Apr 25, 2024

• [RFC] Support reinplaceble ops for custom ops in Inductor

  #124933 opened Apr 25, 2024

• Missing description of Transformer argument "memory_mask" shape in 3D (including the batch dimension) case

  #124931 opened Apr 25, 2024

• Time taken to data loading increased in newer builds (ARM)

  #124922 opened Apr 25, 2024

• Custom Operator Design for torch.compile: Must Output Tensors Always Be Returned?

  #124918 opened Apr 25, 2024

• Exporting torch slice_scatter to onnx Identity

  #124915 opened Apr 25, 2024

• [inductor][cpu]shufflenet_v2_x1_0 QAT performance regression in 2024-04-20 nightly release

  #124913 opened Apr 25, 2024

• DataLoader's pin_memory is default to CUDA if parameter pin_memory_device is not set

  #124908 opened Apr 25, 2024

• 2.3.0 not backward compatible with torchdata

  #124907 opened Apr 25, 2024

• Many cases in distributed/elastic/multiprocessing/redirects_test.py fails when use pytest

  #124906 opened Apr 25, 2024

• Add support for IPC features for PrivateUse devices

  #124902 opened Apr 25, 2024

• `RuntimeError: invalid dtype for bias` when use compile + autocast

  #124901 opened Apr 25, 2024

• [dynamo] Crash when context manager object crosses a graph break

  #124900 opened Apr 25, 2024

• Certain .pyi files are not encoded as UTF-8 in Windows

  #124897 opened Apr 25, 2024

• Windows build step did not fail despite error

  #124886 opened Apr 24, 2024

• ONNX dynamic sized model export with torch.onnx.dynamo_export fails when torch.nn.functional.interpolate is used

  #124884 opened Apr 24, 2024

• Nested wrapper subclasses with torch.compile is broken

  #124878 opened Apr 24, 2024

• SDPA memory efficient kernel returns NaNs when the query and key are different lengths

  #124877 opened Apr 24, 2024

• tensor.dtype.to_complex() crashes kernel after ~100 calls in ipython kernel

  #124868 opened Apr 24, 2024

• support side effects in HOPs?

  #124866 opened Apr 24, 2024

• Deprecate unsupported types in operator registration

  #124863 opened Apr 24, 2024

• torch._dynamo.assume_constant_result does not work outside nn.Module

  #124858 opened Apr 24, 2024

• Use lintrunner-adapters for our adapters when possible

  #124857 opened Apr 24, 2024

• ShapeEnv canonicalization is still over-aggressive

  #124855 opened Apr 24, 2024

• [inductor] unexpected cuda:0 device usage when compiling and runing a model on cuda:1

  #124854 opened Apr 24, 2024

• aten::nonzero calls taking a huge amount of time when using MPS backend vs CPU

  #124850 opened Apr 24, 2024

• Forwardpropagation of high-dimensional tensors through the nn.Linear module becomes multiple times slower since pytorch 2.1.0

  #124838 opened Apr 24, 2024

• [inductor][cpu]RuntimeError: no channels last format strides exist in 1 dimensions in 2024-04-21 nightly release

  #124837 opened Apr 24, 2024

• nn.functional.ELU output differs on MPS vs CPU if input is noncontiguous

  #124834 opened Apr 24, 2024

• MPS RNG state fails to progress immediately after fork_rng

  #124833 opened Apr 24, 2024

• Add privateuse1 check on capturable optimizer

  #124830 opened Apr 24, 2024

• Compiled autograd doesn't support reading attribute from autograd context if the attribute is created in eager forward

  #124827 opened Apr 24, 2024

• CI is downloading model definitions and model weights from external sources

  #124825 opened Apr 24, 2024

• torch.compile error: Unsupported reduction type from torch.float32 to torch.int64

  #124821 opened Apr 24, 2024

• [RFC] Mix and Match CUDA Allocators using Private Pools

  #124807 opened Apr 24, 2024

• Dynamo Export Support for Qwen/Qwen-7B-Chat: Mutating module attribute _ntk_alpha_cached_list during export

  #124796 opened Apr 23, 2024

• Dynamo Export Support for Google/Gemma-2B: Mutating module attribute inv_freq during export

  #124793 opened Apr 23, 2024

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

  #124790 opened Apr 23, 2024

• torch.nn.checkpoint.checkpoint ignores default device in backward() call

  #124788 opened Apr 23, 2024

• Scatter_add limitation when accumulate beyond 2^24 under float32 precision

  #124783 opened Apr 23, 2024

• Desne-sparse broadcasted multiplication fails in the backward pass

  #124778 opened Apr 23, 2024

• Release artifacts for rc releases

  #124759 opened Apr 23, 2024

• DISABLED test_comprehensive_randn_like_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #124758 opened Apr 23, 2024

• DISABLED test_saved_variable_packing_unpacking_did_not_save_original_with_hooks (__main__.TestAutogradWithCompiledAutograd)

  #124757 opened Apr 23, 2024

• Disable PYTORCH_TEST_WITH_DYNAMO=1 tests over the AOTAutograd tests

  #124750 opened Apr 23, 2024

• Improve Dynamo (and other) flaky tests

  #124749 opened Apr 23, 2024

• TestAOTAutograd.test_mem_leak_from_save_for_bw fails locally but not on CI when run with dynamo

  #124747 opened Apr 23, 2024

367 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.

• [dynamo] Automatically convert loop bodies to function calls

  #113538 commented on Apr 26, 2024 • 30 new comments

• [Meta Tensor] fix meta inplace set storage

  #123880 commented on Apr 29, 2024 • 25 new comments

• [vision hash update] update the pinned vision hash

  #123227 commented on Apr 30, 2024 • 24 new comments

• [NT] Make NestedTensor register as having symbolic sizes/strides

  #124687 commented on Apr 27, 2024 • 22 new comments

• [executorch hash update] update the pinned executorch hash

  #123043 commented on Apr 30, 2024 • 21 new comments

• [traced-graph][sparse] propagate sparsity metadata into traced graph

  #117907 commented on Apr 30, 2024 • 20 new comments

• Setup initial testing harness and cache key generation for AOTAutograd Cache

  #124642 commented on Apr 30, 2024 • 19 new comments

• UserWarning: Plan failed with a cudnnException

  #121834 commented on Apr 30, 2024 • 19 new comments

• ARC dynamic rollout

  #124721 commented on Apr 30, 2024 • 18 new comments

• TorchInductor CPU Performance Dashboard

  #93531 commented on Apr 30, 2024 • 16 new comments

• Add CUDA 12.4 workflows

  #121684 commented on Apr 30, 2024 • 16 new comments

• Refresh OpOverloadPacket if a new OpOverload gets added

  #124654 commented on Apr 26, 2024 • 16 new comments

• Add helper function to decompose and inline nn module

  #123683 commented on Apr 29, 2024 • 15 new comments

• fix Invalid call to aoti_torch_tensor_copy_ #123039

  #124037 commented on Apr 30, 2024 • 14 new comments

• Add a cache mechanism to accelerate torch.compile-for-eager

  #116368 commented on Apr 29, 2024 • 12 new comments

• optim.apply_optimizer_in_backward does not account for gradient accumulation

  #124523 commented on Apr 30, 2024 • 11 new comments

• [RFC] Per-Parameter-Sharding FSDP

  #114299 commented on Apr 29, 2024 • 11 new comments

• General MPS op coverage tracking issue

  #77764 commented on Apr 30, 2024 • 10 new comments

• Fakify script object inputs and attributes for non-strict export

  #124239 commented on Apr 30, 2024 • 10 new comments

• [dynamo] Refactor into torch/_inductor/runtime/compile_tasks.py

  #124681 commented on Apr 30, 2024 • 9 new comments

• [inductor] Remove usage of device_interface from _inductor.runtime

  #124592 commented on Apr 30, 2024 • 8 new comments

• [custom_op] use new python custom ops API on prims ops

  #124665 commented on Apr 30, 2024 • 8 new comments

• Verify types in custom op schemas

  #124520 commented on Apr 26, 2024 • 8 new comments

• [DeviceMesh] Fix hash and eq not match

  #123572 commented on Apr 27, 2024 • 7 new comments

• ARC runners timeout during Docker updates

  #124727 commented on Apr 29, 2024 • 7 new comments

• set_default_device/torch.device has performance impact for non-factory functions

  #92701 commented on Apr 24, 2024 • 7 new comments

• [DCP] Introduce async staging extension points

  #122965 commented on Apr 25, 2024 • 7 new comments

• [ROCm] TunableOp improvements

  #124362 commented on Apr 30, 2024 • 7 new comments

• Fixes two build problems on ROCM 6.1 + Ubuntu 22.04

  #118216 commented on Apr 23, 2024 • 6 new comments

• inductor: Add Conv3d support

  #124361 commented on Apr 29, 2024 • 6 new comments

• Allow tensor subclasses and add `torch.serialization.mark_safe_globals` that allows users to allowlist classes for `weights_only` load

  #124331 commented on Apr 30, 2024 • 6 new comments

• Fix typo under torch/_inductor directory

  #119658 commented on Apr 30, 2024 • 6 new comments

• [RFC] Autoload Device Extension

  #122468 commented on Apr 26, 2024 • 6 new comments

• [Memory Snapshot] Add recordAnnotations to capture record_function annotations

  #124179 commented on Apr 26, 2024 • 6 new comments

• Add a variable for some testcases.

  #124708 commented on Apr 29, 2024 • 6 new comments

• [DCP] Adds strict option to DefaultPlanner

  #123869 commented on Apr 26, 2024 • 6 new comments

• [c10d] abort a communicator at most once

  #124436 commented on Apr 29, 2024 • 5 new comments

• [guards][cpp-guards] Optimize NN module getattr guards

  #124522 commented on Apr 30, 2024 • 5 new comments

• Better core binding in torch.backends.xeon.run_cpu when launced from torchrun with --nproc-per-node

  #123711 commented on Apr 30, 2024 • 4 new comments

• Investigate torch.compile Windows support.

  #122094 commented on Apr 30, 2024 • 4 new comments

• Add LR as tensor tests

  #123750 commented on Apr 26, 2024 • 4 new comments

• Initial implementation of Inductor FX Graph Remote Cache

  #124669 commented on Apr 23, 2024 • 4 new comments

• Pytorch Conda nightly build failures due to timeout

  #124667 commented on Apr 29, 2024 • 4 new comments

• torch.Tensor.remainder raises a floating point exception when divisor is -1

  #124644 commented on Apr 29, 2024 • 4 new comments

• RuntimeError: derivative for aten::_scaled_dot_product_flash_attention_backward is not implemented

  #116350 commented on Apr 29, 2024 • 4 new comments

• Enable dynamo-traced optimizer peak memory tests

  #124543 commented on Apr 25, 2024 • 4 new comments

• [dynamo] Support numpy.dtype

  #124481 commented on Apr 24, 2024 • 4 new comments

• Fix the specific setting in documentation to match with elsewhere

  #124463 commented on Apr 27, 2024 • 4 new comments

• Fix for addcdiv contiguous problem

  #124442 commented on Apr 25, 2024 • 4 new comments

• [Quant][PT2E] enable qlinear post op fusion for dynamic quant & qat

  #122667 commented on Apr 30, 2024 • 3 new comments

• torch.compiled model output gets overwritten despite tensor.detach()

  #104435 commented on Apr 24, 2024 • 3 new comments

• Remove caffe2/

  #122527 commented on Apr 30, 2024 • 3 new comments

• [PT2D] Make the speedup benchmark works with DDP + CompiledAutograd

  #121315 commented on Apr 24, 2024 • 3 new comments

• [ONNX] STFT ExportProgram error

  #113504 commented on Apr 26, 2024 • 3 new comments

• DISABLED test_split_with_sizes_aot_autograd_cleans_up_traceback_meta (__main__.AotAutogradFallbackTests)

  #122767 commented on Apr 26, 2024 • 3 new comments

• DISABLED test_split_with_sizes_aot_autograd_cleans_up_traceback_meta_dynamic_shapes (__main__.DynamicShapesAotAutogradFallbackTests)

  #122766 commented on Apr 26, 2024 • 3 new comments

• Implement Copy-on-write (COW) tensors

  #109833 commented on Apr 26, 2024 • 3 new comments

• aten::_linalg_solve_ex.result' is not currently implemented for the MPS

  #98222 commented on Apr 27, 2024 • 3 new comments

• ARM libtorch links to libomp but libomp is no longer bundled

  #124732 commented on Apr 29, 2024 • 3 new comments

• DISABLED test_comprehensive_special_bessel_y1_cuda_int64 (__main__.TestInductorOpInfoCUDA)

  #123919 commented on Apr 30, 2024 • 3 new comments

• DISABLED test_free_activation_memory (__main__.TestCompiledAutograd)

  #123949 commented on Apr 30, 2024 • 3 new comments

• s390x: remove workaround for sleef issue

  #124730 commented on Apr 25, 2024 • 3 new comments

• Fast standalone symbolize for unwinding

  #123966 commented on Apr 25, 2024 • 3 new comments

• [ROCm][CI] upgrade CI to ROCm 6.1

  #124300 commented on Apr 29, 2024 • 3 new comments

• [draft] cuda 124 arm wheel test

  #124112 commented on Apr 26, 2024 • 3 new comments

• [DO NOT MERGE] Test new ROCm CI nodes

  #124424 commented on Apr 30, 2024 • 3 new comments

• [dynamo] Unexpected SymBool appearing in "is_causal" inside scaled_dot_product_attention()

  #124707 commented on Apr 24, 2024 • 3 new comments

• [torch.compile][FlopCounter] AssertionError: Global is not OptimizedModule._orig_mod

  #124196 commented on Apr 23, 2024 • 3 new comments

• [xla hash update] update the pinned xla hash

  #124599 commented on Apr 29, 2024 • 3 new comments

• DISABLED test_mark_non_differentiable (__main__.TestAutogradWithCompiledAutograd)

  #124470 commented on Apr 30, 2024 • 2 new comments

• Add aten._unsafe_masked_index

  #116491 commented on Apr 29, 2024 • 2 new comments

• DISABLED test_buffer_mutation_3_abi_compatible_cuda (__main__.AOTInductorTestABICompatibleCuda)

  #123251 commented on Apr 28, 2024 • 2 new comments

• torch.load with weights_only=True to support pickle protocol 3/4/5

  #118166 commented on Apr 25, 2024 • 2 new comments

• Revisit security implications of #31875

  #111806 commented on Apr 27, 2024 • 2 new comments

• [Environment Variable][1/N] Use thread-safe env variable API in c10

  #119449 commented on Apr 24, 2024 • 2 new comments

• sdp::SDPBackend::flash_attention support PrivateUse1

  #124368 commented on Apr 30, 2024 • 2 new comments

• DISABLED test_isolated_node (__main__.TestAutogradWithCompiledAutograd)

  #124460 commented on Apr 26, 2024 • 2 new comments

• [inductor] switch assume_aligned_inputs to False

  #124336 commented on Apr 29, 2024 • 2 new comments

• Torch compile does not work on python 3.12

  #120233 commented on Apr 26, 2024 • 2 new comments

• Investigate Strictness of torch.compile `is_big_gpu`

  #109489 commented on Apr 26, 2024 • 2 new comments

• Should be able to query a schema for HOPs

  #119592 commented on Apr 26, 2024 • 2 new comments

• CUBLAS_STATUS_EXECUTION_FAILED when calling cublasGemmEx

  #124262 commented on Apr 26, 2024 • 2 new comments

• Optionally use hipblaslt

  #120551 commented on Apr 26, 2024 • 2 new comments

• [Profiler] NCCL collectives text garbled and times reported in ns

  #124102 commented on Apr 24, 2024 • 2 new comments

• Ensure only builtins functions are wrapped in new frame for torch.compile

  #124720 commented on Apr 30, 2024 • 2 new comments

• Avoid using thrust:: directly, use THRUST_NS_QUALIFIER:: instead

  #72582 commented on Apr 30, 2024 • 2 new comments

• RX 6800 GPU reset when using ROCm in Stable Diffusion with Torch backend (not sure if relevant)

  #120775 commented on Apr 30, 2024 • 2 new comments

• GroupNorm & InstanceNorm does not handle channels_last correctly

  #111824 commented on Apr 29, 2024 • 2 new comments

• Update CUDA out of memory mesage with private pool info

  #124673 commented on Apr 24, 2024 • 2 new comments

• Fails to compile with nvidia-cuda-toolkit-12.4.0

  #122169 commented on Apr 30, 2024 • 2 new comments

• DISABLED test_multi_backward (__main__.TestAutogradWithCompiledAutograd)

  #124491 commented on Apr 30, 2024 • 2 new comments

• Unbacked SymInts: Should backwards graph with unbacked SymInts be recompiled with hints

  #124686 commented on Apr 23, 2024 • 2 new comments

• ROCm & Windows Support

  #106608 commented on Apr 29, 2024 • 2 new comments

• torch.compile does not work since 2.2.1 on MacOS for some models

  #124497 commented on Apr 23, 2024 • 2 new comments

• TypeError: unhashable type: non-singleton SymInt in AOTAutograd merge_view_inputs

  #114366 commented on Apr 29, 2024 • 2 new comments

• Correct error message for aten::_local_scalar_dense on meta tensor

  #124554 commented on Apr 24, 2024 • 2 new comments

• benchmark.Compare raises: TypeError: object of type 'NoneType' has no len()

  #63971 commented on Apr 29, 2024 • 2 new comments

• [RFC] PyTorch DistributedTensor

  #88838 commented on Apr 29, 2024 • 2 new comments

• [triton hash update] update the pinned triton hash

  #115529 commented on Apr 29, 2024 • 2 new comments

• Fix common_utils's retry decorator, add run_tests call to test_hub

  #116067 commented on Apr 30, 2024 • 2 new comments

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

  #13246 commented on Apr 28, 2024 • 2 new comments

• [2/N] Non-Tensor: Scalar Support: Add scalar to the cache for eager-through-torch.compile

  #124070 commented on Apr 28, 2024 • 2 new comments

• input.is_sparse() INTERNAL ASSERT FAILED

  #120989 commented on Apr 24, 2024 • 2 new comments

• No factory functions for strided quantized tensors

  #74540 commented on Apr 25, 2024 • 2 new comments

• scatter_reduce method do not support complex number multiplication on CUDA

  #121965 commented on Apr 24, 2024 • 2 new comments

• Placing LSTM model on bfloat16 on GPU causes error

  #88136 commented on Apr 25, 2024 • 2 new comments

• Add 2nd shard to ROCm trunk workflow for core distributed UTs

  #121716 commented on Apr 25, 2024 • 2 new comments

• `torch.distributed` hangs when using `torch.distributed.barrier` before any other communication primitives.

  #124714 commented on Apr 25, 2024 • 2 new comments

• [Inductor][Quant] Change the QConv output scale name

  #124246 commented on Apr 30, 2024 • 2 new comments

• upstream `apex.normalization.FusedRMSNorm`

  #72643 commented on Apr 25, 2024 • 2 new comments

• Grad strides do not match bucket view strides.

  #47163 commented on Apr 25, 2024 • 2 new comments

• Improving format of communication metadata in PyTorch Execution Trace.

  #124674 commented on Apr 25, 2024 • 2 new comments

• Grad strides do not match bucket view strides

  #83909 commented on Apr 25, 2024 • 2 new comments

• Batching rule for `aten::_scaled_dot_product_efficient_attention`

  #102457 commented on Apr 25, 2024 • 2 new comments

• [Performance] Potential Performance optimization for SDPA

  #100270 commented on Apr 26, 2024 • 2 new comments

• Fixed an undefined combination of inputs for torch.fmod.

  #120624 commented on Apr 27, 2024 • 2 new comments

• make tensor data const correct

  #97856 commented on Apr 26, 2024 • 2 new comments

• profiler.export_stacks doesn't return stack trace unless experimental_config is provided

  #100253 commented on Apr 30, 2024 • 1 new comment

• Clang tidy torch csrc16

  #120573 commented on Apr 25, 2024 • 1 new comment

• Skip fx passes for split-cat with Node dims

  #124629 commented on Apr 30, 2024 • 1 new comment

• Tensor.nonzero fails on GPU for tensors containing more than INT_MAX elements

  #51871 commented on Apr 30, 2024 • 1 new comment

• Massive initial memory overhead GPU

  #12873 commented on Apr 30, 2024 • 1 new comment

• Request for deterministic support for reflection_pad2d_backward_cuda

  #98925 commented on Apr 30, 2024 • 1 new comment

• Prefer construction via DLPack to costly element-by-element copy

  #120615 commented on Apr 28, 2024 • 1 new comment

• Upgrade submodule oneDNN to v3.4

  #122472 commented on Apr 30, 2024 • 1 new comment

• RuntimeError: reflection_pad2d_backward_cuda does not have a deterministic implementation

  #123843 commented on Apr 30, 2024 • 1 new comment

• Fix a check message in pickler

  #120701 commented on Apr 27, 2024 • 1 new comment

• While loop autograd

  #124573 commented on Apr 26, 2024 • 1 new comment

• functionalize storage resizing, minimal ppFSDP traceable forward

  #122434 commented on Apr 25, 2024 • 1 new comment

• [dtensor] from_local broadcast use functional collective

  #120457 commented on Apr 24, 2024 • 1 new comment

• Remove dtype check on meta device

  #120634 commented on Apr 26, 2024 • 1 new comment

• Connection closed by peer when using dist.isend in gloo backend

  #75512 commented on Apr 30, 2024 • 1 new comment

• [feature request] Rank-Revealing QR - Adding dgeqp3 support to torch.qr

  #10454 commented on Apr 30, 2024 • 1 new comment

• Some fixups

  #124658 commented on Apr 24, 2024 • 1 new comment

• torch native functions cannot be used with inspect.signature

  #28233 commented on Apr 30, 2024 • 1 new comment

• Make CI less noisy

  #124664 commented on Apr 29, 2024 • 1 new comment

• Cannot re-initialize CUDA in forked subprocess

  #40403 commented on Apr 30, 2024 • 1 new comment

• [Quant][Inductor] Enable lowering of qlinear-binary(-unary) fusion for X86Inductor

  #122593 commented on Apr 29, 2024 • 1 new comment

• Fix DDP no_sync when find_unused_parameters is True

  #124193 commented on Apr 26, 2024 • 1 new comment

• DISABLED test_index (__main__.TestPythonBuiltinOP)

  #119160 commented on Apr 30, 2024 • 1 new comment

• torch.utils.cpp_extension.load recompiling every time

  #124454 commented on Apr 30, 2024 • 1 new comment

• Fix absolute links in pytorch repository and allow it to be proxied

  #101798 commented on Apr 30, 2024 • 1 new comment

• [dynamo] Function => FunctionCtx for placeholder obj

  #120577 commented on Apr 29, 2024 • 1 new comment

• [1/N] Non-Tensor: Scalar Support: Enable aot compile to support aten operations with scalar input like alpha

  #124177 commented on Apr 28, 2024 • 1 new comment

• [tensor] Replace raw loops with std::reduce for size calc.

  #120580 commented on Apr 25, 2024 • 1 new comment

• Understand the oneDNN graph fusion with torch script

  #124458 commented on Apr 30, 2024 • 1 new comment

• [inductor] add cpp builder code.

  #124045 commented on Apr 30, 2024 • 1 new comment

• Prevent cuda:0 context initialization when working on another cuda device

  #124722 commented on Apr 24, 2024 • 1 new comment

• Set simdlen based on ATEN_CPU_CAPABILITY

  #123514 commented on Apr 30, 2024 • 1 new comment

• [WIP] support map impl in pre-dispatch IR

  #120159 commented on Apr 30, 2024 • 1 new comment

• [executorch] Add support for method variant functions in ExecuTorch codegen

  #120840 commented on Apr 30, 2024 • 1 new comment

• Fix dynamo issue "Failed running call_function <built-in method sparse_coo_tensor of type object at 0xDEADBEEF"

  #118192 commented on Apr 28, 2024 • 1 new comment

• ProcessGroupWrapper support custom backend

  #124447 commented on Apr 28, 2024 • 1 new comment

• [dynamo] fix compiling Dataclass construction with default_factory

  #120827 commented on Apr 29, 2024 • 1 new comment

• [MPS] Add support for max_unpool2d

  #118665 commented on Apr 26, 2024 • 1 new comment

• [ROCm] hipSPARSELt Integration

  #124320 commented on Apr 30, 2024 • 1 new comment

• [ROCm] amdsmi library integration

  #119182 commented on Apr 24, 2024 • 1 new comment

• [typing] Rename argument of `nn.Sequential.forward` from `input` to `__input`

  #119209 commented on Apr 28, 2024 • 1 new comment

• Fix missing parameter check in at::batch_norm

  #119361 commented on Apr 29, 2024 • 1 new comment

• Fix stream type to generic in comms default hooks

  #120069 commented on Apr 25, 2024 • 1 new comment

• [Distributed] Add P2P versions of *object_list operations

  #124379 commented on Apr 26, 2024 • 1 new comment

• [Inductor][AMD] Enable pipeliner for Gemm

  #120637 commented on Apr 26, 2024 • 1 new comment

• [Don't merge] Refactor device bound check for xpu code

  #120768 commented on Apr 29, 2024 • 1 new comment

• [fbcode] Upstream parallel fast cat on cpu in OSS cat op

  #120753 commented on Apr 28, 2024 • 1 new comment

• Update expecttest in conda env

  #120711 commented on Apr 27, 2024 • 1 new comment

• DISABLED test_mm_batching (__main__.TestScript)

  #119747 commented on Apr 30, 2024 • 1 new comment

• Add default values to PyTorchMemEffAttention::AttentionKernel::Params members

  #112215 commented on Apr 30, 2024 • 1 new comment

• [sym_shapes][perf] Optimize bound_sympy avoiding sympy equals

  #124211 commented on Apr 23, 2024 • 1 new comment

• Fix `as_strided` functionalization for lazy backend.

  #120435 commented on Apr 28, 2024 • 1 new comment

• [dynamo] Handle np.iinfo/finfo/dtype as input

  #124482 commented on Apr 24, 2024 • 1 new comment

• Add dist hooks support for custom device

  #114730 commented on Apr 27, 2024 • 1 new comment

• Add back non standard shapes test samples for SDPA in common_methods_…

  #115464 commented on Apr 26, 2024 • 1 new comment

• [FAILURE] quantized test

  #120941 commented on Apr 30, 2024 • 1 new comment

• Enable test_embedding_bag_device_* with PYTORCH_TEST_WITH_DYNAMO

  #120884 commented on Apr 29, 2024 • 1 new comment

• Add _to_copy op for jagged NT

  #115749 commented on Apr 29, 2024 • 1 new comment

• [dynamo] fix silent incorrectness caused by variable tracker caching

  #120861 commented on Apr 29, 2024 • 1 new comment

• Some update

  #124450 commented on Apr 29, 2024 • 1 new comment

• [not4land] Batch norm consolidation disable xfails/skips

  #120844 commented on Apr 29, 2024 • 1 new comment

• Add `torch._dynamo.is_fullgraph_compiling` to allow different codepath depending on fullgraph tracing

  #120400 commented on Apr 26, 2024 • 1 new comment

• [WIP] inductor use rand4x

  #117125 commented on Apr 30, 2024 • 1 new comment

• Hacks to work around the fact that ScriptMethod does not have code/signature

  #124449 commented on Apr 29, 2024 • 1 new comment

• [inductor][cpu]GPT2ForSequenceClassification AMP static/dynamic shape default/cpp wrapper single thread accuracy crash

  #123503 commented on Apr 24, 2024 • 1 new comment

• Package manager install on Nvidia Grace Hopper does not make cuda available

  #123835 commented on Apr 24, 2024 • 1 new comment

• HOP dispatch isn't faithful

  #124484 commented on Apr 24, 2024 • 1 new comment

• Placeholder tensor is empty!

  #123171 commented on Apr 24, 2024 • 1 new comment

• Dynamo Export: Support for PixelShuffle

  #124338 commented on Apr 24, 2024 • 1 new comment

• arm64-v8a not compiling due to libpytorch_jni.so

  #51020 commented on Apr 24, 2024 • 1 new comment

• torch.compiler.disable doesn't disable nested functions (also doesn't work as a context manager)

  #123771 commented on Apr 24, 2024 • 1 new comment

• c10::CUDAError

  #67978 commented on Apr 25, 2024 • 1 new comment

• Batch size is hardcoded using torch.jit.trace with LSTMCell

  #59530 commented on Apr 25, 2024 • 1 new comment

• NCCL error of PyTorch 2.1.0 when using multiple gpus

  #113245 commented on Apr 25, 2024 • 1 new comment

• Doesn't work when register hook to torch.nn.MultiheadAttention.out_proj

  #78109 commented on Apr 25, 2024 • 1 new comment

• jit.freeze throws RuntimeError: stack_out && stack_out->size() == 1 INTERNAL ASSERT FAILED at "../torch/csrc/jit/passes/frozen_conv_folding.cpp":281

  #80861 commented on Apr 25, 2024 • 1 new comment

• calling nn.utils.parametrize inside torch.compile leads to error

  #115744 commented on Apr 25, 2024 • 1 new comment

• torch._dynamo.exc.Unsupported: call_function args: UserDefinedObjectVariable(EasyDict)

  #120219 commented on Apr 25, 2024 • 1 new comment

• Update test_cuda.py and test_torch.py optim tests to use OptimizerInfo and optim_db

  #123451 commented on Apr 25, 2024 • 1 new comment

• PyTorch 2.0.0 encountered CUDA error: an illegal memory access was encountered

  #99372 commented on Apr 25, 2024 • 1 new comment

• Transformer Engine Checkpointing Broken on Torch 2.3

  #122946 commented on Apr 25, 2024 • 1 new comment

• [inductor][cpu]pyhpc_turbulent_kinetic_energy AMP multithread static/dynamic shape default/cpp wrapper performance regression

  #123801 commented on Apr 26, 2024 • 1 new comment

• orch.onnx.errors.UnsupportedOperatorError: Exporting the operator 'aten::unflatten' to ONNX opset version 12 is not supported.

  #124716 commented on Apr 26, 2024 • 1 new comment

• masked_fill supports PrivateUse1, when value.device.type is cpu

  #124693 commented on Apr 26, 2024 • 1 new comment

• Discrepancy between CPU->GPU and GPU->CPU data transfer speeds

  #52718 commented on Apr 26, 2024 • 1 new comment

• Advanced indexing with uint8 tensor versus int64 tensor is inconsistent

  #20149 commented on Apr 26, 2024 • 1 new comment

• Fused Linear and Cross-Entropy Loss `torch.nn.functional.linear_cross_entropy`

  #124480 commented on Apr 26, 2024 • 1 new comment

• DISABLED test_inplace_on_view_weak_grad_fn (__main__.TestAutogradWithCompiledAutograd)

  #124453 commented on Apr 26, 2024 • 1 new comment

• Unexpected modification to CPU affinity of Dataloader workers

  #101850 commented on Apr 26, 2024 • 1 new comment

• fusion in fx graph mode did not take care of direct attribute access

  #68892 commented on Apr 26, 2024 • 1 new comment

• DTensor + compile error's during backward when output is non-contiguous

  #118219 commented on Apr 23, 2024 • 1 new comment

• Significant performance degradation with multiprocessing in PyTorch 2.x compared to 1.13.1

  #122626 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_aot_export_module_joint (__main__.TestAOTExport)

  #124166 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_source_multithreaded_complex_work_in_main_thread_True (__main__.TestProfiler)

  #119536 commented on Apr 23, 2024 • 1 new comment

• torch_dispatch has unfaithful behavior w.r.t. wrapped numbers

  #124731 commented on Apr 23, 2024 • 1 new comment

• Importing polars before torch causes a segfault

  #124656 commented on Apr 23, 2024 • 1 new comment

• [torch.compile] torch._dynamo.exc.TorchRuntimeError: Failed running call_function <method 'numpy' of 'torch._C.TensorBase' objects>(*(FakeTensor(..., size=(32, 3, 64, 64)),), **{})

  #124247 commented on Apr 23, 2024 • 1 new comment

• `test_scatter_bf16_cuda` fails on V100

  #118581 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_sparse_tensors (__main__.TestTorchTidyProfiler)

  #124253 commented on Apr 23, 2024 • 1 new comment

• Find a common home for decompositions, perhaps outside of the obliquely named _refs directory

  #124427 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_save_on_cpu_and_checkpoint (__main__.TestAutogradWithCompiledAutograd)

  #124706 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_forward_mode_AD_linalg_lu_cuda_float64 (__main__.TestFwdGradientsCUDA)

  #86774 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_saved_variable_packing_unpacking_did_not_save_original_with_default_hooks (__main__.TestAutogradWithCompiledAutograd)

  #124733 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_saved_tensor_hooks_custom_function_intermediates (__main__.TestAutogradWithCompiledAutograd)

  #124723 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_inplace (__main__.TestAutogradWithCompiledAutograd)

  #124446 commented on Apr 23, 2024 • 1 new comment

• Add BufferDict container

  #37386 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_source_multithreaded_multiple_preexisting_work_in_main_thread_True (__main__.TestProfiler)

  #119576 commented on Apr 23, 2024 • 1 new comment

• DISABLED test_source_multithreaded_open_in_scope_work_in_main_thread_True (__main__.TestProfiler)

  #119668 commented on Apr 23, 2024 • 1 new comment

• aot_export_joint_simple on plain callable (not graph module) doesn't attach stack traces

  #102205 commented on Apr 23, 2024 • 1 new comment

• MPS memory leak in training

  #121113 commented on Apr 24, 2024 • 1 new comment

• [nightly][jit] bad constant exponent (e+38.f) in default_program fused_mul_div_add

  #107503 commented on Apr 24, 2024 • 1 new comment

• Change `GradScaler` to respect an existing `grad_scale` value.

  #123428 commented on Apr 24, 2024 • 1 new comment

• CudaHostAlloc takes a lot of time during training

  #124456 commented on Apr 24, 2024 • 1 new comment

• RuntimeError: derivative for aten::_scaled_dot_product_efficient_attention_backward is not implemented

  #117974 commented on Apr 24, 2024 • 1 new comment

• ImportError `undefined symbol: iJIT_NotifyEvent` encountered when MKL 2024.1 is installed.

  #123097 commented on Apr 24, 2024 • 1 new comment

• [functorch] transforms like jacrev, jacfwd, grad, etc don't work with BatchNorm

  #85533 commented on Apr 24, 2024 • 1 new comment

• Export swallows exception

  #111075 commented on Apr 27, 2024 • 1 new comment

• Multi Scale Deformable Attention Support

  #112827 commented on Apr 27, 2024 • 1 new comment

• FSDP crashes when submodule calls method that isn't `forward()`

  #109385 commented on Apr 28, 2024 • 1 new comment

• [inductor][cpu] FP32/AMP models multiple/single thread static/dynamic shape default/CPP wrapper accuracy crash in 2024-04-14 nightly release

  #124286 commented on Apr 28, 2024 • 1 new comment

• Custom ROCm hip and C++ extensions (replicated from pytorch/tutorials)

  #119429 commented on Apr 29, 2024 • 1 new comment

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

  #24422 commented on Apr 29, 2024 • 1 new comment

• ONNX export is unnecessarily slow (O(N^2))

  #121422 commented on Apr 28, 2024 • 1 new comment

• [RFC] PyTorch next wheel build platform: manylinux-2.28

  #123649 commented on Apr 28, 2024 • 1 new comment

• vec_test_all_types_xxx with dtype c10::complex<float> and c10::complex<double> has failures on division

  #104516 commented on Apr 29, 2024 • 1 new comment

• ImportError: libcudnn.so.8: cannot open shared object file: No such file or directory

  #104259 commented on Apr 29, 2024 • 1 new comment

• RecursionError when running torch.jit.script inside JitTestCase

  #76881 commented on Apr 29, 2024 • 1 new comment

• Add a requirements.txt for windows pip packages

  #103354 commented on Apr 29, 2024 • 1 new comment

• PyTorch Memory Management in GPU-to-CPU Transfers issue

  #124487 commented on Apr 29, 2024 • 1 new comment

• Compile doesn't guard on user NN module attribute

  #124717 commented on Apr 29, 2024 • 1 new comment

• Error: IndexError: map::at When using torch.distributed.all_reduce(tensor)

  #116393 commented on Apr 29, 2024 • 1 new comment

• IndexError: map::at with MPI CUDA collectives

  #114040 commented on Apr 29, 2024 • 1 new comment

• [feature request] np.packbits / np.unpackbits, general BitTensors (maybe can be just tensors with dtype torch.bits8 or have a new dtype torch.bits introduced) and bit packed tensors utilities for saving memory / accesses, support for BitTensors wherever BoolTensors are used

  #32867 commented on Apr 29, 2024 • 1 new comment

• DISABLED test_leaf_assignment (__main__.TestAutogradWithCompiledAutograd)

  #124405 commented on Apr 29, 2024 • 1 new comment

• [FX] Ability to wrap functions in other modules for symbolic tracing

  #53534 commented on Apr 29, 2024 • 1 new comment

• switch more test cases to use MultithreadTestCase

  #108744 commented on Apr 27, 2024 • 1 new comment

• Support using SymBool in arithmetics

  #110738 commented on Apr 27, 2024 • 1 new comment

• Loading traced pytorch model to C++

  #124009 commented on Apr 27, 2024 • 1 new comment

• [RFC] Dynamo Single Step Graph

  #117394 commented on Apr 29, 2024 • 1 new comment

• large model, low memory: need `torch.load` that loads one submodule at a time

  #75242 commented on Apr 29, 2024 • 1 new comment

• torch.onnx： operator 'aten::unflatten' to ONNX is not supported.

  #121301 commented on Apr 26, 2024 • 1 new comment

• Improve behaviour of `torch.linalg.lstsq` on CUDA GPU for rank defficient matrices

  #117122 commented on Apr 27, 2024 • 1 new comment

• Libtorch crashes docker when included in header file

  #124197 commented on Apr 27, 2024 • 1 new comment

• Segfault in TCPStore and FileStore compare_set()

  #123983 commented on Apr 29, 2024 • 1 new comment

• [Meta Tensor] Inplace set storage of meta tensor will alter the storage's nbytes if meta tensor's nbytes is smaller

  #123879 commented on Apr 29, 2024 • 1 new comment

• DISABLED test_mark_non_differentiable_none (__main__.TestAutogradWithCompiledAutograd)

  #124475 commented on Apr 30, 2024 • 1 new comment

• Registering function that takes `const SymInt&` to op that accepts `SymInt` leads to cryptic error

  #124645 commented on Apr 23, 2024 • 0 new comments

• DISABLED test_default_partitioner_output_tensor_shape_tensor (__main__.TestPartitioning)

  #124355 commented on Apr 29, 2024 • 0 new comments

• DISABLED test_contiguous (__main__.TestPartitioning)

  #124323 commented on Apr 29, 2024 • 0 new comments

• DISABLED test_default_partitioner_getitem (__main__.TestPartitioning)

  #124278 commented on Apr 29, 2024 • 0 new comments

• DISABLED test_aot_export_simplified_basic (__main__.TestAOTExport)

  #124254 commented on Apr 29, 2024 • 0 new comments

• DISABLED test_aot_export_multiple_outputs_require_grad_banned (__main__.TestAOTExport)

  #124221 commented on Apr 29, 2024 • 0 new comments

• Conv with permutation on MPS will lead to negative MSE loss

  #124621 commented on Apr 24, 2024 • 0 new comments

• mps bug: failed assertion `[MPSNDArrayDescriptor sliceDimension:withSubrange:] error: subRange.start (6) is not less than length of dimension[0] (6)'

  #96153 commented on Apr 24, 2024 • 0 new comments

• [inductor][cpu]DebertaV2ForQuestionAnswering AMP static/dynamic shape multiple thread default wrapper regression

  #122390 commented on Apr 24, 2024 • 0 new comments

• torch.quantile on MPS doesn't sort values when dim is not None

  #101878 commented on Apr 24, 2024 • 0 new comments

• Foreach tests should xfail on all dtypes that are not supported

  #124726 commented on Apr 23, 2024 • 0 new comments

• [DONOTREVIEW][DTenosr][Test] DTensor 2D sharding

  #124339 commented on Apr 23, 2024 • 0 new comments

• [inductor][cpu]adv_inception_v3, gluon_inception_v3 and inception_v3 AMP performance regression

  #122393 commented on Apr 24, 2024 • 0 new comments

• squash of flight_5 vs flightbase

  #124229 commented on Apr 25, 2024 • 0 new comments

• torch._export can't export resnet50 model

  #124595 commented on Apr 23, 2024 • 0 new comments

• flight51 squashed vs flightbase

  #124236 commented on Apr 25, 2024 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on Apr 30, 2024 • 0 new comments

• fix a typo in the householder_product docs

  #124279 commented on Apr 30, 2024 • 0 new comments

• [Tracker] torch.sparse semi-structured 2.3 beta release

  #115662 commented on Apr 24, 2024 • 0 new comments

• [feature request] Caching allocator diagnostics and memory allocation tracing/visualization

  #1529 commented on Apr 24, 2024 • 0 new comments

• Dynamo-based ONNX Export: Failed to produce a graph during tracing as no tensor operations were found.

  #123973 commented on Apr 23, 2024 • 0 new comments

• DISABLED test_refcounts (__main__.TestTorchTidyProfiler)

  #124220 commented on Apr 29, 2024 • 0 new comments

• `torch.func.functional_call` doesn't work with compiled models

  #97909 commented on Apr 23, 2024 • 0 new comments

• [TESTING] Don't clamp upper to 2

  #124631 commented on Apr 23, 2024 • 0 new comments

• DISABLED test_binary_op_list_error_cases__foreach_add_cuda_int16 (__main__.TestForeachCUDA)

  #124636 commented on Apr 23, 2024 • 0 new comments

• squash of flight_5.3 vs flightbase

  #124672 commented on Apr 25, 2024 • 0 new comments

• DISABLED test_tensor_subclasses (__main__.TestScript)

  #119949 commented on Apr 23, 2024 • 0 new comments

• Output Discrepancy between PyTorch Model and Converted ONNX Model

  #124711 commented on Apr 23, 2024 • 0 new comments

• torch._dynamo.allow_in_graph seems to silently no-op on staticmethods

  #124735 commented on Apr 23, 2024 • 0 new comments

• DISABLED test_resnet18_backward_trace_cpu (__main__.TestPythonKeyCPU)

  #124641 commented on Apr 29, 2024 • 0 new comments

• Updated test_cuda.py optim tests to use OptimizerInfo

  #124563 commented on Apr 28, 2024 • 0 new comments

• [WIP][Inductor Intel GPU backend Upstream] Reuse inductor test for Intel GPU (PART 3)

  #124702 commented on Apr 27, 2024 • 0 new comments

• inductor creates unnecessary buffers

  #124653 commented on Apr 23, 2024 • 0 new comments

• RFC: Turn on no-undefined

  #124545 commented on Apr 30, 2024 • 0 new comments

• DISABLED test_aot_module_simplified_preserves_stack_trace (__main__.TestAOTModuleSimplified)

  #124609 commented on Apr 29, 2024 • 0 new comments

• Stable Diffusion Model Error: torch._dynamo.exc.InternalTorchDynamoError: raw

  #124477 commented on Apr 23, 2024 • 0 new comments

• [supermodules] Remove all supermodule labels

  #124521 commented on Apr 25, 2024 • 0 new comments

• DISABLED test_aot_module_simplified_fake_tensor_gm_raises (__main__.TestAOTModuleSimplified)

  #124590 commented on Apr 29, 2024 • 0 new comments

• [minimizer] Add exclusion function to minimizer base

  #124504 commented on Apr 30, 2024 • 0 new comments

• [dynamo] Allow inlining of hooks for the top module

  #124501 commented on Apr 30, 2024 • 0 new comments

• [dynamo] Support ndarray.dtype attribute access

  #124490 commented on Apr 24, 2024 • 0 new comments

• fix torch.compile with triton kernels under inference_mode

  #124489 commented on Apr 26, 2024 • 0 new comments

• DISABLED test_aot_module_simplified_dynamic (__main__.TestAOTModuleSimplified)

  #124510 commented on Apr 29, 2024 • 0 new comments

• DISABLED test_aot_module_simplified (__main__.TestAOTModuleSimplified)

  #124476 commented on Apr 29, 2024 • 0 new comments

• DISABLED test_aot_dispatch_incorrect_backward (__main__.TestAOTDispatch)

  #124459 commented on Apr 29, 2024 • 0 new comments

• [WIP] [Inductor Intel GPU backend Upstream] Reuse inductor test for Intel GPU (PART 2)

  #124147 commented on Apr 27, 2024 • 0 new comments

• [codemod][lowrisk] Remove extra semi colon from caffe2/c10/core/SymNodeImpl.h

  #123055 commented on Apr 27, 2024 • 0 new comments

• [FSDP2] Eager-Mode Execution Tracker

  #120003 commented on Apr 29, 2024 • 0 new comments

• [onnx.export] Avoid linear loop over symbol_dim_map

  #123029 commented on Apr 25, 2024 • 0 new comments

• FlexAttention isn't using decompositions

  #124643 commented on Apr 24, 2024 • 0 new comments

• Add Gaudi support to benchmarks/dynamo/* benchmark.

  #122960 commented on Apr 24, 2024 • 0 new comments

• [WIP][Inductor Intel GPU backend Upstream] Reuse inductor test for Intel GPU (PART 1)

  #122866 commented on Apr 27, 2024 • 0 new comments

• Make c10::Error empty backtrace as an optional argument

  #122611 commented on Apr 26, 2024 • 0 new comments

• [typing] Rename argument of `nn.Sequential.forward` from `input` to `__input`

  #119208 commented on Apr 28, 2024 • 0 new comments

• Avoid always building stack trace strings in c10::Error

  #122086 commented on Apr 26, 2024 • 0 new comments

• [WIP] Arm64 Enablement

  #117274 commented on Apr 27, 2024 • 0 new comments

• [FSDP] Use generic device handle instead of cuda

  #121620 commented on Apr 24, 2024 • 0 new comments

• [Inductor Cutlass backend] DO NOT REVIEW - to be split up

  #121492 commented on Apr 25, 2024 • 0 new comments

• custom ops should have needs_fixed_stride_order by default

  #124647 commented on Apr 25, 2024 • 0 new comments

• Conflict between ``torch.func`` transformations and ``torch.jit.trace``

  #98724 commented on Apr 25, 2024 • 0 new comments

• Add `ciflow/inductor` for test only changes

  #118206 commented on Apr 23, 2024 • 0 new comments

• lintrunner should fail on badly formatted docstrings

  #102227 commented on Apr 25, 2024 • 0 new comments

• Bfloat16 tensor .numpy() support

  #90574 commented on Apr 25, 2024 • 0 new comments

• [ONNX] stft export fails with dynamo_export

  #113067 commented on Apr 25, 2024 • 0 new comments

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

  #119195 commented on Apr 25, 2024 • 0 new comments

• [dynamo] Validate check_fn

  #118448 commented on Apr 25, 2024 • 0 new comments

• CUDAGraph Tree TORCH_CHECK failed when NCCL operator exists.

  #124391 commented on Apr 26, 2024 • 0 new comments

• [PT2] Return int32 indices in max_pool2d_with_indices

  #103785 commented on Apr 26, 2024 • 0 new comments

• torch.normal ignores default_device

  #122886 commented on Apr 26, 2024 • 0 new comments

• DISABLED test_non_contiguous_tensors_nn_ConvTranspose1d_cuda_complex32 (__main__.TestModuleCUDA)

  #81732 commented on Apr 26, 2024 • 0 new comments

• Switch batch norm stack to consolidated ops

  #119496 commented on Apr 30, 2024 • 0 new comments

• S390x binaries

  #120398 commented on Apr 27, 2024 • 0 new comments

• [FSDP] Removed clamp to `NO_SHARD` for world size 1

  #120334 commented on Apr 24, 2024 • 0 new comments

• Fix implicit fallthroughs where it is simple to do so in caffe2/

  #119700 commented on Apr 27, 2024 • 0 new comments

• [FSDP] Add device in pin_memory argument

  #119878 commented on Apr 24, 2024 • 0 new comments

• [CI] CPU Inductor codepath for AVX2/Default is not tested in CI

  #123224 commented on Apr 24, 2024 • 0 new comments

• [inductor] Enable fx graph caching by default

  #124091 commented on Apr 30, 2024 • 0 new comments

• cpu performance for int4mm kernels

  #122813 commented on Apr 24, 2024 • 0 new comments

• [WIP][inductor] refine loop split logic

  #124060 commented on Apr 30, 2024 • 0 new comments

• Inconsistent results when training a model containing SyncBatchNorm with multiple GPUs

  #124680 commented on Apr 24, 2024 • 0 new comments

• [Inductor] [Quant] Enable lowering of quant per tensor and refactor quant pattern

  #124041 commented on Apr 30, 2024 • 0 new comments

• Add scaled_dot_product_attention "scale" argument to nn.MultiHeadAttention

  #124718 commented on Apr 24, 2024 • 0 new comments

• Migrating from setup.py install/develop to leverage pip standards

  #124027 commented on Apr 28, 2024 • 0 new comments

• [inductor][cpp] GEMM template

  #124021 commented on Apr 30, 2024 • 0 new comments

• Questions about parameter initialization, especially with torch.bfloat16 precision

  #124719 commented on Apr 24, 2024 • 0 new comments

• Fix constant propagation pass

  #114471 commented on Apr 26, 2024 • 0 new comments

• [discussion] Route pointwise Conv1d/Conv2d to matmul? (also in eager)

  #116506 commented on Apr 24, 2024 • 0 new comments

• Initial LR Scheduler composability tests

  #123753 commented on Apr 25, 2024 • 0 new comments

• Fix user warning for tensor LR

  #123752 commented on Apr 25, 2024 • 0 new comments

• Swap warning counter to flag in LRScheduler

  #123751 commented on Apr 25, 2024 • 0 new comments

• Add decomposition for slice_scatter

  #123744 commented on Apr 26, 2024 • 0 new comments

• Dynamo x autograd.Function: graph breaks on all the staticmethods on autograd.Function

  #118397 commented on Apr 24, 2024 • 0 new comments

• [debug] a debug PR to test perf regression due to triton

  #123694 commented on Apr 30, 2024 • 0 new comments

• Dynamo x autograd.Function: graph breaks on freevars in forward

  #118394 commented on Apr 24, 2024 • 0 new comments

• Use _unsafe_masked_index in masked_scatter decomposition

  #123667 commented on Apr 26, 2024 • 0 new comments

• Improve decomposition for constand_pad_nd

  #123661 commented on Apr 26, 2024 • 0 new comments

• Dynamo x autograd.Function: silently ignores all of the ctx.methods

  #118396 commented on Apr 24, 2024 • 0 new comments

• Automated submodule update: FBGEMM

  #115316 commented on Apr 30, 2024 • 0 new comments

• Rename TorchDynamo -> Dyanamo in the dynamo tutorial doc

  #123431 commented on Apr 27, 2024 • 0 new comments

• Reenable dim for python 3.12

  #123384 commented on Apr 24, 2024 • 0 new comments

• [Dynamic Shapes] Fix error handling for indirectly fully constrained dynamic dimensions

  #123293 commented on Apr 30, 2024 • 0 new comments

• [CI] Node-20 update

  #122115 commented on Apr 24, 2024 • 0 new comments

• Add mode to MemoryDep to track atomic accumulates

  #123223 commented on Apr 26, 2024 • 0 new comments

• Decompositions for upsample linear backward

  #123222 commented on Apr 26, 2024 • 0 new comments