70 Pull requests opened by 46 people

• Add pybind support for is_hpu property of Tensor object

  #132228 opened Jul 31, 2024

• [CI][expriment] try linux.arm64.m7g.metal

  #132232 opened Jul 31, 2024

• Optimize sort kernel for contiguous tensors

  #132236 opened Jul 31, 2024

• make functorch CSE respect mutations as barriers (like fsdp.set_)

  #132243 opened Jul 31, 2024

• [TD] Fuzz testing?

  #132282 opened Jul 31, 2024

• [c10d][experimental] Add _abort_process_group

  #132291 opened Jul 31, 2024

• Fix inf value reduction in non persistent reduction for scans

  #132293 opened Jul 31, 2024

• Run cudagraphs on AOTAutograd cache hit

  #132294 opened Jul 31, 2024

• [PT2][Optimus] Add missing example value for introduced nodes

  #132297 opened Jul 31, 2024

• [ONNX] Define the `TORCH_ONNX_USE_EXPORTED_PROGRAM` flag

  #132299 opened Jul 31, 2024

• [torchbind] don't warning for certain skippable methods.

  #132306 opened Jul 31, 2024

• Split of "[reland] [export] fix zero arg export in training_ir and constant tensor handling"

  #132307 opened Jul 31, 2024

• [BE][typing] fix types in common pruning

  #132309 opened Jul 31, 2024

• [DeviceMesh] Add supports for non-continuous slicing

  #132310 opened Jul 31, 2024

• [DeviceMesh] Update slicing documentation to include nD and non-continuous slicing

  #132311 opened Jul 31, 2024

• [CUDA] `is_bf16_supported()` should not crash if there are no GPUs

  #132313 opened Jul 31, 2024

• [export][pre-dispatch] add test case for user-defined triton kernels

  #132317 opened Jul 31, 2024

• [dynamo] Treat attr of unspecialized buiitin nn modules as static

  #132318 opened Jul 31, 2024

• [AOTI][refactor] Move set_cpp_kernel to base class

  #132319 opened Jul 31, 2024

• [AOTI][refactor] Consolidate how python_kernel_name is set

  #132320 opened Jul 31, 2024

• fsdp.set_: convey to functionalization that it mutates storage

  #132322 opened Jul 31, 2024

• [AOTI][tooling][1/n] Add intermediate value debug printer

  #132323 opened Jul 31, 2024

• [pyre][ptt] Prepare to enable type checking by default caffe2/[a-s]* directories

  #132324 opened Jul 31, 2024

• [inductor] fix UndefinedTensorImpl singleton can't export on Windows.

  #132326 opened Jul 31, 2024

• NOCOMMIT test Pyre CI for Pytorch

  #132327 opened Jul 31, 2024

• Ban decorator usage of dynamo_timed

  #132328 opened Jul 31, 2024

• [dynamo] fix add_push_null callsites with CALL_FUNCTION_EX

  #132329 opened Jul 31, 2024

• [dynamo] support inspect.signature.bind

  #132330 opened Jul 31, 2024

• [dynamo] Add line number to _warn_capture_scalar_outputs()

  #132333 opened Jul 31, 2024

• [dynamo] Track params/buffers and mark them as static

  #132334 opened Jul 31, 2024

• Add None return type to init

  #132335 opened Jul 31, 2024

• Add Adafactor foreach impl

  #132336 opened Jul 31, 2024

• move torch._functionalize APIs to pybind. add one for marking storage mutations

  #132337 opened Jul 31, 2024

• [test] dynamo: dont hold onto real tensors in the __dict__ of any graph inputs

  #132338 opened Jul 31, 2024

• [DeviceMesh] Remove parent mesh concept from _MeshEnv and replace by root mesh

  #132339 opened Jul 31, 2024

• Only make wait_tensor as a side_effect op

  #132340 opened Jul 31, 2024

• Only make wait_tensor as a side_effect op

  #132341 opened Jul 31, 2024

• [CP] Extend CP to support load-blancing shards

  #132342 opened Jul 31, 2024

• Fix file lock issue in AotCodeCompiler

  #132343 opened Jul 31, 2024

• Cast inputs to low precision kernels in emulate low precision mode

  #132345 opened Jul 31, 2024

• [AOTI] Fix complex not defined

  #132347 opened Jul 31, 2024

• Set weights_only=False in export custom object load

  #132348 opened Jul 31, 2024

• Fix all RuntimeErrors during weights_only load from being erroneously reported with the weights_only message

  #132349 opened Jul 31, 2024

• Add None return type to init -- functorch and torchgen

  #132351 opened Jul 31, 2024

• Add None return type to init -- tests

  #132352 opened Jul 31, 2024

• [export] Apply CIA override only if implemented

  #132353 opened Jul 31, 2024

• [Inductor] Small updates to B2B-GEMM

  #132354 opened Jul 31, 2024

• [MPS] Add regression test for memory leak in `nn.Linear`

  #132355 opened Aug 1, 2024

• [NJT][flop counter] attention: if offsets are fake, use max seqlen

  #132356 opened Aug 1, 2024

• [pytorch][counters] Pybind for WaitCounter

  #132357 opened Aug 1, 2024

• [WIP][partitioners] handle nodes with >1 output

  #132359 opened Aug 1, 2024

• Remove some alias for c10::optional

  #132361 opened Aug 1, 2024

• [Dynamo] Support abc.MutableMapping.get

  #132363 opened Aug 1, 2024

• Revert "[aoti] Fix float16 and bfloat16 for generated GPU code (#131437)"

  #132365 opened Aug 1, 2024

• [AOTI][refactor] Update MKLDNN ops cpp wrapper support

  #132367 opened Aug 1, 2024

• [pipelining] Make test_schedule quiet

  #132369 opened Aug 1, 2024

• [WIP] refactor distributed code

  #132371 opened Aug 1, 2024

• Change deprecate warning on dispatch_on_subclass to warn once

  #132374 opened Aug 1, 2024

• add testcase for kineto on-demand profiling, check #131020

  #132375 opened Aug 1, 2024

• Add None return type to init -- tests rest

  #132376 opened Aug 1, 2024

• [Easy] Fix argument name collision in `HigherOrderOperator` dispatched functions

  #132377 opened Aug 1, 2024

• Move slow tests to be in repo

  #132379 opened Aug 1, 2024

• Reduce number of guards introduced by check_cudnn_tensor_shapes when cudnn version is higher enough

  #132384 opened Aug 1, 2024

• [inductor] cpp codegen alignas for all OSs.

  #132387 opened Aug 1, 2024

• Apply loop split optimization in codegen_node

  #132389 opened Aug 1, 2024

• Update torch-xpu-ops pin (ATen XPU implementation)

  #132390 opened Aug 1, 2024

• [3/N][dtensor] Strided Sharding offset calculation util

  #132391 opened Aug 1, 2024

• add src map to data-dependent errors

  #132393 opened Aug 1, 2024

• [inductor] make restrict_keyword cross OSs.

  #132394 opened Aug 1, 2024

• [11/N] Use std::nullopt and std::optional

  #132396 opened Aug 1, 2024

19 Issues closed by 13 people

• DISABLED test_inplace_grad_div_floor_rounding_cuda_float64 (__main__.TestBwdGradientsCUDA)

  #100520 closed Aug 1, 2024

• triton kernels (and maybe custom ops) that mutate multiple inputs silently incorrect with torch.compile

  #132196 closed Aug 1, 2024

• torch.Tensor.uniform_ causes invalid syntax in InternalTorchDynamoError

  #131019 closed Jul 31, 2024

• Document newness of `.inference_mode()` ?

  #132288 closed Jul 31, 2024

• [AOTI] error: ‘bfloat16’ was not declared in this scope

  #122986 closed Jul 31, 2024

• Unexpected Behavior with torch.mean and torch.sum on Empty Dimensions

  #131457 closed Jul 31, 2024

• Even when force_disable_caches we still attempt to upload to remote cache

  #132241 closed Jul 31, 2024

• xpu: intel conda channel is not available

  #131802 closed Jul 31, 2024

• tolist() on single item

  #132142 closed Jul 31, 2024

• Nightly windows wheels build failure Jul 31. 2024

  #132295 closed Jul 31, 2024

• [CUDA][Inductor][Pooling] `test_comprehensive_nn_functional_max_pool2d_cuda` appears broken on H100

  #132199 closed Jul 31, 2024

• Mutating global variable during inlining a function from imported module breaks in dynamo

  #132165 closed Jul 31, 2024

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

  #132298 closed Jul 31, 2024

• `torch.export` Fails with `nn.MSELoss` Due to No Tensor Operations Found During Tracing

  #132141 closed Jul 31, 2024

• torch._dynamo.exc.Unsupported: builtin: str [<class 'torch._dynamo.variables.user_defined.UserDefinedObjectVariable'>] False

  #130301 closed Jul 31, 2024

• Numpy Compatibility Issue with PyTorch DataLoaders (RuntimeError: Numpy is not available)

  #132179 closed Jul 31, 2024

• DISABLED test_fully_shard_post_optim_event_overlap (__main__.TestFullyShardOverlap)

  #131081 closed Jul 31, 2024

• AssertionError: Empty array is not supported in C

  #131335 closed Jul 31, 2024

• RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [1, 1, 881, 609]], which is output 0 of AsStridedBackward0, is at version 2; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

  #132149 closed Jul 31, 2024

69 Issues opened by 30 people

• tensor.to(device) not copying data correctly between two GPUs

  #132397 opened Aug 1, 2024

• `torch.multinomial` generates incorrect distribution

  #132395 opened Aug 1, 2024

• `amp.custom_fwd` interacts poorly with custom `autograd.Function` with `setup_context` separated from `forward`

  #132388 opened Aug 1, 2024

• torch cpu float16 range is not aligned with scipy on polygamma.

  #132386 opened Aug 1, 2024

• crnn model from pth to torchscript problem

  #132385 opened Aug 1, 2024

• dynamic tensor shape of vmap dimention

  #132381 opened Aug 1, 2024

• [inductor][cpu] LayoutLMForSequenceClassification amp single thread accuracy failure

  #132380 opened Aug 1, 2024

• per sample gradient with torch.autograd.grad

  #132378 opened Aug 1, 2024

• [BUG] C++ error in `torch.compile` when `dynamic=True`

  #132373 opened Aug 1, 2024

• Bizarre segmentation fault on MacOS M3 when sklearn is imported

  #132372 opened Aug 1, 2024

• Switch to new name `oneDNN` instead of `MKLDNN`

  #132368 opened Aug 1, 2024

• PyTorch's Distributed Checkpoint Cannot Save a Parameter of Size 1

  #132366 opened Aug 1, 2024

• Dynamo bug when handling KeyError

  #132362 opened Aug 1, 2024

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

  #132360 opened Aug 1, 2024

• public API for checking local tensor type

  #132358 opened Aug 1, 2024

• [MPS] Memory leak in `nn.Linear`

  #132332 opened Jul 31, 2024

• DISABLED test_profile_memory (__main__.CppThreadTest)

  #132331 opened Jul 31, 2024

• Support Enum for dictionaries in TorchScript

  #132315 opened Jul 31, 2024

• is_bf_16_supported() should return False when running on CPU **NOT AN ERROR**

  #132303 opened Jul 31, 2024

• flash attention triton kernel x pt2 silently incorrect

  #132301 opened Jul 31, 2024

• CUDA Memory Explosion -- Conv3D

  #132300 opened Jul 31, 2024

• Can we avoid RecordStream for P2P ops?

  #132292 opened Jul 31, 2024

• 25+ compile tests were disabled as flaky

  #132290 opened Jul 31, 2024

• test_torch.py::TestTorchDeviceTypeCUDA::test_deepcopy_scalar_cuda_float32 and others leaking memory in inductor

  #132287 opened Jul 31, 2024

• [inductor][cpu] Reformer multiple thread CPP wrapper performance passed with static shape but failed with dynamic shape

  #132286 opened Jul 31, 2024

• On RoCM, some torch.compile fails with 0 mismatched elements

  #132283 opened Jul 31, 2024

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

  #132281 opened Jul 31, 2024

• DISABLED test_compile_forward_deg2rad_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132280 opened Jul 31, 2024

• DISABLED test_compile_backward_tan_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132279 opened Jul 31, 2024

• DISABLED test_compile_backward___radd___cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132278 opened Jul 31, 2024

• DISABLED test_compile_backward_div_floor_rounding_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132277 opened Jul 31, 2024

• DISABLED test_compile_backward_nan_to_num_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132276 opened Jul 31, 2024

• DISABLED test_compile_backward_bfloat16_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132275 opened Jul 31, 2024

• DISABLED test_compile_backward_half_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132274 opened Jul 31, 2024

• DISABLED test_compile_forward_asin_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132273 opened Jul 31, 2024

• DISABLED test_compile_forward_sgn_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132272 opened Jul 31, 2024

• DISABLED test_compile_backward_nn_functional_relu_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132271 opened Jul 31, 2024

• DISABLED test_compile_forward_nn_functional_relu_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132270 opened Jul 31, 2024

• DISABLED test_compile_forward_special_scaled_modified_bessel_k0_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132269 opened Jul 31, 2024

• DISABLED test_compile_backward_sinh_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132268 opened Jul 31, 2024

• DISABLED test_compile_backward_special_xlog1py_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132267 opened Jul 31, 2024

• DISABLED test_compile_forward_tan_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132265 opened Jul 31, 2024

• DISABLED test_compile_forward_angle_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132266 opened Jul 31, 2024

• DISABLED test_compile_backward_mul_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132264 opened Jul 31, 2024

• DISABLED test_compile_forward_conj_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132263 opened Jul 31, 2024

• DISABLED test_compile_forward_log10_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132262 opened Jul 31, 2024

• DISABLED test_compile_backward_log10_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132261 opened Jul 31, 2024

• DISABLED test_compile_backward_float_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132260 opened Jul 31, 2024

• DISABLED test_compile_forward_max_binary_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132259 opened Jul 31, 2024

• DISABLED test_compile_forward_fmax_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132258 opened Jul 31, 2024

• DISABLED test_compile_forward_min_binary_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132257 opened Jul 31, 2024

• DISABLED test_compile_backward_sigmoid_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132256 opened Jul 31, 2024

• DISABLED test_compile_backward_floor_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132255 opened Jul 31, 2024

• DISABLED test_compile_backward_ceil_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132254 opened Jul 31, 2024

• DISABLED test_compile_forward_ge_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132253 opened Jul 31, 2024

• DISABLED test_compile_backward_special_log_ndtr_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132252 opened Jul 31, 2024

• DISABLED test_compile_backward_positive_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132251 opened Jul 31, 2024

• DISABLED test_compile_backward_lgamma_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132250 opened Jul 31, 2024

• DISABLED test_compile_backward_cos_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132249 opened Jul 31, 2024

• DISABLED test_compile_backward_abs_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132248 opened Jul 31, 2024

• DISABLED test_compile_forward_abs_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132247 opened Jul 31, 2024

• DISABLED test_compile_backward_conj_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132246 opened Jul 31, 2024

• DISABLED test_compile_forward_float_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132245 opened Jul 31, 2024

• DISABLED test_compile_forward_positive_cuda_float32 (__main__.TestNestedTensorOpInfoCUDA)

  #132244 opened Jul 31, 2024

• Inductor autotuner time_taken_ns is not accurate

  #132242 opened Jul 31, 2024

• torch.nn.functional.interpolate completely broken with torch.jit.script and torch.fx

  #132240 opened Jul 31, 2024

• hasattr tracing for PythonModuleVariable is unsupported #129742

  #132237 opened Jul 31, 2024

• tensors, created with factory methods inside Subclasses.__torch_dispatch___ are not fakified

  #132235 opened Jul 31, 2024

• torch.nn.Bilinear weights initialization reason

  #132231 opened Jul 31, 2024

239 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.

• [RFC][c10d] Add a new API for timeout extend for one local rank and the timeout will reset when the first collective finishes

  #130905 commented on Aug 1, 2024 • 22 new comments

• Inductor-CPU WoQ int8 GEMM micro-kernel with scale epilogue

  #131887 commented on Aug 1, 2024 • 22 new comments

• [NJT] Support Chunk backward for simple cases

  #132193 commented on Aug 1, 2024 • 12 new comments

• [inductor][triton] improved kernel-level benchmarking

  #130926 commented on Aug 1, 2024 • 10 new comments

• AutoHeuristic: mixed_mm heuristic for A100

  #131613 commented on Aug 1, 2024 • 7 new comments

• [12/N] Fix clang-tidy warnings in jit

  #132209 commented on Jul 31, 2024 • 7 new comments

• Feature test ability in c10

  #131248 commented on Aug 1, 2024 • 4 new comments

• Construct NJT without graph breaks

  #130292 commented on Jul 31, 2024 • 4 new comments

• Add instruction count benchmark to run on pull requests

  #131475 commented on Aug 1, 2024 • 3 new comments

• typing ir.py - part 2

  #131846 commented on Aug 1, 2024 • 3 new comments

• [c10d][Log] Use pg_id instead of pg_name for logging prefix

  #132058 commented on Jul 31, 2024 • 3 new comments

• [traced-graph][sparse] propagate sparsity in fx graph

  #131920 commented on Aug 1, 2024 • 3 new comments

• [nccl] Wrap nccl code update with version check

  #130419 commented on Aug 1, 2024 • 2 new comments

• [pipelining] Add schedule runtime for lowered schedule

  #130488 commented on Aug 1, 2024 • 2 new comments

• SparseCsrCUDA: cuDSS backend for linalg.solve

  #129856 commented on Jul 31, 2024 • 2 new comments

• [PP] Forward only schedule

  #132177 commented on Aug 1, 2024 • 2 new comments

• Use FakeTensor cache for subclass inner tensors

  #131803 commented on Aug 1, 2024 • 2 new comments

• [Intel GPU] Allow XPU device in copy, cdist, index_put_impl

  #130088 commented on Jul 31, 2024 • 2 new comments

• [AOTI] Fix number type for AOTI

  #132180 commented on Aug 1, 2024 • 1 new comment

• [AOTI] Fix bfloat16 in CPU

  #132150 commented on Aug 1, 2024 • 1 new comment

• [ROCm] Add AMDSMI support for UUID input

  #129741 commented on Aug 1, 2024 • 1 new comment

• Add registration mechanism for aoti model runner

  #131638 commented on Jul 31, 2024 • 1 new comment

• Generalization of distributed UT content to enable non cuda device execution

  #131758 commented on Jul 31, 2024 • 1 new comment

• [test/torch_np] Fix usages of deprecated NumPy 2.0 APIs in numpy_tests

  #131909 commented on Jul 31, 2024 • 1 new comment

• [inductor] Replace torch.allclose with torch.testing.assert_close in test_fx_fusion

  #130618 commented on Aug 1, 2024 • 1 new comment

• [C10D] Clarify warning for concurrent PG usage

  #131895 commented on Jul 31, 2024 • 1 new comment

• [CUDA][SDPA] Fix expect export on sm90+

  #132194 commented on Aug 1, 2024 • 1 new comment

• [2/N] Fix clang-tidy warnings in aten/src/ATen/native/*.{cpp,h}

  #131834 commented on Aug 1, 2024 • 1 new comment

• adds kl for multinomial

  #130899 commented on Jul 31, 2024 • 1 new comment

• fix a typo in the householder_product docs

  #124279 commented on Jul 31, 2024 • 1 new comment

• Update SavedTensorHooks TLS stack to use SafePyObject

  #131700 commented on Jul 31, 2024 • 1 new comment

• WIP batch B,W into BW

  #131762 commented on Aug 1, 2024 • 0 new comments

• [Fix]: prim::If with multiple outputs and input return directly

  #131779 commented on Jul 31, 2024 • 0 new comments

• build head~1 and run pr-time benchmarks on it

  #131751 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: Heuristic that ranks choices for mm

  #131714 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: Enable explicit support for ranking

  #131710 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: mixed_mm h100 heuristic

  #131790 commented on Aug 1, 2024 • 0 new comments

• Add get_optin_feature() to allow opt-in to amz2023

  #131792 commented on Jul 31, 2024 • 0 new comments

• fit(...) method of nn.Module class

  #131806 commented on Aug 1, 2024 • 0 new comments

• [inductor] benchmarking collective

  #131819 commented on Jul 31, 2024 • 0 new comments

• Skip frame if torch dispatch mode enabled

  #131828 commented on Aug 1, 2024 • 0 new comments

• Fix flaky test_non_contiguous_input_mm_plus_mm and test_precompilations

  #131835 commented on Aug 1, 2024 • 0 new comments

• Call _safe_softmax from sdpa math path

  #131844 commented on Aug 1, 2024 • 0 new comments

• Update fused flash cpu kernel to set masked out rows to 0

  #131863 commented on Aug 1, 2024 • 0 new comments

• Update fused mem-eff kernel to set masked out rows to 0

  #131867 commented on Aug 1, 2024 • 0 new comments

• Add a private _safe_softmax

  #131060 commented on Aug 1, 2024 • 0 new comments

• Implement masked_select op for NestedTensors

  #131069 commented on Jul 31, 2024 • 0 new comments

• [aoti] Remove mutable tensor assertion

  #131074 commented on Jul 31, 2024 • 0 new comments

• Add cmake option USE_SYSTEM_FBGEMM

  #131282 commented on Aug 1, 2024 • 0 new comments

• [inductor] support vec for atomic add

  #131314 commented on Aug 1, 2024 • 0 new comments

• [CP] Rewrite ring attention backward algorithm and enablement APIs

  #131351 commented on Aug 1, 2024 • 0 new comments

• [DONT REVIEW] Prototype DeviceMesh.create_view_dim()

  #131352 commented on Aug 1, 2024 • 0 new comments

• Fix constant propagation in builtins and UserClasses

  #131354 commented on Jul 31, 2024 • 0 new comments

• [FSDP][dtensor] add FSDP2+TP distributed state dict test

  #131408 commented on Aug 1, 2024 • 0 new comments

• [ts-migration][1/N]: Add prim::Loop for constant number of iterations and condition

  #131418 commented on Aug 1, 2024 • 0 new comments

• [executorch hash update] update the pinned executorch hash

  #131420 commented on Aug 1, 2024 • 0 new comments

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

  #131481 commented on Jul 31, 2024 • 0 new comments

• [CUDA][CUTLASS] Fixes for CUTLASS upgrade

  #131493 commented on Aug 1, 2024 • 0 new comments

• [ts-migration]: Add support for aten::append

  #131548 commented on Jul 31, 2024 • 0 new comments

• Add explicit GQA support.

  #131559 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: tuned_mm

  #131615 commented on Aug 1, 2024 • 0 new comments

• inductor mm autotuning: add back previously pruned configs

  #131616 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: script to generate data for mm

  #131617 commented on Aug 1, 2024 • 0 new comments

• Correct sample creation of torch.histogram in UT op_db to align PyTorch defined operator semantics

  #131630 commented on Jul 31, 2024 • 0 new comments

• Add Sleef Implementation for maximum Kernel for ARM

  #131642 commented on Aug 1, 2024 • 0 new comments

• [pt2e][quant] Ensure BN node is erased after convert

  #131651 commented on Aug 1, 2024 • 0 new comments

• [do not land] Test warm start compile with changes

  #131660 commented on Jul 31, 2024 • 0 new comments

• AutoHeuristic: Support ranking/pruning choices

  #131705 commented on Aug 1, 2024 • 0 new comments

• Add device to create_cpu_state_dict api to make it backend agnostic

  #131868 commented on Jul 31, 2024 • 0 new comments

• [torch.special] Adding betainc with backward operation

  #132135 commented on Jul 31, 2024 • 0 new comments

• simplify return string

  #132144 commented on Jul 31, 2024 • 0 new comments

• Speedup int8mm_kernel with RVV

  #132146 commented on Jul 31, 2024 • 0 new comments

• Fix autotuning for flex_decoding

  #132157 commented on Aug 1, 2024 • 0 new comments

• [pytorch][counters] Pybind for WaitCounter

  #132167 commented on Jul 31, 2024 • 0 new comments

• disable the new cpp_builder args UT

  #132168 commented on Jul 31, 2024 • 0 new comments

• Update the FQN for auto_functionalized HOO.

  #132171 commented on Jul 31, 2024 • 0 new comments

• [export] change deepcopy to copy in _replace_set_grad_with_hop pass..

  #132181 commented on Aug 1, 2024 • 0 new comments

• Reland "[1/2] PT2 Inductor ComboKernels - Foreach cases (#124969)"

  #132182 commented on Jul 31, 2024 • 0 new comments

• [PT2] Port remove_noop to PT2 pre_grad passes

  #132183 commented on Aug 1, 2024 • 0 new comments

• [dtensor][debug] adding js script to pytorch github so that i can host the browser visualizer on pytorch

  #132185 commented on Jul 31, 2024 • 0 new comments

• [TS2E] Remove reference to torch.onnx internals

  #132186 commented on Aug 1, 2024 • 0 new comments

• C++ network flow implementation in c10

  #132188 commented on Aug 1, 2024 • 0 new comments

• Add optional exact_ordering parameter to the DataLoader

  #132189 commented on Aug 1, 2024 • 0 new comments

• [AOTI] Fix a typo in ExternKernel.codegen_const_args

  #132191 commented on Jul 31, 2024 • 0 new comments

• [fx] python_code(verbose=True): show size/strides for all tensors

  #132192 commented on Aug 1, 2024 • 0 new comments

• [TESTING] mark dynamo output_graph code log as "warning"

  #132195 commented on Aug 1, 2024 • 0 new comments

• [DTensor] Added naive replicate strategy for more diagonal ops

  #132201 commented on Jul 31, 2024 • 0 new comments

• Enable CUDA 12.4.1

  #132202 commented on Aug 1, 2024 • 0 new comments

• [WIP] Introduce a device-agnostic runtime API design

  #132204 commented on Aug 1, 2024 • 0 new comments

• use spawn as default start method to create dataloader subprocess

  #132210 commented on Jul 31, 2024 • 0 new comments

• [pytorch/mtia] MTIA equivalent of torch.<device>.get_device_properties

  #132211 commented on Jul 31, 2024 • 0 new comments

• introduce regression to be detected

  #132215 commented on Aug 1, 2024 • 0 new comments

• Populate submodules of `torch._C` to `sys.modules` recursively

  #132216 commented on Aug 1, 2024 • 0 new comments

• [Intel GPU] Remove special dispatch logic for xpu in adaptive_avg_pooling

  #132217 commented on Aug 1, 2024 • 0 new comments

• Workaround to SDPA bug in MacOS15

  #132220 commented on Jul 31, 2024 • 0 new comments

• [CI] Update CPU inductor smoke test model list and target

  #132221 commented on Aug 1, 2024 • 0 new comments

• [WIP] Enable Windows Arm64

  #132225 commented on Aug 1, 2024 • 0 new comments

• Upgrade submodule oneDNN to v3.5.1

  #131877 commented on Aug 1, 2024 • 0 new comments

• Add oneDNN BRGEMM support

  #131878 commented on Aug 1, 2024 • 0 new comments

• Use brgemm for Half flash attention kernel

  #131879 commented on Aug 1, 2024 • 0 new comments

• Add `padding_side` to `pad_sequence` with `"left"` and `"right"` options (`"right"` as default)

  #131884 commented on Aug 1, 2024 • 0 new comments

• [Inductor][CPP] Add vectorization support for double

  #131886 commented on Jul 31, 2024 • 0 new comments

• support Conv BN folding for inline inbuilt modules

  #131888 commented on Jul 31, 2024 • 0 new comments

• Enable FlashAttention on Windows

  #131906 commented on Jul 31, 2024 • 0 new comments

• [export] Support "custom" metadata field.

  #131912 commented on Jul 31, 2024 • 0 new comments

• Fix symbolic nested int printing

  #131916 commented on Jul 31, 2024 • 0 new comments

• Optimize test transformers

  #131919 commented on Jul 31, 2024 • 0 new comments

• Run performance test non-alternately

  #131935 commented on Jul 31, 2024 • 0 new comments

• Add compiler bisector

  #131936 commented on Aug 1, 2024 • 0 new comments

• Fix sum() forward for NJT

  #131945 commented on Jul 31, 2024 • 0 new comments

• retry distributed tests on new AMI

  #131946 commented on Jul 31, 2024 • 0 new comments

• wip

  #131984 commented on Aug 1, 2024 • 0 new comments

• Various fix39

  #131989 commented on Jul 31, 2024 • 0 new comments

• [BE] Make maybe_aliasing_or_mutating proper tag

  #131990 commented on Aug 1, 2024 • 0 new comments

• [mtia][sdpa] MTIA SDPA dispatch via _fused_sdp_choice_stub

  #132008 commented on Aug 1, 2024 • 0 new comments

• [Inductor][FlexAttention] Add kwarg to top level for users to specify kernel params

  #132015 commented on Aug 1, 2024 • 0 new comments

• Preserve source_fn_stack in the training IR decomp

  #132033 commented on Aug 1, 2024 • 0 new comments

• Add clone_* variants for wrapper functions for better cProfile

  #132073 commented on Jul 31, 2024 • 0 new comments

• fastpath FunctionalTensor sizes()

  #132084 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: utils to make things easily reproducible

  #132088 commented on Aug 1, 2024 • 0 new comments

• [dynamo][dynamic] Treat buffers as static shape objects

  #132098 commented on Jul 31, 2024 • 0 new comments

• add script to compare results and fail at regression

  #132099 commented on Aug 1, 2024 • 0 new comments

• Add OpInfo for _convert_weight_to_int4pack

  #132112 commented on Jul 31, 2024 • 0 new comments

• [export] Fix serialization of OpOverload w/ SymInt outputs

  #132126 commented on Jul 31, 2024 • 0 new comments

• Add try except for _maybe_evaluate_static call in IndexPropagation

  #132128 commented on Aug 1, 2024 • 0 new comments

• torch.compile mode="max-autotune" precision appears to be lower

  #96693 commented on Jul 31, 2024 • 0 new comments

• Missing float8 storage

  #131196 commented on Jul 31, 2024 • 0 new comments

• torch.ops.fsdp.set_ on input doesn't actually modify the input (under torch.compile)

  #132197 commented on Jul 31, 2024 • 0 new comments

• backward of adaptive max pool (adaptive_max_pool2d_backward_cuda) doesn't have a deterministic implementation

  #131972 commented on Jul 31, 2024 • 0 new comments

• `torch.nn.functional.max_unpool2d`'s behavior is different on cpu and gpu on torch 2.5.0.dev20240708+cu121

  #132041 commented on Jul 31, 2024 • 0 new comments

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

  #123649 commented on Jul 31, 2024 • 0 new comments

• Static shape guards are put on tensors even if they never actually get used

  #131893 commented on Jul 31, 2024 • 0 new comments

• `all_reduce` hangs in minimal multi-node script

  #131781 commented on Jul 31, 2024 • 0 new comments

• [typing] Add missing `__all__` to modules.

  #131765 commented on Jul 31, 2024 • 0 new comments

• Can not torch.export.export with torch.arange

  #131889 commented on Jul 31, 2024 • 0 new comments

• Cpp-wrapper mode issue tracker

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

• [Inductor] Test failure in test_comprehensive_nn_functional_max_pool2d_cuda

  #131072 commented on Jul 31, 2024 • 0 new comments

• inspect.signature.bind is not supported

  #93760 commented on Aug 1, 2024 • 0 new comments

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

  #122169 commented on Aug 1, 2024 • 0 new comments

• [AudioLM] Graph break: 'skip function Random in file python3.10/random.py'

  #121349 commented on Aug 1, 2024 • 0 new comments

• [AudioLM] Graph break: 'inline in skipfiles: Random.randrange'

  #121350 commented on Aug 1, 2024 • 0 new comments

• torch.triu() may returns wrong values using MPS

  #100005 commented on Aug 1, 2024 • 0 new comments

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

  #124897 commented on Aug 1, 2024 • 0 new comments

• torch.nn.Transformer.generate_square_subsequent_mask gives nans instead of zeros when running on mps

  #116170 commented on Aug 1, 2024 • 0 new comments

• LibTorch cannot load PyTorch exported model

  #47917 commented on Aug 1, 2024 • 0 new comments

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

  #131958 commented on Aug 1, 2024 • 0 new comments

• Failed to export module with dynamic shapes

  #131897 commented on Aug 1, 2024 • 0 new comments

• [RFC] Use CUDA graphs by default on torch.compile

  #121968 commented on Aug 1, 2024 • 0 new comments

• aten::bucketize.Scalar doesn't work in compile mode

  #132222 commented on Aug 1, 2024 • 0 new comments

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

  #123771 commented on Aug 1, 2024 • 0 new comments

• [tracker] Open issue with inline_inbuilt_nn_modules

  #131696 commented on Aug 1, 2024 • 0 new comments

• LibTorch, Error in 'xxx': free(): invalid pointer

  #30507 commented on Aug 1, 2024 • 0 new comments

• RECORD_FUNCTION compilation error

  #131339 commented on Aug 1, 2024 • 0 new comments

• General MPS op coverage tracking issue

  #77764 commented on Aug 1, 2024 • 0 new comments

• performance drop because batching rule for aten::_scaled_dot_product_attention_math is not yet implemented

  #110525 commented on Jul 31, 2024 • 0 new comments

• "Adaptive pool MPS: input sizes must be divisible by output sizes", I keep getting this error even when I try to adjust for size

  #97109 commented on Jul 31, 2024 • 0 new comments

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

  #120219 commented on Jul 31, 2024 • 0 new comments

• [DOC] fix `mkl-static` install channel issue.

  #132103 commented on Jul 31, 2024 • 0 new comments

• `torch.nn.transformer.forward` returns incorrect value inside `torch.no_grad()` blocks.

  #132136 commented on Jul 31, 2024 • 0 new comments

• Broken link in doc.

  #132178 commented on Jul 31, 2024 • 0 new comments

• Adding betainc with backward operation

  #132133 commented on Jul 31, 2024 • 0 new comments

• NotImplementedError in some APIs with Nested Jagged Tensors

  #132212 commented on Jul 31, 2024 • 0 new comments

• [profiler] CUDA runtime op has wrong device time because of lazy init

  #132218 commented on Jul 31, 2024 • 0 new comments

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

  #24422 commented on Jul 31, 2024 • 0 new comments

• AssertionError: call_args and arg_types do not match

  #131337 commented on Jul 31, 2024 • 0 new comments

• Drop Connect Layer

  #131688 commented on Jul 31, 2024 • 0 new comments

• torch.onnx.export 2GiB limit

  #132205 commented on Jul 31, 2024 • 0 new comments

• TorchInductor CPU Performance Dashboard

  #93531 commented on Jul 31, 2024 • 0 new comments

• Expand Examples for torch.autograd.functional.jacobian

  #132140 commented on Jul 31, 2024 • 0 new comments

• Inconsistency between `torch.get_device` and `torch.Tensor.get_device` with `__torch_function__`

  #131944 commented on Jul 31, 2024 • 0 new comments

• DISABLED test_comprehensive__unsafe_masked_index_cuda_int32 (__main__.TestInductorOpInfoCUDA)

  #131118 commented on Jul 31, 2024 • 0 new comments

• Batching rule for `aten::_scaled_dot_product_efficient_attention`

  #102457 commented on Jul 31, 2024 • 0 new comments

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

  #126856 commented on Jul 31, 2024 • 0 new comments

• EventList still using old API breaks on Python 3.12

  #132227 commented on Jul 31, 2024 • 0 new comments

• torch.compile for torch.cumsum returns nans sometimes with all inf inputs

  #132107 commented on Jul 31, 2024 • 0 new comments

• Tests stopped working after merging #126905 on Arm64

  #132132 commented on Jul 31, 2024 • 0 new comments

• INTERNAL ASSERT FAILED at "../torch/csrc/autograd/python_torch_functions_manual.cpp":661 when returning a constant tensor in the forward method

  #132134 commented on Jul 31, 2024 • 0 new comments

• compile: `torch._subclasses.fake_tensor.FakeTensor` does not inherit from user-defined `torch.Tensor` subclass

  #132148 commented on Jul 31, 2024 • 0 new comments

• [dynamo] Context manager

  #132154 commented on Jul 31, 2024 • 0 new comments

• torch.ops.fsdp.set_ with torch.compile silently incorrect

  #132200 commented on Jul 31, 2024 • 0 new comments

• why not enable expandable_segments by default？

  #130330 commented on Jul 31, 2024 • 0 new comments

• TORCHELASTIC_RESTART_COUNT doesn't seem to be broadcasted to all worker

  #108158 commented on Jul 31, 2024 • 0 new comments

• FusedKernelCPU failed to delete generated dll files on Windows

  #50260 commented on Aug 1, 2024 • 0 new comments

• [ROCm] Tunableop record untuned

  #128813 commented on Aug 1, 2024 • 0 new comments

• Scale XBLOCK in triton reduction configs to avoid hitting max grid

  #128826 commented on Aug 1, 2024 • 0 new comments

• [reland][ROCm] TunableOp for gemm_and_bias

  #128919 commented on Aug 1, 2024 • 0 new comments

• Always use high precision for SDPA math backend

  #128922 commented on Aug 1, 2024 • 0 new comments

• Fix recent build error on ppc64le

  #129736 commented on Aug 1, 2024 • 0 new comments

• [BE][Easy][13/19] enforce style for empty lines in import segments in `test/j*/`

  #129764 commented on Jul 31, 2024 • 0 new comments

• [BE][Easy][17/19] enforce style for empty lines in import segments in `torch/[a-c]*/` and `torch/[e-n]*/`

  #129769 commented on Jul 31, 2024 • 0 new comments

• [BE][Easy][19/19] enforce style for empty lines in import segments in `torch/[o-z]*/`

  #129771 commented on Jul 31, 2024 • 0 new comments

• [Pipelining] Add schedule unshard/reshard pass

  #129810 commented on Aug 1, 2024 • 0 new comments

• Support built-in id function for TensorVariable on parameters

  #130100 commented on Aug 1, 2024 • 0 new comments

• Support XPU ABI=0 build

  #130110 commented on Aug 1, 2024 • 0 new comments

• Enable AOTI Eager to support int list

  #130231 commented on Jul 31, 2024 • 0 new comments

• [2/N][dtensor] Strided Sharding shard_to_replicate

  #130239 commented on Aug 1, 2024 • 0 new comments

• More appropriate socket errors and debug messages

  #130347 commented on Jul 31, 2024 • 0 new comments

• [pipelining] Add schedule send/recv pass

  #130378 commented on Aug 1, 2024 • 0 new comments

• AutoHeuristic: Introduce script to collect data for flex_attention

  #130398 commented on Aug 1, 2024 • 0 new comments

• flash_attn: limit compilation parallelism due to high memory requirements

  #130443 commented on Jul 31, 2024 • 0 new comments

• [ROCm][CK][Inductor] Enable addmm for CK backend to gemm max autotune

  #130576 commented on Jul 31, 2024 • 0 new comments

• [ROCm] ROCm triton pin update

  #130625 commented on Jul 31, 2024 • 0 new comments

• Gqa benchmark

  #130634 commented on Jul 31, 2024 • 0 new comments

• [MPS] Add support for MPS int4 (groupwise) and int8 (per channel) kernels starting with macOS 15.0

  #130715 commented on Aug 1, 2024 • 0 new comments

• [Inductor] support masked vectorization for the tail_loop of the 2d tiles kernel

  #130724 commented on Jul 31, 2024 • 0 new comments

• [FSDP][dtensor] use _StridedShard to represent nested sharding for correct full_tensor() result

  #130760 commented on Aug 1, 2024 • 0 new comments

• WIP implement batching of send/recv

  #130860 commented on Aug 1, 2024 • 0 new comments

• Support IPC for Expandable Segments

  #130890 commented on Aug 1, 2024 • 0 new comments

• Add decomposition for squeeze_copy

  #130941 commented on Jul 31, 2024 • 0 new comments

• Refactor process_inputs outside of create_aot_dispatcher_function

  #130962 commented on Jul 31, 2024 • 0 new comments

• [inductor] support vectorization for torch.any(bool) -> bool

  #131017 commented on Aug 1, 2024 • 0 new comments

• BuildExtension breaks build because of absolute path

  #132130 commented on Aug 1, 2024 • 0 new comments

• 关于yolov5在mac设备上使用pytorch的mps加速出现的各种问题

  #132226 commented on Aug 1, 2024 • 0 new comments

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

  #124245 commented on Aug 1, 2024 • 0 new comments

• Including AdaBound in the list of Optimizers.

  #46809 commented on Aug 1, 2024 • 0 new comments

• DISABLED test_comprehensive_nn_functional_nll_loss_cuda_float32 (__main__.TestDecompCUDA)

  #117732 commented on Aug 1, 2024 • 0 new comments

• OSError: [WinError 126] The specified module could not be found.

  #131662 commented on Aug 1, 2024 • 0 new comments

• xpu: efficientnet inference underperforms ipex

  #132176 commented on Aug 1, 2024 • 0 new comments

• M1 mps issue

  #89708 commented on Aug 1, 2024 • 0 new comments

• NCCL watchdog thread terminated with exception

  #113128 commented on Aug 1, 2024 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on Jul 31, 2024 • 0 new comments

• Automated submodule update: FBGEMM

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

• [draft] numpy 2.0.0rc1 test

  #121979 commented on Aug 1, 2024 • 0 new comments

• Add MaskedTensor passthrough: unfold, F.Unfold, F.Fold, stack

  #125262 commented on Jul 31, 2024 • 0 new comments

• [vision hash update] update the pinned vision hash

  #125806 commented on Aug 1, 2024 • 0 new comments

• refine fp32 precision api

  #125888 commented on Aug 1, 2024 • 0 new comments

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

  #126050 commented on Aug 1, 2024 • 0 new comments

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

  #126054 commented on Aug 1, 2024 • 0 new comments

• Introduce test skip markers for Sandcastle

  #126273 commented on Aug 1, 2024 • 0 new comments

• [inductor] Use full bit widths for bf16/fp16 vectorization

  #126502 commented on Jul 31, 2024 • 0 new comments

• [Inductor] support masked vectorization for the tail_loop

  #126526 commented on Jul 31, 2024 • 0 new comments

• [1/N][dtensor] introduce StridedShard placement type and _split_tensor() logic

  #126697 commented on Aug 1, 2024 • 0 new comments

• Default meta device to use swap_tensors in nn.Module._apply (.to_empty and .to('meta'))

  #126819 commented on Jul 31, 2024 • 0 new comments

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

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

• [inductor] enable bf32 test for mkldnn conv

  #127293 commented on Aug 1, 2024 • 0 new comments

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

  #127294 commented on Aug 1, 2024 • 0 new comments

• Updating Module Tracker

  #127624 commented on Jul 31, 2024 • 0 new comments

• [Testing only] Flip default on weights_only

  #127627 commented on Aug 1, 2024 • 0 new comments

• [MPS] Add native strided API for MPSNDArray starting with macOS 15

  #128393 commented on Aug 1, 2024 • 0 new comments