7 Pull requests merged by 5 people

• [EZ] Pin scipy to 1.12 for Py-3.12

  #127322 merged May 29, 2024

• Update hf_BirdBird periodic-dynamo-benchmarks results

  #127312 merged May 29, 2024

• Put back "[Release only] Release 2.3 start using triton package from pypi""

  #127290 merged May 28, 2024

• [DSD] Fix to remove non_persistent buffer in distributed state dict (#125337)

  #127219 merged May 27, 2024

• [DSD] Add a test to verify FSDP lazy initialization case (#127069)

  #127130 merged May 27, 2024

• [DCP][state_dict] Remove the check of FSDP has root (#121544)

  #126557 merged May 27, 2024

• [DSD] Correctly handle _extra_state (#125336)

  #126567 merged May 27, 2024

196 Pull requests opened by 110 people

• Warn unused functions

  #127185 opened May 26, 2024

• Quick Fix on #126854, deepcopy `lr` and other possible `base_parameters`

  #127190 opened May 26, 2024

• [inductor][cpp] BF16 AMX micro-gemm support

  #127195 opened May 26, 2024

• Reduce number of samples in {svd,pca}_lowrank OpInfos

  #127199 opened May 26, 2024

• Implement a generic function scheduler

  #127200 opened May 26, 2024

• [Inductor][CPP] Enable int8 GEMM Template

  #127206 opened May 27, 2024

• Add error checking for boolean beta and alpha for fake tensor impl of torch.addr

  #127207 opened May 27, 2024

• [Inductor] Skip model_fail_to_load and eager_fail_to_run models in inductor benchmarks test

  #127210 opened May 27, 2024

• [Inductor][CPP] Fix FP16 GEMM Template UT failure with FP16 instruction support

  #127211 opened May 27, 2024

• [Inductor][CPP] Fallback QLinear Binaryfusion from postop sum to binary add when others is view

  #127212 opened May 27, 2024

• Enable OP: convtranspose into operator benchmark

  #127216 opened May 27, 2024

• Add OpInfo entry for as_strided_copy

  #127231 opened May 27, 2024

• Add OpInfo entry for add_alias

  #127232 opened May 27, 2024

• [Clang Tidy] Fix misc-header-include-cycle errors in clang-tidy and ignore some files

  #127233 opened May 27, 2024

• [easy?] Move AsyncCompile to a different file

  #127235 opened May 27, 2024

• [Traceable FSDP2] Make ._unsharded_param creation traceable

  #127238 opened May 27, 2024

• [Do not review] Top of Traceable FSDP2 stack - to be broken up

  #127239 opened May 27, 2024

• [MPS] Fused Adam & AdamW

  #127242 opened May 27, 2024

• [Traceable FSDP2] Make ._unsharded_param creation traceable

  #127243 opened May 28, 2024

• [DO NOT REVIEW][NOT USED] Add queue_callback support

  #127244 opened May 28, 2024

• [Traceable FSDP2] Make ._unsharded_param creation traceable

  #127245 opened May 28, 2024

• [DO NOT REVIEW][NOT USED] Add queue_callback support

  #127246 opened May 28, 2024

• Add Dynamo support for run_with_rng_state HOP

  #127247 opened May 28, 2024

• [Traceable FSDP2] Improve FSDPManagedNNModuleVariable support

  #127249 opened May 28, 2024

• fix code

  #127252 opened May 28, 2024

• Support SetVariable mutation

  #127253 opened May 28, 2024

• Check hasattr before comparing source name

  #127254 opened May 28, 2024

• [Traceable FSDP2] Workaround in nn_module_proxy()

  #127255 opened May 28, 2024

• [WIP] Improve __bool__ access handling for UserDefinedObjectVariable

  #127257 opened May 28, 2024

• Trace TensorVariable attribute mutation if call_setattr is called

  #127258 opened May 28, 2024

• Add support for register_post_accumulate_grad_hook

  #127259 opened May 28, 2024

• Add tracing support for out-variant custom ops that return None

  #127260 opened May 28, 2024

• [NOT USED] support inplace all_gather

  #127261 opened May 28, 2024

• [Doc] fix some typos (found by codespell and typos)

  #127267 opened May 28, 2024

• Add new Inductor-friendly op _fp8_mm and lowering

  #127268 opened May 28, 2024

• [quant] Enable auto code completion by import during type checking

  #127269 opened May 28, 2024

• add xpu for amp

  #127276 opened May 28, 2024

• Enable deterministic support for oneDNN

  #127277 opened May 28, 2024

• update amp example to device-agnostic

  #127278 opened May 28, 2024

• add xpu to torch.compile

  #127279 opened May 28, 2024

• add xpu to torch.tensors

  #127280 opened May 28, 2024

• Select Runner Label Dynamically

  #127287 opened May 28, 2024

• [inductor] enable bf32 test for mkldnn conv

  #127293 opened May 28, 2024

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

  #127294 opened May 28, 2024

• functorch x torch.compile: Remove functools.wraps hack

  #127302 opened May 28, 2024

• Remove parts of AOTDedupWrapper logic since we have dynamo dedup

  #127306 opened May 28, 2024

• [halide-backend] Add test shard

  #127308 opened May 28, 2024

• Use by-column algorithm for fp16/bf16 CPUBlas gemm_transb kernels

  #127318 opened May 28, 2024

• Remove tensor storage_offset/storage_bytes from the cache key

  #127319 opened May 28, 2024

• [PT2Quant] Migrate move_exported_model_to_eval to torch.export._trace._export

  #127323 opened May 28, 2024

• [export] Fix fake mode detection with empty inputs.

  #127327 opened May 28, 2024

• [do-not-review][wip] remove TLS cudagraph tree manager for CA

  #127330 opened May 28, 2024

• [wip] mark compiled autograd params as static addresses for cudagraphs

  #127331 opened May 28, 2024

• [pipelining] rewrite interleaved 1f1b

  #127332 opened May 28, 2024

• [Inductor] Emit strided block pointer from ModularIndexing and FloorDiv

  #127342 opened May 28, 2024

• add auto-functionalize support for mutable list[Tensor]

  #127347 opened May 28, 2024

• Retire torch.distributed.pipeline

  #127354 opened May 28, 2024

• [Functorch][cuDNN] Bump tolerances for `test_vmapjvpvjp`

  #127355 opened May 28, 2024

• [dtensor][debug] added c10d reduce_scatter_ and reduce_scatter_tensor_coalesced tracing_ to CommDebugMode

  #127358 opened May 28, 2024

• [dtensor][debug] added c10d alltoall_ and alltoall_base_ to CommDebugMode

  #127360 opened May 28, 2024

• [c10d] guard gpu context during abort

  #127363 opened May 29, 2024

• UFMT format on test_fake_tesnor.py test_futures.py test_fx.py

  #127369 opened May 29, 2024

• Migrate CalculateSmallVectorDefaultInlinedElements to constexpr Function for SmallVector

  #127370 opened May 29, 2024

• Remove cuda check in the CUDAGraph destrcutor

  #127382 opened May 29, 2024

• Masked scale meta function registration #119984

  #127389 opened May 29, 2024

• Build SYCL kernels for ATen XPU ops on Native Windows (take 2)

  #127390 opened May 29, 2024

• [sym_shapes] Allow check_is_size for backed symints

  #127395 opened May 29, 2024

• [inductor] Enable subprocess-based parallel compile internally

  #127401 opened May 29, 2024

• Handle unpacking during TorchScript to ExportedProgram conversion

  #127419 opened May 29, 2024

• [c10d] Add commCreateFromRanks to c10d

  #127421 opened May 29, 2024

• [inductor] Take absolute value of strides when picking loop order

  #127425 opened May 29, 2024

• Implement Graph Transform Observer

  #127427 opened May 29, 2024

• Add NaturalDiv to distinguish from FloorDiv/TruncDiv

  #127430 opened May 29, 2024

• [DRAFT] nested tensor subclass support

  #127431 opened May 29, 2024

• [export] provide refine function for automatically accepting dynamic shapes suggested fixes

  #127436 opened May 29, 2024

• [jit] Validate mobile module fields parsed by flatbuffer loader

  #127437 opened May 29, 2024

• Update to cuda 12.4.1

  #127439 opened May 29, 2024

• [TESTING] Run inductor CI on custom Triton pin

  #127450 opened May 29, 2024

• force_stride_order on fused_all_gather_matmul/fused_matmul_reduce_scatter's operands to avoid a copy due to layout transformation

  #127454 opened May 29, 2024

• Improve the scheduling for fused_matmul_reduce_scatter

  #127455 opened May 29, 2024

• fix post_grad pattern

  #127457 opened May 29, 2024

• Add typing annotations to pattern_matcher.py

  #127458 opened May 29, 2024

• documentation for pattern_matcher.py

  #127459 opened May 29, 2024

• Add registry for TorchScript to ExportedProgram conversion

  #127464 opened May 29, 2024

• update test_reformer_train test to handle nn module inlining

  #127467 opened May 29, 2024

• Add new convenience checks for PyTorch operators/kernels

  #127469 opened May 29, 2024

• Save quantization_tag in export graph serialization

  #127473 opened May 29, 2024

• [pipelining] Stress test schedules with multi iters

  #127475 opened May 29, 2024

• Onboard ARM bfloat16 to gemm-by-dot-product-for-gemm_transa_ infrastructure

  #127477 opened May 29, 2024

• Patch ARM Half use_gemv_fast_path gate to avoid kernel duplication

  #127478 opened May 29, 2024

• [pipelining] Add forward-only tests to mimic inference

  #127479 opened May 29, 2024

• Onboard ARM bfloat16 to gemv fast path

  #127484 opened May 29, 2024

• reset dynamo in test_do_not_skip_side_effects unit test loop to avoid dynamo cache limit hit

  #127487 opened May 30, 2024

• Use bfdot instruction in ARM bfloat16 dot product if compiling with it available

  #127488 opened May 30, 2024

• [caffe2][be] migrate global static initializer

  #127489 opened May 30, 2024

• [NestedTensor] Extend coverage for unbind when ragged_idx != 1

  #127493 opened May 30, 2024

• Properly detect nested torch function args

  #127496 opened May 30, 2024

• Check unused variables in tests

  #127498 opened May 30, 2024

• Use less A100.large shards for torchao perf benchmark

  #127499 opened May 30, 2024

• [cpuinfo] bump cpuinfo to the latest to support amx isa check

  #127505 opened May 30, 2024

• [halide-backend] Add GPU support

  #127506 opened May 30, 2024

• [Intel GPU]Enable fp64 GEMM

  #127507 opened May 30, 2024

• [Intel GPU]Enable fp64 double GEMM

  #127508 opened May 30, 2024

• [BE][ptd_fb_test][1/N] Enable testslide

  #127512 opened May 30, 2024

• Add linker script optimization flag to CMAKE rule for CUDA ARM wheel

  #127514 opened May 30, 2024

• Supervisor as a torchrun rendezvous impl

  #127515 opened May 30, 2024

• Try setting memory budget to 0.5 by default

  #127520 opened May 30, 2024

• [TORCH_FA2_flash_api] Update total_q to the reshaped query 0th dimension

  #127524 opened May 30, 2024

• Validate file and handle exceptions for weights_only unpickler

  #127526 opened May 30, 2024

• [ROCm] Fix error in torch.cuda initialisation if amdsmi is not available

  #127528 opened May 30, 2024

• [WIP] Add decomposition for aten.bernoulli.p

  #127537 opened May 30, 2024

• Always simplify sympy expressions before printing.

  #127543 opened May 30, 2024

• Handle aten::__contains__ during TorchScript to ExportedProgram conversion

  #127544 opened May 30, 2024

• Add noqa to prevent lint warnings

  #127545 opened May 30, 2024

• Update doc for nn.Module.load_state_dict(assign=True) to be more clear about semantic

  #127549 opened May 30, 2024

• add wider support for input mutations during backward in torch.compile, as long as they are hidden from autograd

  #127551 opened May 30, 2024

• update test_ddp_graphs to reflect inlined nn module behaviour, ddp graph splitting not triggered when modules inlined.

  #127553 opened May 30, 2024

• [export] only add guard-related runtime asserts once, silence if needed

  #127554 opened May 30, 2024

• Add profiler annotation for fused_all_gather_matmul and fused_matmul_reduce_scatter

  #127556 opened May 30, 2024

• [pipelining] test pipeline_order in schedule

  #127559 opened May 30, 2024

• Remove unstable ARC jobs

  #127563 opened May 30, 2024

• Joy dev

  #127564 opened May 30, 2024

• [Split Build] Make libtorch_global_deps accessible from libtorch wheel

  #127570 opened May 30, 2024

• [DO NOT MERGE] Testing new runners

  #127573 opened May 30, 2024

• [AOTInductor] [Tooling] Update NaN and INF Checker for AOTInductor

  #127574 opened May 30, 2024

• Use freshly traced jit-traced module to be used in export analysis

  #127577 opened May 30, 2024

• [DO NOT MERGE] Testing additional LF config labels

  #127579 opened May 30, 2024

• Handle custom op during TorchScript to ExportedProgram conversion

  #127580 opened May 30, 2024

• [inductor] fix redis-related env vars in remote_cache.py

  #127583 opened May 30, 2024

• [FSDP2] Fix submesh slicing to enable 3D parallelism

  #127585 opened May 31, 2024

• [WIP] mark NestedInts as symints instead of symfloats

  #127587 opened May 31, 2024

• [CI] Comment hf_T5_generate, hf_GPT2 and timm_efficientnet in inductor cpu smoketest for performance unstable issue

  #127588 opened May 31, 2024

• [Do not merge] [Test] Test builder cudnn v9 change

  #127589 opened May 31, 2024

• [Quant][Inductor] Add get_mutation_names in QLinearPointwiseBinaryPT2E IR

  #127592 opened May 31, 2024

• [ts migration] support aten::dim, aten::len, aten::__getitem__

  #127593 opened May 31, 2024

• Disable one of the GPT2 SDPA patterns for single-thread case

  #127594 opened May 31, 2024

• WIP: fake tensor SymInt support

  #127596 opened May 31, 2024

• test linear add bias

  #127597 opened May 31, 2024

• [Inductor][CPP] Support more than one LocalBuffer

  #127598 opened May 31, 2024

• [fbgemm_gpu] remove deleted embedding_backward_dense_host references

  #127599 opened May 31, 2024

• Don't find statically linked libs in TorchConfig.cmake.in

  #127601 opened May 31, 2024

• [WIP] Reuse UT for Intel GPU backend [Part1]

  #127602 opened May 31, 2024

• [export][unflatten] More strictly respect scope when removing inputs

  #127607 opened May 31, 2024

• Autoselect default device in FSDP construction.

  #127609 opened May 31, 2024

• [AOTI] align data_size of the constants

  #127610 opened May 31, 2024

• [CI] disable td for xpu ci test by default

  #127611 opened May 31, 2024

• [inductor] custom do_bench_gpu to reduce max-autotune overhead

  #127613 opened May 31, 2024

• Add functionality to make ViewAndMutationData (slightly more) cache safe

  #127618 opened May 31, 2024

• Validate tensor storage before a direct access to it

  #127619 opened May 31, 2024

• [caffe2][be][2/n] migrate gloabl static initializer

  #127620 opened May 31, 2024

• Updating Module Tracker

  #127624 opened May 31, 2024

• [Testing only] Flip default on weights_only

  #127627 opened May 31, 2024

• [RFC] Introduce Checkpointable for DCP (#127540)

  #127628 opened May 31, 2024

• [Inductor UT][Intel GPU] Skip test case which doesn't currently work on the XPU stack but newly re-enabled by community.

  #127629 opened May 31, 2024

• [dtensor][debug] created example test to print module parameters

  #127631 opened May 31, 2024

• wip test change

  #127632 opened May 31, 2024

• [export] Handle serializing duplicate getitem nodes

  #127633 opened May 31, 2024

• [aten_cuda/flash_attn] Add typename to template argument Kernel_trait…

  #127634 opened May 31, 2024

• [DSD] Fixes various bugs for broadcast_from_rank0

  #127635 opened May 31, 2024

• [inductor] parallel-compile: call triton_key() before forking

  #127639 opened May 31, 2024

• [ONNX] Add quantized layer norm op to opset 17

  #127640 opened May 31, 2024

• View specialization

  #127641 opened May 31, 2024

• [ATen][Native] fixes sparse SPMV on aarch64

  #127642 opened May 31, 2024

• [FSDP] keep paras in torch.distributed.checkpoint.state_dict.set_optimizer_state_dict

  #127644 opened May 31, 2024

• GGML inspired int8 MM Metal shader

  #127646 opened May 31, 2024

• Test jit

  #127647 opened May 31, 2024

• [PT2][Optimus] Improve group batch fusion with same parent/users fusion enablement

  #127648 opened May 31, 2024

• Relax Symbol check on debug_compile mode

  #127650 opened May 31, 2024

• [Perf] Provide API to retrieve less data from recorder

  #127651 opened May 31, 2024

• [CI] Gen xml during run

  #127653 opened May 31, 2024

• [c10d][BE] fix test_init_pg_and_rpc_with_same_socket

  #127654 opened May 31, 2024

• [Caffe2] Remove Caffe2 proto and other files

  #127655 opened May 31, 2024

• [ts-migration] support aten::__is__, aten::__isnot__, aten::__not__, profiler::_record_function_enter_new, profiler::_record_function_exit

  #127656 opened May 31, 2024

• [inductor] simplify indexing

  #127661 opened Jun 1, 2024

• Allow symint inputs to aten.expand_copy and aten.view_copy

  #127662 opened Jun 1, 2024

• Inductor: Allow small sizes of m for mixed mm autotuning

  #127663 opened Jun 1, 2024

• adjust thresholds for gluon_inception_v3, beit_base_patch16_224, phli…

  #127664 opened Jun 1, 2024

• [c10d] add a simple test to demonstrate the user usage of collectives

  #127665 opened Jun 1, 2024

• [wip][cudagraphs] static input params flag

  #127668 opened Jun 1, 2024

• [DO NOT MERGE] Fuzzkatt/12 4 inductor tolerance fixes debug

  #127669 opened Jun 1, 2024

• declare 'static' if the function is not intended to be used outside o…

  #127670 opened Jun 1, 2024

• [Inductor][CI][CUDA 12.4] Update dynamic_inductor_timm_training.csv - change gluon_inception_v3 from fail_accuracy to pass

  #127672 opened Jun 1, 2024

• [pipelining][review-only] Simple 1F1B schedule

  #127673 opened Jun 1, 2024

• [AOTI] Switch to use shim v2

  #127674 opened Jun 1, 2024

• [WIP] add optional boolean arg "true" to cumsum signature

  #127675 opened Jun 1, 2024

• [Inductor][Flex-attention] Support different sequence lengths for Query and Key/Value

  #127678 opened Jun 1, 2024

• [symbolic shapes] if symbol not in var_to_range default to unknown range

  #127681 opened Jun 1, 2024

• wip

  #127682 opened Jun 1, 2024

• Simplify CMake code

  #127683 opened Jun 1, 2024

• [BE][Ez]: Enable ruff PYI019

  #127684 opened Jun 1, 2024

• [BE][Ez]: Apply PYI059 - Generic always come last

  #127685 opened Jun 1, 2024

• [inductor][cpp] align dtype of local buffer with the global one

  #127687 opened Jun 1, 2024

• [BE]: Enable ruff TCH rules and autofixes for better imports

  #127688 opened Jun 1, 2024

• [BE] wrap deprecated function/class with `typing_extensions.deprecated`

  #127689 opened Jun 1, 2024

• Deprecate `torch._utils.is_compiling()` and `torch._dynamo.external_utils.is_compiling()`

  #127690 opened Jun 1, 2024

• Retry of D58015187 Move AsyncCompile to a different file

  #127691 opened Jun 1, 2024

112 Issues closed by 33 people

• TreeSpec equal logic return False for the same tree in str format

  #127447 closed Jun 1, 2024

• Upgrade MacOS runner to 14

  #127490 closed Jun 1, 2024

• UNSTABLE trunk / macos-py3-arm64

  #127542 closed Jun 1, 2024

• When testing the scalar version, test_torchinductor.py will fail

  #126763 closed Jun 1, 2024

• [Feature] `_foreach_copy_` supports different src/dst dtype with fusion

  #115171 closed Jun 1, 2024

• Segmentation fault between Numpy and Pytorch using torch.bmm

  #93161 closed May 31, 2024

• Type conversion between float/complex

  #97888 closed May 31, 2024

• PyTorch Distributed Load Updates or Returns `state_dict`

  #125096 closed May 31, 2024

• [NestedTensor] RelaxedUnspecConstraint failures due to mark_dynamic in NT constructor

  #127097 closed May 31, 2024

• Build without MKL is not possible when MKL is installed

  #32407 closed May 31, 2024

• [RFC] rename allow_in_graph to unsafe_allow_in_graph

  #122189 closed May 31, 2024

• dcp resharding does not work for optimizer state_dict

  #91205 closed May 31, 2024

• Remove `coordinator_rank` from public fns in distributed checkpointing, e.g. `load`

  #119205 closed May 31, 2024

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

  #124942 closed May 31, 2024

• DCP sees 1/2 of the expected size of each tensor in 3D parallel

  #126595 closed May 31, 2024

• Loading Old Checkpoints with DTensor

  #127351 closed May 31, 2024

• The training always freezes after some epochs.

  #22671 closed May 31, 2024

• [AOTI][cpp_wrapper] test_triton_heuristics.py::test_artificial_grid_cpp_wrapper disabled

  #123210 closed May 31, 2024

• [ONNX] export.export + dynamo_export fail on GLU because decompositions are not run

  #125894 closed May 31, 2024

• dynamo breaks when getting attributes of builtins

  #127172 closed May 31, 2024

• [optim] add missing 'maximize' parameter to LBFGS, NAdam and RAdam optimizers

  #126642 closed May 31, 2024

• [ONNX] Memory leak when exporting a jit model to onnx

  #82532 closed May 31, 2024

• [ONNX] Simplify fake tensor export logic to get rid of the redundant model argument

  #127141 closed May 31, 2024

• Do we need an N-dim sub-DeviceMesh?

  #126530 closed May 30, 2024

• Fix torch._dynamo.exc.Unsupported: builtin: index when nn module inlining enabled.

  #127426 closed May 30, 2024

• Incorrect output from inductor: tile -> as_strided -> add

  #127474 closed May 30, 2024

• saved_tensors_hooks auto delete custom attributes

  #126676 closed May 30, 2024

• Add a `descending` flag to `linalg.eigh` and `linalg.svd`

  #58034 closed May 30, 2024

• UNSTABLE periodic / linux-jammy-xpu-py3.8 / test (default)

  #127539 closed May 30, 2024

• Error while parsing dispatch dictionary for NativeFunctions YAML in torchgen

  #127405 closed May 30, 2024

• Error while parsing NativeFunction from YAML in torchgen

  #127406 closed May 30, 2024

• Error while parsing DeviceCheckType for NativeFunctions YAML in torchgen

  #127407 closed May 30, 2024

• Error while parsing function signature for NativeFunctions YAML in torchgen

  #127408 closed May 30, 2024

• Missing structured delegate function in YAML in torchgen

  #127409 closed May 30, 2024

• Error while parsing Yaml file in torchgen

  #127410 closed May 30, 2024

• Error while parsing precomputed field for NativeFunctions YAML in torchgen

  #127411 closed May 30, 2024

• Error while parsing function parameter with default value for NativeFunctions YAML in torchgen

  #127404 closed May 30, 2024

• Inconsistent gradients of Conv2d layers when training the same model using CPU and GPU

  #127226 closed May 30, 2024

• DISABLED test_non_contiguous_input_addmm (__main__.TestMaxAutotune)

  #126176 closed May 30, 2024

• DISABLED test_activations_bfloat16_half_cpu_cpu_float16 (__main__.TestNNDeviceTypeCPU)

  #126177 closed May 30, 2024

• DISABLED test_quantization_doc_fx (__main__.TestQuantizationDocs)

  #125670 closed May 30, 2024

• DISABLED test_quantization_doc_ptsq (__main__.TestQuantizationDocs)

  #125669 closed May 30, 2024

• DISABLED test_quantization_doc_custom (__main__.TestQuantizationDocs)

  #125668 closed May 30, 2024

• DISABLED test_quantization_doc_ptdq (__main__.TestQuantizationDocs)

  #125667 closed May 30, 2024

• PyTorch nightly docs build hasn't run since 4/8?

  #127527 closed May 30, 2024

• Check for error messages on torch.compile with pybind'ed functions

  #126799 closed May 30, 2024

• torch.svd_lowrank does not work for complex matrices.

  #122188 closed May 30, 2024

• error: ‘False’ was not declared in this scope

  #127392 closed May 30, 2024

• [Inductor] Masked `tl.load` operations should explicitly include `other` if the masked out values are expected to be used

  #126535 closed May 30, 2024

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

  #119588 closed May 30, 2024

• Document torch.Tensor legacy constructor

  #122408 closed May 29, 2024

• torch.fx.passes.split_module.split_module doesn't support dynamic shapes

  #103539 closed May 29, 2024

• [While_loop] How to use layer like `torch.nn.BatchNorm2d` with while_loop?

  #127320 closed May 29, 2024

• Add `str` type to `device` parameter of `torch.cuda.get_device_name()` on the doc

  #126400 closed May 29, 2024

• [MPS] `.to('mps')` zeroes out elements in tensors taking up >=2^32 bytes

  #96716 closed May 29, 2024

• Fix "failed running function instead of failed running module" when nn module inlining enabled.

  #125605 closed May 29, 2024

• [docs] scaled_dot_product_attention is_causal description is misleading

  #126873 closed May 29, 2024

• Fix torch._dynamo.exc.Unsupported: call_id withh args (UnspecializedNNModuleVariable() when TORCHDYNAMO_INLINE_INBUILT_NN_MODULES=1

  #127095 closed May 29, 2024

• running opcheck leads to `Fail to import hypothesis in common_utils, tests are not derandomized` print

  #126871 closed May 29, 2024

• opcheck has dependency on expecttest, which is not a pytorch runtime dependency, leading to "module not found" error message

  #126870 closed May 29, 2024

• UNSTABLE linux-binary-manywheel / manywheel-py3_8-cuda11_8

  #104727 closed May 29, 2024

• UNSTABLE linux-binary-manywheel / manywheel-py3_8-cuda12_1-test / test

  #127288 closed May 29, 2024

• MAX-Autotune Compilation Time Regression Due To Added MM Configs

  #125687 closed May 29, 2024

• UNSTABLE linux-binary-manywheel / manywheel-py3_8-cuda12_4-test / test

  #127289 closed May 29, 2024

• AOTI doesn't produce meaningful response w/ CPU backend on Linux x86

  #123990 closed May 29, 2024

• [torch.compile] `index_select` out of bound read

  #121251 closed May 29, 2024

• [pipelining] Add back support for multi-use parameters/buffers

  #126626 closed May 29, 2024

• install cuda version always get cpuonly

  #106565 closed May 28, 2024

• test_view_dynamic_zero_dim no longer testing zero input

  #105066 closed May 28, 2024

• DISABLED test_some_outputs_dont_require_grad_view (__main__.TestAOTAutograd)

  #125593 closed May 28, 2024

• DISABLED test_some_output_requires_grad_input_doesnt (__main__.TestAOTAutograd)

  #125402 closed May 28, 2024

• DISABLED test_schema_correctness_nn_functional_conv3d_cuda_complex128 (__main__.TestSchemaCheckModeOpInfoCUDA)

  #114573 closed May 28, 2024

• DISABLED test_view_and_inplace_view (__main__.TestAOTAutograd)

  #125671 closed May 28, 2024

• DISABLED test_aot_export_multiple_outputs_require_grad_banned (__main__.TestAOTExport)

  #124221 closed May 28, 2024

• Profiler reports different # of Calls depending on group_by_stack_n

  #83737 closed May 28, 2024

• DISABLED test_profiler (__main__.TestJit)

  #65521 closed May 28, 2024

• In profiler, recorded block's total time can be less than the operators within the block

  #43868 closed May 28, 2024

• RuntimeError: MPS device does not support bmm for non-float inputs

  #127178 closed May 28, 2024

• Data race in RecordFunction::callbackShouldRun

  #58452 closed May 28, 2024

• Segmentation fault when importing `sklearn.model_selection`

  #127192 closed May 28, 2024

• UNSTABLE pull / linux-focal-cuda12.4-py3.10-gcc9 / build

  #127108 closed May 28, 2024

• pyyaml dependency error on Mac with source installation

  #127158 closed May 28, 2024

• Overly demanding PyTorch CLA

  #127285 closed May 28, 2024

• A UserWarning occurs after CBAM attention is added

  #127198 closed May 28, 2024

• When training done, the mode output same result each tensor input. ( I tried many way to debug, but can't find any way to fix it, so i guess this is a bug )

  #127177 closed May 28, 2024

• CLA required despite being existing contributor

  #127272 closed May 28, 2024

• Tensor slicing reducing dimensionality of tensor

  #127236 closed May 28, 2024

• torch.linalg.inv returns constantly the identity matrix for input on GPU

  #127281 closed May 28, 2024

• Tensor Parallel cannot work when tp mesh size is 1

  #127213 closed May 28, 2024

• Symbolic shapes unable to reason: Ne(Mod(u0*u2 + u1*u2, u0 + u1), 0)

  #125307 closed May 28, 2024

• fatal: not a git repository: '.git'

  #127188 closed May 28, 2024

• DISABLED test_transformerdecoder (__main__.TestNN)

  #126043 closed May 28, 2024

• DISABLED test_retrace_export_cond_simple_cuda_float32 (__main__.TestHOPCUDA)

  #123564 closed May 27, 2024

• DISABLED test_conv_empty_input_cpu_float32 (__main__.TestNNDeviceTypeCPU)

  #126091 closed May 27, 2024

• DISABLED test_activations_bfloat16_half_cpu_cpu_bfloat16 (__main__.TestNNDeviceTypeCPU)

  #126090 closed May 27, 2024

• DISABLED test_correctness_Adam_use_closure_False_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)

  #126076 closed May 27, 2024

• DISABLED test_correctness_Rprop_use_closure_False_cuda_float32 (__main__.CompiledOptimizerParityTestsCUDA)

  #126077 closed May 27, 2024

• [Distributed] gloo backend, barrier operation is even slower than broadcast

  #127179 closed May 27, 2024

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

  #124922 closed May 27, 2024

• Importing torch after TensorFlow results in std::runtime_error

  #101152 closed May 27, 2024

• Improve discoverability of meta function registration in documentation

  #126337 closed May 27, 2024

• DISABLED test_min_cut_partitioner_output_tensor_shape_tensor (__main__.TestPartitioning)

  #104326 closed May 27, 2024

• DISABLED test_unbacked_cat_backwards_cuda (__main__.TestInductorDynamicCUDA)

  #125019 closed May 27, 2024

• DISABLED test_full_tensor_sync (__main__.DTensorTest)

  #125366 closed May 27, 2024

• DISABLED test_depthwise_convolution (__main__.DistConvolutionOpsTest)

  #125565 closed May 27, 2024

• DISABLED test_pad_dynamic_cuda (__main__.TestInductorDynamicCUDA)

  #124276 closed May 27, 2024

• DISABLED test_all_gather_object (__main__.TestObjectCollectives)

  #125566 closed May 27, 2024

• DISABLED test_dtensor_device_mesh_device_conversion (__main__.DTensorMeshTest)

  #100126 closed May 27, 2024

• Dynamo benchmarks direct arg passing doesn't work

  #126543 closed May 26, 2024

• ERROR: Could not find a version that satisfies the requirement torch (from versions: none)

  #127181 closed May 26, 2024

• torch compile failure with bool inputs with CPP backend

  #126824 closed May 25, 2024

• CMake -DBUILD_PYTHON=ON doesn't make & install Python bindings.

  #126831 closed May 25, 2024

104 Issues opened by 79 people

• UNSTABLE inductor / cuda12.4-py3.10-gcc9-sm86 / test (dynamic_inductor_timm)

  #127680 opened Jun 1, 2024

• RuntimeError: `jit.freeze` fails to find externally assigned attributes

  #127679 opened Jun 1, 2024

• Key error in index_propagation when looking up dynamic shape vr

  #127677 opened Jun 1, 2024

• module 'torch.mps' has no attribute 'device'

  #127676 opened Jun 1, 2024

• Add line number to ` _warn_capture_scalar_outputs():`

  #127667 opened Jun 1, 2024

• Strange clamp assert error when building on Fedora 40/gcc 14 in IndexKernel.hip

  #127666 opened Jun 1, 2024

• Inductor generates unnecessary allocation + copy operations for custom ops with mutable inputs

  #127660 opened May 31, 2024

• DISABLED test_workspace_allocation_error (__main__.CudaGraphTreeTests)

  #127657 opened May 31, 2024

• Illegal memory access resulted from pointwise autotuning of a cat-like kernel

  #127652 opened May 31, 2024

• OneCycleLR Example

  #127649 opened May 31, 2024

• [export] Errors out when unflattening TorchTitan

  #127643 opened May 31, 2024

• c++ library written with a lot of errors

  #127638 opened May 31, 2024

• SyntaxError: unterminated string literal (detected at line 1) (<unknown>, line 1)

  #127637 opened May 31, 2024

• dynamo minifier test test_cpu_cuda_module_after_dynamo fail with nn module inlining.

  #127636 opened May 31, 2024

• Fix accuracy regression for cspdarknet53 or flakiness associated with cu121 (and potentially cu124)

  #127626 opened May 31, 2024

• Segfault, possibly due to recursion limit

  #127622 opened May 31, 2024

• custom_op API: better type anntation for Tuple

  #127621 opened May 31, 2024

• TORCH_LIBRARY breaks when passing (unexpanded) macro as namespace argument

  #127615 opened May 31, 2024

• [Bug] Data on CPUs Are Not Synchronized Before Subsequent Operations

  #127612 opened May 31, 2024

• Error in gradient of CTCLoss

  #127608 opened May 31, 2024

• ` to be a quantized tensor. Is this likely due to missing support for quantized `onnx::Pad`

  #127606 opened May 31, 2024

• RuntimeError: CUDA error: an illegal memory access was encountered with specific input shape and Conv1d

  #127605 opened May 31, 2024

• The data was supposed to be on the GPU, but was mistakenly placed on the CPU.

  #127603 opened May 31, 2024

• pip install torch==2.1.0 fails to be found

  #127591 opened May 31, 2024

• Equivalent index expression results in 50% perf difference in a triton kernel

  #127581 opened May 30, 2024

• AOTAutograd: allow input mutations in the bw that occur under no_grad

  #127572 opened May 30, 2024

• [export] Cannot mutate tensors with frozen storage

  #127571 opened May 30, 2024

• Nightly s390 manywheel-py3_8-cpu-s390x-build is failing

  #127568 opened May 30, 2024

• AOTAutograd: the partitioner should not move mutations from the backward into the forward.

  #127561 opened May 30, 2024

• [libtorch] duplicate symbol of vtable

  #127560 opened May 30, 2024

• Reserving this issue to test something, don't mind me

  #127555 opened May 30, 2024

• ddp graph splitting not happening when nn modules inlined.

  #127552 opened May 30, 2024

• TestNestedTensorSubclassCPU::test_chunk_cpu fails in debug mode

  #127546 opened May 30, 2024

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

  #127541 opened May 30, 2024

• DISABLED test_cusparse_multiple_threads_same_device (__main__.TestCuda)

  #127536 opened May 30, 2024

• Can we have a Y-Split module for torch.nn.Sequential?

  #127535 opened May 30, 2024

• Compilation from source fails (PYTORCH 1.13.1)

  #127534 opened May 30, 2024

• CUDA 12.5

  #127532 opened May 30, 2024

• Triton `kernel.run` segfaults when passed non-default stream

  #127531 opened May 30, 2024

• Run rstcheck on modified docstrings and docs as additional linter

  #127530 opened May 30, 2024

• Multiple unhandled exceptions in weights_only unpickler

  #127525 opened May 30, 2024

• SDPA memory efficient and flash attention kernels don't work with singleton dimensions

  #127523 opened May 30, 2024

• Modeling ViT does not support quantized models

  #127521 opened May 30, 2024

• [RFC] Load and register rendezvous backends dynamically as plugins at runtime

  #127519 opened May 30, 2024

• THPVariable_Check(list_elem) INTERNAL ASSERT FAILED

  #127518 opened May 30, 2024

• With pytest8.2.0 or later, test cases under test/distributed/ execute will meet issue "object has no attribute 'runTest'. Did you mean: 'run_test'"

  #127517 opened May 30, 2024

• [AOTI] Conv-BN folding on CPU not working anymore after benchmark script change in https://github.com/pytorch/pytorch/pull/123403

  #127513 opened May 30, 2024

• `assume_constant_result` does not work with method of `UnspecializedNNModuleVariable`

  #127509 opened May 30, 2024

• [BUG] Using custome backend for torch.compile give nothing outputs

  #127502 opened May 30, 2024

• Avoid Having to Register Op For ExternKernelChoice of Aten Refs

  #127500 opened May 30, 2024

• torch.compiler docstrings can be derived from _dynamo and _inductor docs

  #127497 opened May 30, 2024

• Multiplying a sparse CSR tensor by a strided vector nondeterministically fails on MacOS ARM64

  #127491 opened May 30, 2024

• investigate :torch._dynamo.exc.Unsupported: cache_size_limit reached when inlining nn module for test_do_not_skip_side_effects

  #127483 opened May 29, 2024

• Cannot fakeify a tensor who's .grad field is a tensor subclass

  #127470 opened May 29, 2024

• Add way to annotate a raw triton kernel using the triton kernel HOP so that it becomes make_fx-traceable

  #127452 opened May 29, 2024

• PyTorch version agnostic C++ extensions

  #127445 opened May 29, 2024

• address TODO: model is somehow not being freed when z3 is available

  #127444 opened May 29, 2024

• Ensure the Python custom ops tutorial actually runs in PyTorch 2.4

  #127443 opened May 29, 2024

• Release 2.3.1 validations checklist and cherry-picks

  #127441 opened May 29, 2024

• UNSTABLE inductor / cuda12.1-py3.10-gcc9-sm86 / test (dynamic_inductor_timm)

  #127438 opened May 29, 2024

• Crash in flatbuffer_loader during mobile model load from python API

  #127434 opened May 29, 2024

• index_select op not implemented on Vulkan backend

  #127422 opened May 29, 2024

• Fix torch._dynamo.exc.Unsupported: call_method GetAttrVariable(UnspecializedNNModuleVariable(CheckpointWrapper), __dict__) __contains__ [ConstantVariable()] {}

  #127416 opened May 29, 2024

• [inductor][cpu]abnormal performance improvement and accuracy drop for huggingface suit static/dynamic quantization

  #127402 opened May 29, 2024

• sys.maxsize special case doesn't work if you slightly offset the ranges

  #127396 opened May 29, 2024

• Unable to quantize resnet50 model with post training static quantization

  #127391 opened May 29, 2024

• torch=2.1, import torch; fails AttributeError: '_OpNamespace' 'aten' object has no attribute 'sym_constrain_range_for_size'

  #127388 opened May 29, 2024

• Backwards pass through Beta distribution rsample gives inf for 4 < alpha - 2**16 < 1040, beta = 3/2

  #127387 opened May 29, 2024

• 4 GPT2 can't run into `_scaled_dot_product_flash_attention_for_cpu` using AOTI due to export related change in https://github.com/pytorch/pytorch/pull/123732

  #127383 opened May 29, 2024

• Different tensor strides can result in surprisingly large discrepancies in Conv2d outputs

  #127375 opened May 29, 2024

• Torch.compile produces Exception: Please convert all Tensors to FakeTensors first or instantiate

  #127374 opened May 29, 2024

• DISABLED test_dtensor_op_db_bmm_cpu_float32 (__main__.TestDTensorOpsCPU)

  #127373 opened May 29, 2024

• [Feature Request] switch amx isa detection in onednn to cpuinfo

  #127368 opened May 29, 2024

• torch.export.export() throws out an error when dealing weighttying model.

  #127357 opened May 28, 2024

• `flake8: noqa` disables flake8 linter for the whole file and it's not obvious

  #127352 opened May 28, 2024

• Dynamo should prune non-live captured variables

  #127350 opened May 28, 2024

• ~PyTorch Docathon H1 2024!~

  #127345 opened May 28, 2024

• DISABLED test_comprehensive_fft_ifft_cuda_float64 (__main__.TestInductorOpInfoCUDA)

  #127344 opened May 28, 2024

• DISABLED test_arange2_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests)

  #127343 opened May 28, 2024

• Add option for custom ops to automatically get a FakeTensor kernel (during static shapes)

  #127337 opened May 28, 2024

• Inductor fails at assert self.symbol_to_source.get(expr)

  #127328 opened May 28, 2024

• torch.compile reorder_for_compute_comm_overlap sink_waits pass does not work

  #127324 opened May 28, 2024

• torch.compile (inductor) bug random signed number generation

  #127310 opened May 28, 2024

• I don’t know if it’s a problem with cuda or pytorch

  #127299 opened May 28, 2024

• Fused AdamW maybe should accept lr_dict directly?

  #127284 opened May 28, 2024

• [BUG][JIT] `torch.jit.script` is not compatible with `DeprecationWarning` and `FutureWarning`

  #127283 opened May 28, 2024

• torch.fx.symbolic_trace doesn't support many Callable types

  #127282 opened May 28, 2024

• Python sometimes crashes inexplicably, with "/var/log/kernel. log" displaying index errors

  #127275 opened May 28, 2024

• DISABLED test__int_mm_k_32_n_16_use_transpose_a_True_use_transpose_b_False_cuda (__main__.TestLinalgCUDA)

  #127273 opened May 28, 2024

• ImportError: cannot import name 'triton_key' from 'triton.compiler.compiler' when using triton 2.3.x

  #127271 opened May 28, 2024

• Not supporting higher protobuf versions

  #127270 opened May 28, 2024

• make_graphed_callables don't work with FSDP at all even on a simple network

  #127225 opened May 27, 2024

• RuntimeError: grid_sampler_2d_cpu not implemented for Half

  #127224 opened May 27, 2024

• [BUG] torch.linalg.lstsq returning wrong result.

  #127223 opened May 27, 2024

• Migrate `CalculateSmallVectorDefaultInlinedElements` to constexpr Function for SmallVector

  #127222 opened May 27, 2024

• lacking checking for ConvTranspose's parameters when running with GPUs

  #127221 opened May 27, 2024

• Comparing dynamic shapes fails with KeyError

  #127217 opened May 27, 2024

• No way to config profiling scope in torch.autograd.profile

  #127215 opened May 27, 2024

• DISABLED test_large_mmaped_weights_non_abi_compatible_cuda (__main__.AOTInductorTestNonABICompatibleCuda)

  #127202 opened May 27, 2024

• Tensors of the same index must be on the same device and the same dtype except step tensors that can be CPU and float32 notwithstanding.

  #127197 opened May 26, 2024

• torch.topk results differ on CPU and CUDA

  #127196 opened May 26, 2024

• Should `torch.Size` convert np.ndarrays to lists of ints?

  #127194 opened May 26, 2024

• Unify async_save and sync_save in state_dict_saver from distributed checkpointing

  #127191 opened May 26, 2024

• `store_param_remainders` from Apex DistributedFusedAdam

  #127189 opened May 26, 2024

383 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.

• Complete revamp of float/promotion sympy handling

  #126905 commented on May 31, 2024 • 78 new comments

• [CI] Add freezing for cpu inductor accuracy test in inductor CI

  #124715 commented on May 31, 2024 • 38 new comments

• [Autograd] Cond Higher-Order Operation

  #126911 commented on Jun 1, 2024 • 28 new comments

• [quant][pt2e][quantizer] Support `set_module_name_qconfig` in X86InductorQuantizer

  #126044 commented on May 31, 2024 • 28 new comments

• [Inductor] support masked vectorization for the tail_loop

  #126526 commented on May 30, 2024 • 24 new comments

• [executorch hash update] update the pinned executorch hash

  #123043 commented on Jun 1, 2024 • 21 new comments

• [vision hash update] update the pinned vision hash

  #125806 commented on Jun 1, 2024 • 21 new comments

• ROCm 6.x appears Cannot find CO in the bundle libhipblaslt.so for ISA

  #127169 commented on May 30, 2024 • 18 new comments

• [CI] add xpu test in periodic workflow

  #126410 commented on Jun 1, 2024 • 17 new comments

• Memory Tracker for tracking Module wise memory

  #124688 commented on Jun 1, 2024 • 16 new comments

• TorchInductor CPU Performance Dashboard

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

• [BE]: Update cudnn to 9.1.0.70

  #123475 commented on Jun 1, 2024 • 15 new comments

• [aoti] Add initial custom op support

  #127034 commented on May 31, 2024 • 15 new comments

• [dtensor] local_map UX change: keep func signature and be compatible with Tensor input

  #126924 commented on Jun 1, 2024 • 12 new comments

• [inductor] online softmax

  #127011 commented on May 28, 2024 • 12 new comments

• [inductor][cpp] bf16/fp16 gemm template computed with fp32 w/o epilogue fusion

  #126068 commented on Jun 1, 2024 • 12 new comments

• Opt model save and load

  #126374 commented on Jun 1, 2024 • 11 new comments

• [RFC] Add support for device extension autoloading

  #127074 commented on Jun 1, 2024 • 11 new comments

• torch.export.export() fails with dynamic shapes when more than one shape is dynamic

  #126127 commented on May 31, 2024 • 11 new comments

• Re-implement pin_memory to be device-agnostic by leveraging the Accelerator concept

  #126376 commented on May 31, 2024 • 10 new comments

• General MPS op coverage tracking issue

  #77764 commented on Jun 1, 2024 • 10 new comments

• T5 -small Dynamic quantization in graviton3

  #127062 commented on May 30, 2024 • 9 new comments

• Enable UFMT on test_shape_ops.py test_show_pickle.py test_sort_and_select.py

  #127165 commented on May 31, 2024 • 9 new comments

• fix constant folding with buffer mutation

  #123909 commented on May 28, 2024 • 9 new comments

• Beef up the allow_in_graph docs

  #127117 commented on Jun 1, 2024 • 9 new comments

• [torchbind] always fakify script object by default in non-strict export

  #127116 commented on Jun 1, 2024 • 9 new comments

• [Inductor][CPP] Add Min/Max with VecMask

  #126841 commented on May 27, 2024 • 8 new comments

• skip hf_T5_generate in dynamic shape test

  #121129 commented on May 31, 2024 • 8 new comments

• Set simdlen based on ATEN_CPU_CAPABILITY

  #123514 commented on May 30, 2024 • 8 new comments

• First version of AOTAutogradCache

  #126791 commented on May 31, 2024 • 8 new comments

• enable device index check for all device types

  #126767 commented on May 31, 2024 • 8 new comments

• LambdaLR has incorrect multiplicative behavior when using torch.tensor LR

  #126854 commented on May 30, 2024 • 8 new comments

• Extend Fake Tensor Caching to Symints

  #126411 commented on May 28, 2024 • 8 new comments

• CUDNN not detected in official nightly devel image. We need to use `12.4.1` upstream images

  #126005 commented on May 31, 2024 • 8 new comments

• Update _dedup_save_plans.py

  #126569 commented on Jun 1, 2024 • 8 new comments

• [ROCm] TunableOp improvements

  #124362 commented on Jun 1, 2024 • 7 new comments

• Enable UFMT format on test/test_nn.py

  #126663 commented on May 31, 2024 • 7 new comments

• [dtensor] implement scatter op with simple replication

  #126713 commented on Jun 1, 2024 • 7 new comments

• Fix decorators skipping NCCL tests

  #122397 commented on May 29, 2024 • 7 new comments

• sdp::SDPBackend::flash_attention support PrivateUse1

  #126392 commented on May 31, 2024 • 7 new comments

• [inductor] add cpp builder code. (take 2)

  #125849 commented on Jun 1, 2024 • 7 new comments

• [feature request] `quantized::linear_dynamic` on CUDA/eager, and other quantized and low-level int8 operators (matmul, gemm etc) on CUDA + integrate LLM.int8 + integrate ZeroQuant?

  #69364 commented on May 30, 2024 • 7 new comments

• [cuDNN][SDPA] Remove `TORCH_CUDNN_SDPA_ENABLED=1`, enable cuDNN SDPA by default on H100 and 2nd on other archs >= sm80

  #125343 commented on Jun 1, 2024 • 7 new comments

• [dynamo][nn-modules] Trace through nn.Module dunder methods for UnspecializedNNModule

  #126578 commented on Jun 1, 2024 • 7 new comments

• Support aten operations with out tensor

  #124926 commented on May 31, 2024 • 7 new comments

• Make TraceUtils.h to be device-agnostic

  #126969 commented on Jun 1, 2024 • 6 new comments

• Move the build of AOTriton to base ROCM docker image.

  #127012 commented on Jun 1, 2024 • 6 new comments

• extend `nonzero` to int64

  #125850 commented on May 31, 2024 • 6 new comments

• [CUDNN] Remove defunct cuDNN V8 API build flag

  #120006 commented on Jun 1, 2024 • 6 new comments

• [WIP] Warn on future divergent behavior for conditional views

  #126129 commented on Jun 1, 2024 • 6 new comments

• Feature: Implement support for `cudnn_batch_norm_out` kernel to replace the autogen approach.

  #123020 commented on May 29, 2024 • 6 new comments

• [Submodule] Remove glog dependency

  #126768 commented on Jun 1, 2024 • 6 new comments

• PT2 Inductor ComboKernels

  #124969 commented on Jun 1, 2024 • 6 new comments

• [compiled autograd][cudagraphs] Inputs runtime wrapper to move cpu scalars to cuda

  #125382 commented on Jun 1, 2024 • 5 new comments

• Introduce Inductor passes to micro-pipeline all-gather-matmul and matmul-reduce-scatter in certain cases

  #126598 commented on Jun 1, 2024 • 5 new comments

• CUDA 12.4 CI Inductor Issues

  #126692 commented on Jun 1, 2024 • 5 new comments

• [DTensor] `clip_grad_norm_` follow-ups

  #121020 commented on May 31, 2024 • 5 new comments

• Compiling extension on pytorch nighlty image is broken

  #125879 commented on Jun 1, 2024 • 5 new comments

• PyTorch's Execution Trace Observer Producing Invalid JSON

  #126703 commented on Jun 1, 2024 • 5 new comments

• refine fp32 precision api

  #125888 commented on May 30, 2024 • 5 new comments

• [Inductor][CPP] Enable Local Buffer for Outer loop fusion

  #126967 commented on May 31, 2024 • 5 new comments

• [ONNX] Skip assertion nodes

  #126889 commented on May 29, 2024 • 5 new comments

• [BE]: Update NCCL submodule to 2.21.5

  #124014 commented on May 31, 2024 • 5 new comments

• [Split Build] Test split build in CI

  #126699 commented on Jun 1, 2024 • 5 new comments

• Pytorch 2.0 installation tutorial does not work under Macbook

  #96073 commented on May 30, 2024 • 4 new comments

• [RFC] Per-Parameter-Sharding FSDP

  #114299 commented on May 26, 2024 • 4 new comments

• [dynamo] Automatically convert loop bodies to function calls

  #113538 commented on May 31, 2024 • 4 new comments

• MPS device can train or evaluate models producing unacceptable output due to "fast math" optimization

  #84936 commented on May 31, 2024 • 4 new comments

• PyTorch DataLoader improvements for Iterable Dataset

  #127072 commented on Jun 1, 2024 • 4 new comments

• Refresh OpOverloadPacket if a new OpOverload gets added

  #126863 commented on May 29, 2024 • 4 new comments

• dynamo doesn't support `__torch_function__` on non-tensor classes

  #127174 commented on May 31, 2024 • 4 new comments

• IPEX as TorchDynamo Backend Performance Dashboard

  #101273 commented on May 27, 2024 • 4 new comments

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

  #126050 commented on May 31, 2024 • 4 new comments

• add core tag to linalg_vector_norm

  #125789 commented on May 31, 2024 • 4 new comments

• [torchbind] support torch.compile with aot_eager backend

  #127114 commented on May 31, 2024 • 4 new comments

• Setting static arguments in `torch.compile()`

  #121299 commented on May 29, 2024 • 4 new comments

• `torch.compiler.allow_in_graph` does not create a `call_module` op in fx.Graph in torch 2.3.0

  #126566 commented on May 29, 2024 • 4 new comments

• Add aten._unsafe_masked_index

  #116491 commented on May 30, 2024 • 4 new comments

• [dtensor] standardize multi mesh-dim strategy with utils

  #126712 commented on Jun 1, 2024 • 3 new comments

• [xla hash update] update the pinned xla hash

  #126672 commented on May 27, 2024 • 3 new comments

• fix multiprocessing.spawn doc issue

  #126902 commented on May 28, 2024 • 3 new comments

• Triton Matrix Multiplication example invalid results (return zeros) on Volta

  #127157 commented on May 30, 2024 • 3 new comments

• Move MKLDNN Specific IR to Separate File

  #126504 commented on May 27, 2024 • 3 new comments

• [inductor] `aten.index_put_` runtime shape mismatch on H100 but not on A100

  #126614 commented on May 30, 2024 • 3 new comments

• make_graphed_callables fails when Module or function does not contain parameters

  #124582 commented on May 28, 2024 • 3 new comments

• [dynamo][numpy] Add unsigned integer dtypes

  #125717 commented on May 31, 2024 • 3 new comments

• Invalidate StorageImpl instances when tensor is overwritten with cudagraphs

  #125264 commented on Jun 1, 2024 • 3 new comments

• delete inductor config.trace.compile_profile

  #127143 commented on May 29, 2024 • 3 new comments

• Add unfold support for MaskedTensor

  #125262 commented on May 30, 2024 • 3 new comments

• add uuid in cudaDeviceProperties

  #125083 commented on May 29, 2024 • 3 new comments

• move label_to_label.yml to an issue?

  #126714 commented on May 29, 2024 • 3 new comments

• [dynamo] Trace through invalid bool tensor operations properly

  #127003 commented on Jun 1, 2024 • 3 new comments

• [Dynamo] Log backward graph compilation metrics

  #126629 commented on May 31, 2024 • 3 new comments

• [NJT] Allow construction of NJT within graph using offsets from inputs

  #124624 commented on May 28, 2024 • 3 new comments

• [v2.3.1] Release Tracker

  #125425 commented on May 28, 2024 • 3 new comments

• ROCm loses some supported GPUs by requiring hipblaslt

  #119081 commented on May 31, 2024 • 3 new comments

• `@functools.wraps` graph breaks in many cases where we should be able to handle it

  #123365 commented on May 28, 2024 • 3 new comments

• Made some minor improvements to flexattention perf + added more autotune configs

  #126811 commented on May 28, 2024 • 3 new comments

• [Inductor][CPP] Add ne with VecMask

  #126940 commented on May 27, 2024 • 3 new comments

• s390x: build s390x binaries on each pull request

  #125399 commented on May 29, 2024 • 2 new comments

• Allow multiple cudagraph recordings per compiled graph

  #126822 commented on Jun 1, 2024 • 2 new comments

• inplace parameter in dropouts should function as expected regardless of the value of training(or train) paramter

  #126755 commented on May 31, 2024 • 2 new comments

• [XPU] Add xpu support of `make triton`

  #126513 commented on May 29, 2024 • 2 new comments

• [triton hash update] update the pinned triton hash

  #115529 commented on May 27, 2024 • 2 new comments

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

  #124480 commented on May 31, 2024 • 2 new comments

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

  #120219 commented on Jun 1, 2024 • 2 new comments

• [Bazel] Build shared libraries to reduce total binary size

  #126031 commented on May 30, 2024 • 2 new comments

• [nn.functional] Set default value of Optional params of batch_norm to None

  #122717 commented on May 28, 2024 • 2 new comments

• [inductor][cpu]detectron2_maskrcnn_r_101_fpn detectron2_maskrcnn_r_50_c4 accuracy check failed when freezing flag is on.

  #127073 commented on Jun 1, 2024 • 2 new comments

• Disable atomic_add fallback for cpu in index_put lowering

  #122873 commented on May 27, 2024 • 2 new comments

• Add Ability to Skip Hipification When Building CUDA Extension

  #127015 commented on May 28, 2024 • 2 new comments

• [SDPA/memeff] Backport changes from xFormers to PT

  #127090 commented on May 31, 2024 • 2 new comments

• xpu: implement grid_sample op for XPU (fallback to CPU not possible for fp16 and bf16)

  #127002 commented on May 30, 2024 • 2 new comments

• ```FlopCounterMode``` returns 0 when inference mode is on during forwardpropagation.

  #126268 commented on May 30, 2024 • 2 new comments

• [PT2E Quantization] `prepare_pt2e` produces inconsistent data types for primitive int

  #127076 commented on May 30, 2024 • 2 new comments

• Move loop ordering after fusion

  #126255 commented on May 30, 2024 • 2 new comments

• BUG: `torch.pow(-0.0)` wrongly returns `-0.0` (should be +0.0)

  #127163 commented on May 28, 2024 • 2 new comments

• libtorch mimalloc Debug Warning with WinError

  #125840 commented on May 30, 2024 • 2 new comments

• [VIT Model] [perf Degradation] [X86] [ARM] torch.compile + weight prepacking results in perf degradation for VIT Transformer model

  #126391 commented on May 28, 2024 • 2 new comments

• [feature][cudagraph] API to clear a bad recording

  #127147 commented on May 29, 2024 • 2 new comments

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

  #123503 commented on May 30, 2024 • 2 new comments

• NCCL watchdog thread terminated with exception

  #113128 commented on May 30, 2024 • 2 new comments

• pytorch is built without _GLIBCXX_USE_CXX11_ABI and can cause std::regex crashes (probably)

  #50779 commented on May 26, 2024 • 2 new comments

• fixed torch.where and torch.nonzero issue

  #123024 commented on May 28, 2024 • 1 new comment

• Parts of https://download.pytorch.org not updated anymore, broken links and missing versions

  #121837 commented on May 29, 2024 • 1 new comment

• Register lowering for `_foreach_norm.Scalar`

  #123051 commented on May 30, 2024 • 1 new comment

• [BASE_MODULE] removing base_module for glow/fb/test:test_utils

  #123092 commented on May 31, 2024 • 1 new comment

• [TP] Add tests for wildcard support

  #123101 commented on May 31, 2024 • 1 new comment

• [Pytorch] Replace nn.module call to `train()` for executorch models (#757)

  #122532 commented on Jun 1, 2024 • 1 new comment

• add testing support for Vectorized<Half>

  #123132 commented on Jun 1, 2024 • 1 new comment

• Add quantized.linear_unpacked_dynamic_fp16

  #122509 commented on May 26, 2024 • 1 new comment

• [WIP] Add comm fusion pass

  #122505 commented on May 26, 2024 • 1 new comment

• dynamo_export fails for for NVIDIA NeMo framework classes decorated with wrapt package

  #125462 commented on May 29, 2024 • 1 new comment

• torch._inductor.config.max_autotune_gemm_backends = "TRITON" crashes with Convolution layer

  #125728 commented on May 29, 2024 • 1 new comment

• Add DtypeContext

  #122481 commented on May 28, 2024 • 1 new comment

• Torch 2.1 compile + FSDP (mixed precision) + LlamaForCausalLM: `RuntimeError: attempting to assign a gradient with dtype 'c10::BFloat16' to a tensor with dtype 'float'.`

  #111317 commented on May 29, 2024 • 1 new comment

• [export] Restore original placeholder names

  #122452 commented on May 27, 2024 • 1 new comment

• segfault when registering op with mismatched schema

  #127102 commented on May 29, 2024 • 1 new comment

• Ensure profiler record function callback always increments "record function id".

  #123755 commented on May 27, 2024 • 1 new comment

• [TEST] assume no aligned inputs

  #122159 commented on May 28, 2024 • 1 new comment

• Fix edge cases for gather in inductor

  #126893 commented on May 29, 2024 • 1 new comment

• [example] Example of properly optimizing mm=>layernorm=>gelu

  #122687 commented on May 25, 2024 • 1 new comment

• Modified pointless_convert pass to only apply if it's removing truly useless conversion

  #122689 commented on May 27, 2024 • 1 new comment

• [fx] Preserve Fx graph node order in partitioner across runs

  #122696 commented on May 27, 2024 • 1 new comment

• [wip][export] Add symint input support

  #122712 commented on May 25, 2024 • 1 new comment

• [POC][FSDP2] Ran parameter post acc grad hooks manually

  #122719 commented on May 26, 2024 • 1 new comment

• Allow flight recorder to call elasped_time again.

  #122731 commented on May 27, 2024 • 1 new comment

• [DRAFT] add int4_mm support for ARM using SIMDe

  #122764 commented on May 26, 2024 • 1 new comment

• add testing support for Vectorized<Half>

  #122798 commented on May 27, 2024 • 1 new comment

• [Inductor][Optimus] Fix group batch fusion for AFOC

  #122839 commented on May 27, 2024 • 1 new comment

• [ATen-Vulkan][EZ] Small fixes: fix gpu size calculation and Half scalartype ctype mapping

  #122861 commented on Jun 1, 2024 • 1 new comment

• [dynamo/inductor] guard on whether a tensor is a view, assume non-views are aligned

  #122867 commented on May 27, 2024 • 1 new comment

• Libtorch build for ROCM error: “aten/src/THH” not exist

  #126640 commented on May 29, 2024 • 1 new comment

• Optimize upsample_bilinear2d on channels last format

  #122879 commented on May 31, 2024 • 1 new comment

• [inductor][Autotune] Add matrix_instr_nonkdim to triton_meta (#122852)

  #122906 commented on May 27, 2024 • 1 new comment

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

  #122667 commented on May 28, 2024 • 1 new comment

• [ROCm] skip some SDPA tests

  #122914 commented on May 29, 2024 • 1 new comment

• [AOTInductor] Support quantized linear on CPU with fbgemm

  #122939 commented on May 30, 2024 • 1 new comment

• Fix fuse linear bn for affine=False

  #122537 commented on Jun 1, 2024 • 1 new comment

• [dynamo] Dont run Dynamo on assert* functions

  #122993 commented on May 29, 2024 • 1 new comment

• [dynamo] Detect functionalized functions with HOPs with mutations at tracing time

  #126988 commented on May 29, 2024 • 1 new comment

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

  #124497 commented on May 27, 2024 • 1 new comment

• [checkpoint] Clean up selective activation checkpoint and make public

  #125795 commented on May 31, 2024 • 1 new comment

• JAX + PyTorch produces `OMP: Error #13: Assertion failure at kmp_affinity.cpp(532)`

  #97635 commented on May 27, 2024 • 1 new comment

• Run CompositeImplicit on torch fns during dynamo tracing

  #125808 commented on Jun 1, 2024 • 1 new comment

• Use sleef on macOS Apple silicon by default

  #126509 commented on May 28, 2024 • 1 new comment

• _convert_input_to_fake doesn't work when there are no inputs

  #126770 commented on May 27, 2024 • 1 new comment

• [NT] Implementing Multi-Head Attention with NestedTensors

  #125214 commented on May 27, 2024 • 1 new comment

• ❓Different results between normal batching and `vmap` while using lower precision (e.g., bfloat16)

  #125534 commented on May 27, 2024 • 1 new comment

• Support for stereo audio data in from torch.utils.tensorboard.SummaryWriter

  #56360 commented on May 27, 2024 • 1 new comment

• Remove redundant device guard in Resize.h

  #126498 commented on May 31, 2024 • 1 new comment

• Box constraints for optimizers

  #22281 commented on May 27, 2024 • 1 new comment

• NVML_SUCCESS == r INTERNAL ASSERT FAILED at "../c10/cuda/CUDACachingAllocator.cpp":830, please report a bug to PyTorch.

  #122068 commented on May 27, 2024 • 1 new comment

• TorchDynamo ONNX Export does not work as expected with masking (ScatterElements)

  #126856 commented on May 27, 2024 • 1 new comment

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

  #123973 commented on May 26, 2024 • 1 new comment

• Update triton pin to improve throughput w/ assert

  #126098 commented on May 30, 2024 • 1 new comment

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

  #124565 commented on May 26, 2024 • 1 new comment

• Specify version

  #105460 commented on May 26, 2024 • 1 new comment

• [inductor][codegen] Codegen constexpr globals and constexpr annotated globals correctly.

  #126195 commented on May 30, 2024 • 1 new comment

• [aoti] Add support for more custom op input/output types

  #126215 commented on May 28, 2024 • 1 new comment

• [Traceable FSDP][Compiled Autograd] Add queue_callback() support

  #126366 commented on May 29, 2024 • 1 new comment

• Added memory budget to partitioner

  #126320 commented on Jun 1, 2024 • 1 new comment

• Optimize Vectorized<float> exp() with neon simd instructions

  #126612 commented on May 29, 2024 • 1 new comment

• [torchbind] remove test cases that don't fakify script objects

  #127113 commented on May 31, 2024 • 1 new comment

• DISABLED test_large_sampler_indices (__main__.TestDataLoaderPersistentWorkers)

  #117784 commented on May 28, 2024 • 1 new comment

• [dynamo] Handle inplace op aliasing errors

  #126474 commented on May 28, 2024 • 1 new comment

• torch.compile crash - Aborted exit code 134

  #125804 commented on May 28, 2024 • 1 new comment

• Many tests under test/distributed/elastic not running in OSS CI

  #124317 commented on May 28, 2024 • 1 new comment

• Tracker: Slow gradcheck failures possibly indicating incorrect gradients

  #80411 commented on May 28, 2024 • 1 new comment

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

  #127176 commented on May 28, 2024 • 1 new comment

• `torch.einsum` docs don't mention that `opt_einsum` must be installed separately

  #127109 commented on May 28, 2024 • 1 new comment

• Modularize aten parameter parser and checker

  #125308 commented on May 31, 2024 • 1 new comment

• [dynamo] Improve support for hasattr for custom __getattr__

  #125340 commented on Jun 1, 2024 • 1 new comment

• A huge difference between the results of torch.round() on the GPU compared to its results on the CPU and other DL libraries

  #126975 commented on May 28, 2024 • 1 new comment

• torch_dispatch mode silent incorrectness with torch.compile

  #115653 commented on May 28, 2024 • 1 new comment

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

  #122169 commented on May 28, 2024 • 1 new comment

• [inductor][cpp] support bf16/fp16 gemm template epilogue fusion

  #126545 commented on Jun 1, 2024 • 1 new comment

• add wait_tensor to ops to skip in DCE pass

  #126541 commented on May 31, 2024 • 1 new comment

• aten.bernoulli.p is missing in core aten IR opset but does not get decomposed

  #105519 commented on May 28, 2024 • 1 new comment

• CUDA Extension Bug for Pytorch 2.2.0

  #118842 commented on May 27, 2024 • 1 new comment

• [1/N] Dynamic Shape: To enable export.aot_compile to switch between static shape and dynamic shape

  #126517 commented on May 31, 2024 • 1 new comment

• Segmentation Fault in `torch.nn.Conv1d` starting from torch 2.2.0

  #121222 commented on May 27, 2024 • 1 new comment

• dynamo test: remove the cuda hardcoding and detect device dynamically

  #125761 commented on May 31, 2024 • 1 new comment

• QAT using pytorch-quantization cause accuracy loss after exporting to onnx

  #120166 commented on May 31, 2024 • 1 new comment

• [inductor][cpu]tnt_s_patch16_224 and functorch_dp_cifar10 fP32 static default wrapper performance regression

  #119178 commented on May 30, 2024 • 1 new comment

• [MPS] Improve the performance of torch.linear()

  #91737 commented on May 31, 2024 • 1 new comment

• Why is AvgPool2D taking longer than Conv2D for the same input?

  #93188 commented on May 31, 2024 • 1 new comment

• [inductor][cpu]speech_transformer AMP single/multiple thread static/dynamic shape CPP/default wrapper performance regression in 2024-05-12 nightly release

  #126274 commented on May 30, 2024 • 1 new comment

• UNSTABLE inductor / linux-jammy-cpu-py3.8-gcc11-inductor / test (inductor_torchbench_cpu_smoketest_perf)

  #126993 commented on May 30, 2024 • 1 new comment

• RuntimeError: [1] is setting up NCCL communicator and retreiving ncclUniqueId from [0] via c10d key-value store by key '0', but store->get('0') got error: Timeout waiting for key: default_pg/0/0 after 1800000 ms

  #82091 commented on May 30, 2024 • 1 new comment

• [Split Build] Test split build in pull CI workflow

  #126813 commented on Jun 1, 2024 • 1 new comment

• Cudnn 9.1.1 is out!

  #119400 commented on May 31, 2024 • 1 new comment

• [DCP] `set_model_state_dict` errors on compiled module with non-persistent buffer

  #122792 commented on May 31, 2024 • 1 new comment

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

  #124850 commented on May 31, 2024 • 1 new comment

• Support backward hook optimizers in FSDP

  #98419 commented on May 30, 2024 • 1 new comment

• Investigate torch.compile Windows support.

  #122094 commented on May 31, 2024 • 1 new comment

• Previous version not found

  #107611 commented on May 31, 2024 • 1 new comment

• FSDP does not work on GLOO backend

  #74041 commented on May 31, 2024 • 1 new comment

• [bug] Dynamo graph break when using pyton module `heapq` (manipulates with `list`s), although succeeds when placing `heapq.py` near the test script

  #106885 commented on May 31, 2024 • 1 new comment

• torch.compile error

  #124044 commented on May 31, 2024 • 1 new comment

• xpu: provide a way to debug explicit CPU fallback

  #126488 commented on May 30, 2024 • 1 new comment

• xpu: python hangs on exit after check for xpu on multi-dev system

  #126259 commented on May 30, 2024 • 1 new comment

• torch.compile generates wrong code on CPU and compiled code replaces original function

  #126848 commented on May 30, 2024 • 1 new comment

• tensordict functional calls with nn.Module silently gives the wrong (non-functional) result

  #127173 commented on May 30, 2024 • 1 new comment

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

  #123097 commented on May 30, 2024 • 1 new comment

• Pytorch dataloader not loading first-available data with multiple workers

  #105203 commented on May 30, 2024 • 1 new comment

• [ONNX] view(dtype=dtype) is not supported by both onnx.export and onnx.dynamo_export

  #126921 commented on May 30, 2024 • 1 new comment

• Lowering after pointwise cat can lead to uncontiguous memory accesses

  #124002 commented on May 30, 2024 • 1 new comment

• The doc of `diagonal()` doesn't have such an explanation

  #126827 commented on May 30, 2024 • 1 new comment

• The dynamic compilation does not handle composite shapes with multiple dimensions very effectively

  #127162 commented on May 30, 2024 • 1 new comment

• MultiheadAttention returns NaNs when need_weights=False for long sequences with a mask that ignores old tokens

  #127055 commented on May 30, 2024 • 1 new comment

• [RFC] Enable PyTorch XPU on Native Windows on Intel GPUs

  #126719 commented on May 30, 2024 • 1 new comment

• [dynamo] Dynamic slicing on data-dependent value is not supported (regression from legacy ONNX)

  #127154 commented on May 31, 2024 • 1 new comment

• Multiple Inputs/Outputs with torch.onnx.dynamo_export

  #127128 commented on May 31, 2024 • 1 new comment

• [ONNX] pad_sequence() is not exportable, with neither legacy onnx.export nor with dynamo_export

  #127153 commented on May 31, 2024 • 1 new comment

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

  #127167 commented on May 31, 2024 • 1 new comment

• [WIP] fix calls to `super().__getattr__` nn.Module in dynamo

  #126875 commented on May 31, 2024 • 1 new comment

• UserWarning: Plan failed with a cudnnException: CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR

  #126605 commented on May 30, 2024 • 1 new comment

• [torch.compile]: Enhanced Error Reporting and Performance Canary Mode

  #126644 commented on May 31, 2024 • 1 new comment

• Add parameter "half_pixel_center =False" to the Bilinear function

  #48389 commented on May 30, 2024 • 1 new comment

• FSDP `use_orig_params` + full sharding results in missing parameters in gathered state dict

  #126310 commented on May 31, 2024 • 1 new comment

• DISABLED test_comprehensive_special_bessel_y1_cuda_int32 (__main__.TestInductorOpInfoCUDA)

  #127080 commented on May 29, 2024 • 1 new comment

• [fx] Preserve Fx graph node order in partitioner across runs

  #115621 commented on May 29, 2024 • 1 new comment

• test_python_shard.bat -> .sh

  #117194 commented on May 31, 2024 • 1 new comment

• RuntimeError in torch.istft with center=False: Window Overlap Add Issue

  #118507 commented on May 29, 2024 • 1 new comment

• Make watchdog sleep interval auto-adaptive

  #118293 commented on May 26, 2024 • 1 new comment

• [optim] Add support for complex tensors in SparseAdam

  #118653 commented on Jun 1, 2024 • 1 new comment

• Fix jagged NT softmax semantics

  #119459 commented on May 31, 2024 • 1 new comment

• Add decompositions for copy variants of view ops

  #119889 commented on May 30, 2024 • 1 new comment

• [Caffe2]Remove Caffe2 scripts and benchmarks

  #126747 commented on May 30, 2024 • 1 new comment

• Change ATEN generator argument type to const std::optional<Generator>&

  #120076 commented on May 25, 2024 • 1 new comment

• Delete lazy ddp optimizer

  #120727 commented on May 26, 2024 • 1 new comment

• turned on matrix-multiplication => matrix-vector multiplication always on if reduction-dim is contiguous

  #120954 commented on May 27, 2024 • 1 new comment

• DISABLED test_circular_dependencies (__main__.TestImports)

  #110040 commented on May 29, 2024 • 1 new comment

• fix graph deepcopy to also copy `_tracer_extras`

  #121171 commented on May 28, 2024 • 1 new comment

• [Inductor] Add prologue fusion

  #121211 commented on Jun 1, 2024 • 1 new comment

• [NJT] Actually inline NT torch function during dynamo

  #121445 commented on Jun 1, 2024 • 1 new comment

• [DDP] Bucket handling: make first bucket size equal to bucket_cap_mb if it was set

  #121640 commented on May 28, 2024 • 1 new comment

• [no-ci] test pr

  #122006 commented on May 28, 2024 • 1 new comment

• [feature request] torch.kthvalue to support a new argument largest

  #29398 commented on May 30, 2024 • 1 new comment

• [dynamo] Fake tensor impl for Tensor.add_ not checking for errors

  #127049 commented on Jun 1, 2024 • 1 new comment

• RuntimeError: NVML_SUCCESS == r INTERNAL ASSERT FAILED at "/opt/pytorch/pytorch/c10/cuda/CUDACachingAllocator.cpp":830, please report a bug to PyTorch.

  #123834 commented on May 29, 2024 • 1 new comment

• Support tracing through _get_current_dispatch_mode_stack

  #126789 commented on May 28, 2024 • 1 new comment

• Compile with non-default mode + triton kernel fails

  #126864 commented on May 29, 2024 • 1 new comment

• [benchmark] Rename the count field FunctionCount

  #105471 commented on May 29, 2024 • 1 new comment

• Switch to CXX11 ABI

  #126778 commented on May 31, 2024 • 0 new comments

• some logging changes

  #127005 commented on May 29, 2024 • 0 new comments

• [Split Build] Use single command to build both wheels

  #126590 commented on May 29, 2024 • 0 new comments

• [halide-backend] Initial implementation of HalideKernel and HalideScheduling

  #126417 commented on Jun 1, 2024 • 0 new comments

• [do-not-review][aot] explicitly pass number of static inputs (buffer/params) to backend

  #127006 commented on Jun 1, 2024 • 0 new comments

• Allow overriding per-dim group options via _MeshEnv.set_dim_group_options

  #126599 commented on Jun 1, 2024 • 0 new comments

• [Inductor] support masked vectorization for the tail_loop

  #127166 commented on May 26, 2024 • 0 new comments

• Enable LSAN

  #127171 commented on May 29, 2024 • 0 new comments

• [Split Build] Test non-linux manywheel builds

  #126929 commented on May 29, 2024 • 0 new comments

• [Split Build][WIP] Change global env variables to cmake args

  #126930 commented on May 29, 2024 • 0 new comments

• [inductor] [for benchmarking] Add a subprocess-based parallel compile

  #127088 commented on May 27, 2024 • 0 new comments

• Enable optimized dynamic quantization on aarch64

  #126687 commented on May 30, 2024 • 0 new comments

• Set NO_DELAY flag in TCPStore

  #127042 commented on Jun 1, 2024 • 0 new comments

• [2/N] Dynamic Shape: Enable dynamic shape support for aoti_eager

  #126883 commented on May 31, 2024 • 0 new comments

• [FSDP2] Added `share_comm_ctx`

  #127032 commented on May 28, 2024 • 0 new comments

• avoid size() dispatch on FunctionalTensor

  #126784 commented on May 31, 2024 • 0 new comments

• [inductor] Enable subprocess-based parallel compile as the default

  #126817 commented on May 29, 2024 • 0 new comments

• Collect static parameter metadata in aot

  #126820 commented on May 30, 2024 • 0 new comments

• Remove unused arg to GraphLowering

  #126821 commented on May 30, 2024 • 0 new comments

• [TD] Test removal on sm86

  #127131 commented on May 29, 2024 • 0 new comments

• [wip, dynamo] trace through FunctionalTensor

  #126936 commented on May 30, 2024 • 0 new comments

• Default traceable subclasses to use swap_tensors path for load_state_dict

  #126788 commented on May 30, 2024 • 0 new comments

• Enable XPU operator codegen via backend whitelist

  #126977 commented on May 29, 2024 • 0 new comments

• reset dynamo cache before each test

  #126586 commented on May 25, 2024 • 0 new comments

• [export] add nonstrict, retrace, serdes patching for _trace._export tests

  #126810 commented on May 28, 2024 • 0 new comments

• Initial commit of flight recorder analyzer

  #126726 commented on Jun 1, 2024 • 0 new comments

• Add private escape hatches to fall back to pre-swap tensors behavior

  #126984 commented on May 30, 2024 • 0 new comments

• [inductor][cpp] add vectorization support for double

  #126858 commented on May 29, 2024 • 0 new comments

• [CI] Ensure inductor/test_cpu_cpp_wrapper is actually run in inductor_cpp_wrapper_abi_compatible

  #126717 commented on Jun 1, 2024 • 0 new comments

• Avoid accessing storage in wrapper tensor

  #126878 commented on May 31, 2024 • 0 new comments

• HealthcheckNCCL

  #127044 commented on May 29, 2024 • 0 new comments

• Save backward graphs lazily to cache

  #126999 commented on May 31, 2024 • 0 new comments

• [RFC] Intel GPU Upstreaming

  #114723 commented on May 27, 2024 • 0 new comments

• Faster Pytorch dequantize() + matmul for quantized models

  #115985 commented on May 29, 2024 • 0 new comments

• [MPS] F.conv1d and F.conv2d produce incorrect gradients when minibatch >= 2^16

  #96225 commented on May 29, 2024 • 0 new comments

• DISABLED test_bmm_multithreaded (__main__.TestTorch)

  #125240 commented on May 29, 2024 • 0 new comments

• DISABLED test_memory_format_type_cuda (__main__.TestTorchDeviceTypeCUDA)

  #126954 commented on May 29, 2024 • 0 new comments

• Choose better configs for `tuned_mixed_mm`

  #127056 commented on May 29, 2024 • 0 new comments

• xpu: support torch.utils.data.DataLoader(pin_memory_device='xpu')

  #126491 commented on May 30, 2024 • 0 new comments

• xpu: can't build XPU backend without sourcing oneAPI environment variables (/opt/intel/oneapi/setvars.sh)

  #127008 commented on May 30, 2024 • 0 new comments

• When testing the scalar version, test_open_device_registration will fail

  #126372 commented on May 30, 2024 • 0 new comments

• When testing the scalar version, test_AllenaiLongformerBase_repro will fail

  #126262 commented on May 30, 2024 • 0 new comments

• [Inductor] Kill Mutation Layout

  #118570 commented on May 30, 2024 • 0 new comments

• Lowering for the Average Pooling 3D backward operation

  #127101 commented on May 30, 2024 • 0 new comments

• UNSTABLE inductor / cuda12.1-py3.10-gcc9-sm86 / test (inductor_timm)

  #126884 commented on May 30, 2024 • 0 new comments

• Invalid Reference to Class

  #99107 commented on May 30, 2024 • 0 new comments

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

  #82677 commented on May 30, 2024 • 0 new comments

• [ONNX] Do not include concrete tensors when we add initializers

  #127140 commented on May 31, 2024 • 0 new comments

• [ONNX] Support external tensors in ONNXProgram.save()

  #127142 commented on May 31, 2024 • 0 new comments

• [ONNX] ExportedProgram weight serialization support

  #127138 commented on May 31, 2024 • 0 new comments

• [ONNX] Move the graph building logic

  #127139 commented on May 31, 2024 • 0 new comments

• [AOTI] Using `AOTI_TORCH_CHECK` will cause performance drop on several models compared with using `TORCH_CHECK`

  #126665 commented on May 31, 2024 • 0 new comments

• Cpp-wrapper mode issue tracker

  #117363 commented on May 31, 2024 • 0 new comments

• torchrun fails to run on Windows 11

  #108602 commented on May 31, 2024 • 0 new comments

• What is the processing principle when the complex64 input tensor contains nan or inf for addition?

  #127075 commented on May 27, 2024 • 0 new comments

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

  #24422 commented on May 28, 2024 • 0 new comments

• randn generates different output for 4x4 tensor size sliced to match shape of direct 2x4 or 4x2 and compare output

  #127164 commented on May 28, 2024 • 0 new comments

• UNSTABLE pull / linux-focal-cuda12.4-py3.10-gcc9-sm86 / build

  #127104 commented on May 28, 2024 • 0 new comments

• function signature for multiprocessing.spawn is multiprocessing.spawn.spawn

  #126899 commented on May 28, 2024 • 0 new comments

• Segmentation error for torch==2.2.1 on MacOs

  #121101 commented on May 28, 2024 • 0 new comments

• [DSD] keep 'initial_lr' in `torch.distributed.checkpoint.state_dict.set_optimizer_state_dict`

  #126948 commented on May 28, 2024 • 0 new comments

• Python 3.12 CPU Build system cannot find MKL libraries

  #119557 commented on May 28, 2024 • 0 new comments

• Unexpected MYPY linter errors on CI

  #126361 commented on May 28, 2024 • 0 new comments

• Make the `sccache` cache easily available to all pytorch contributors in readonly mode

  #125297 commented on May 28, 2024 • 0 new comments

• [RFC] Autoload Device Extension

  #122468 commented on May 29, 2024 • 0 new comments

• [RFC] Raising minimal glibc support to: glibc2_28 . Deprecation support for Amazon Linux 2 support for PyTorch Release 2.5

  #126551 commented on May 29, 2024 • 0 new comments

• Add warning messages to provide info about expected performance improvement using cuda for a specific model

  #126874 commented on May 29, 2024 • 0 new comments

• Support for no-frills FP8 matmuls

  #123761 commented on May 29, 2024 • 0 new comments

• Inaccurate filename due to test class metaprogramming that doesn't set __file__ and __name__

  #125467 commented on May 29, 2024 • 0 new comments

• torch.cuda.BoolTensor uses 8 bits per element, not 1 bit as reported by element_size()

  #41571 commented on May 29, 2024 • 0 new comments

• [question] How hard would it be to implement 4-bit precision training?

  #49298 commented on May 29, 2024 • 0 new comments

• No factory functions for strided quantized tensors

  #74540 commented on May 29, 2024 • 0 new comments

• [discussion, idea] Batched, vectorized base64 decoding / encoding + maybe RLE decoding / encoding

  #90560 commented on May 29, 2024 • 0 new comments

• Inplace fused (leaky)relu+(leaky)dropout for memory savings (I think, can be made fully allocation-less if never fully allocating random mask in FlashAttention style and recover the mask from the output)

  #92927 commented on May 29, 2024 • 0 new comments

• [feature request] Specialized memory layouts and wide blocked/tiled dtypes for cublasLt/onednn: e.g. torch.float16x32 / torch.int8x32 / torch.bits1x512 (akin to torch.quint2x4)

  #104702 commented on May 29, 2024 • 0 new comments

• The performance of multiplication of two matrices is different between window and linux

  #26345 commented on May 31, 2024 • 0 new comments

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

  #124147 commented on May 31, 2024 • 0 new comments

• [DO NOT MERGE] Test new ROCm CI nodes

  #124424 commented on May 31, 2024 • 0 new comments

• [dynamo] Support ndarray.dtype attribute access

  #124490 commented on May 31, 2024 • 0 new comments

• While loop autograd

  #124573 commented on Jun 1, 2024 • 0 new comments

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

  #124702 commented on May 31, 2024 • 0 new comments

• [WIP][Inductor] Update Intel GPU Triton commit pin.

  #124842 commented on May 31, 2024 • 0 new comments

• Add Efficient Attention support on ROCM

  #124885 commented on May 28, 2024 • 0 new comments

• S390x ci periodic tests

  #125401 commented on May 30, 2024 • 0 new comments

• [Inductor][ROCm] Composable Kernel backend for Inductor

  #125453 commented on Jun 1, 2024 • 0 new comments

• Uses memory pools for mixing CUDA allocators

  #125722 commented on May 29, 2024 • 0 new comments

• [AOTI][not for review] Test cpp_wrapper mode

  #125733 commented on May 29, 2024 • 0 new comments

• [CI]enable AMP accuracy test for inductor on SPR CPU

  #125748 commented on May 29, 2024 • 0 new comments

• Separate AOTI Eager utils as a single file

  #125819 commented on May 31, 2024 • 0 new comments

• [3/N] Non-Tensor: Support string parameter for aten operations

  #125831 commented on May 31, 2024 • 0 new comments

• [4/N] Non-Tensor: Support layout, device and dtype for aten operations

  #125897 commented on May 31, 2024 • 0 new comments

• Fix tensor subclass + dynamic shapes in torch.compile + aot autograd

  #125941 commented on May 31, 2024 • 0 new comments

• Remove deprecated _aminmax operator

  #125995 commented on May 28, 2024 • 0 new comments

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

  #126054 commented on May 30, 2024 • 0 new comments

• Enable UFMT format on test/quantization

  #126152 commented on May 27, 2024 • 0 new comments

• [wip][inductor] move loop ordering after fusion

  #126254 commented on May 29, 2024 • 0 new comments

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

  #126304 commented on Jun 1, 2024 • 0 new comments

• Support Delay Loading of c10.dll in when using libtorch as a thirdparty library.

  #105058 commented on May 31, 2024 • 0 new comments

• Cost & performance estimation for Windows Arm64 compilation

  #92302 commented on May 31, 2024 • 0 new comments

• torch.nn.AdaptiveAvgPool2d lacks checking of input dimension

  #126673 commented on May 31, 2024 • 0 new comments

• Broken Link and unfinished sentence in Frequently Asked Questions

  #126367 commented on May 31, 2024 • 0 new comments

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

  #76191 commented on May 31, 2024 • 0 new comments

• [BE] wrap deprecated function/class with `typing_extensions.deprecated` for better IDE integration

  #126888 commented on Jun 1, 2024 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on Jun 1, 2024 • 0 new comments

• [POC][pytree] test flattening dict in sorted order

  #115014 commented on May 29, 2024 • 0 new comments

• Automated submodule update: FBGEMM

  #115316 commented on Jun 1, 2024 • 0 new comments

• Factory function and basic .sizes() support for C++ NestedTensor

  #117905 commented on May 28, 2024 • 0 new comments

• Switch batch norm stack to consolidated ops

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

• [NJT] Store vec on nested ints

  #119976 commented on May 28, 2024 • 0 new comments

• [NJT] Factory function support

  #119977 commented on May 28, 2024 • 0 new comments

• [AMD] Turn on hipblaslt by default

  #120480 commented on May 31, 2024 • 0 new comments

• Optionally use hipblaslt

  #120551 commented on May 30, 2024 • 0 new comments

• Enable clang-tidy on c10/util/Float8*.h

  #120573 commented on May 26, 2024 • 0 new comments

• [do not review]

  #120881 commented on May 28, 2024 • 0 new comments

• Adds general tensor equality subsystem

  #121481 commented on May 28, 2024 • 0 new comments

• [draft] python 3.13 test

  #121979 commented on May 31, 2024 • 0 new comments

• Improve decomposition for constand_pad_nd

  #123661 commented on May 29, 2024 • 0 new comments

• Add auto-tuning for sparse semi-structured MM operator

  #123742 commented on May 30, 2024 • 0 new comments