Platforms: xpu This test was disabled because it is failing on main branch (recent examples). cc @mruberry @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben @jataylo @gujinghui @fe...

Platforms: inductor This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME...

Platforms: linux This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 7 workflow(s) with 7 failures and 7 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THI...

Platforms: dynamo, linux, mac, macos, rocm, slow, asan This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 4 workflow(s) with 12 failures and 4 successes. Debugging instructions (after clicking on the...

Platforms: linux, mac, macos, rocm, asan, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 7 workflow(s) with 21 failures and 7 successes. Debugging instructions (after clicking on the recent s...

Platforms: dynamo This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME TH...

🐛 Describe the bug Hello @janeyx99 I am one of owners of https://github.com/amd/quark and I am super excited in converting our runtime-compiled Torch C++ extensions into pre-compiled libraries shipped with Quark using the PyTorch's stable ABI When I checked our code that uses torch C++, I have identified APIs that i...

We are performing B200 runner maintenance one by one for the six runners. Jobs may encounter failures during this time and may need to be restarted. cc @malfet @pytorch/pytorch-dev-infra @huydhn @atalman @ptrblck @eqy @tinglvv

🐛 Describe the bug The torch==2.13.0+rocm7.2 wheel fails the torch.compile smoke test with RuntimeError: Can't detect vectorized ISA for CPU when it is imported/run on a non-ROCm image (e.g. pytorch/almalinux-builder:cpu-main), where torch.cuda.is_available() is False and the compile check exercises the CPU / Induct...

🐛 Describe the bug I first reported this privately as a security advisory (GHSA-ppvg-47fv-5f4w). It was closed under the SECURITY.md policy. In that thread I asked whether it might still be worth tracking as a documentation/API-hardening issue and whether you would prefer that I open a regular issue; I did not recei...

🚀 The feature, motivation and pitch While porting torchvision to stable ABI, I noticed that ops that return nothing and exist only for their side effects cannot be registered safely. StableLibrary::def only takes a schema string so the op gets FROM_SCHEMA alias analysis, and TorchScript DCE deletes calls to an op wi...

🐛 Describe the bug The documentation of asarray states: device (torch.device, optional) – the device of the returned tensor. Default: None, which causes the device of obj to be used. Or, if obj is a Python sequence, the current default device will be used. The described behaviour is coherent with the specification o...

🐛 Describe the bug This issue is to have a centralized place to list and track work on adding support to new ops for the MPS backend. Most requested ops (extracted from comments to this issue and #77764) PyTorch MPS Ops Project : Project to track all the ops for MPS backend. There are a very large number of operator...

🐛 Describe the bug Reproducible Example import torch def main(): pool = torch.nn.MaxPool2d(2, return_indices=True) out, indices = pool(torch.randn(1, 10, 10, 10)) unpool = torch.nn.MaxUnpool2d(2) _ = torch.jit.trace(unpool, (out, indices)) if __name__ == "__main__": main() error message: TypeError: cond must be a bo...

🐛 Describe the bug When utilizing torch.compile with graph caching enabled (TORCHINDUCTOR_FX_GRAPH_CACHE=1) on Python 3.13 and PyTorch 2.14 Nightly, the compilation disk cache subsystem crashes during the serialization phase of tensor subclass metadata. Specifically, when a custom tensor subclass (such as an advance...

🐛 Describe the bug Capturing N training iterations (forward + backward + AdamW(capturable=True).step()) of TWO independent trainers into one torch.cuda.CUDAGraph produces a graph whose replay crashes. In the script below each trainer owns M=4 plain nn.Modules called in a Python loop - no functorch anywhere - and the...

Platforms: linux, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 6 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NO...

Platforms: dynamo, linux, mac, macos, rocm, slow, asan This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 8 workflow(s) with 24 failures and 8 successes. Debugging instructions (after clicking on the...

Job vllm-test / vllm-x-pytorch-test / test (vllm_multi_model_test_28_failure_test, 1, 1, linux.g6.4xlarge.experimental.nvidia.gpu, unstable) Symptom 13 failed / 23 passed / 29 skipped. All 13 failures are vLLM multimodal pooling model tests, and every one dies the same way: vLLM EngineCore fails to start with Runtim...

Possible / low-confidence. add_M64_N64_K*_cpu_bwd*_BACKWARD, 9 cases, ~200-400x slower vs the prior baseline, stable across runs. Flagged as low-confidence: a real 400x regression in add backward would be catastrophic and caught in training, so this looks more like a backward-pass measurement artifact than a genuine...

Possible regression. embedding and embeddingbag, ~48 cases, ~70-90x slower vs the prior baseline, stable across runs. Flagged as possible because embeddingbag time is roughly flat vs bag count (10 vs 2300 bags both ~8500us), which suggests the number may be dominated by a fixed per-call cost rather than the embeddin...

Possible / low-confidence. Two isolated single-shape cases flagged by the op-benchmark CPU rebaseline, each stable across two runs but not part of a broader family pattern. index_add__M8_N32_K1_dim0_cpu_dtypetorch.float32 -- ~2.4x baddbmm_B128_M64_N32_K64_cpu_dtypetorch.float32 -- ~2.0x. Note: other baddbmm shapes a...

Two quantization families flagged by the rebaseline, both stable across runs. qcat (quantized cat), 18 cases, ~3-4x slower -- likely regression. fake_quantize per-tensor/channel (original_kernel/learnable_kernel), ~700x slower -- possible regression or measurement artifact; the magnitude is extreme and needs an owne...

🐛 Describe the bug TL;DR On the _PipelineScheduleRuntime path (which ScheduleInterleaved1F1B now uses), the communication lowering places each RECV_F at the position where its sender is scheduled by a dependency simulation, not at the timestep where the data is actually produced/needed. https://github.com/pytorch/py...

Likely regression. 8 cases of eager transcendental unary math are ~3-5x slower vs the prior baseline, stable across two runs. Suspected: the vectorized/Sleef CPU math path. Please confirm. sinh_M512_N512_cpu old 81.5 -> 432.0 / 388.9 (~5x) log10__M512_N512_cpu old 11.8 -> 46.2 / 45.1 (~4x) log2__M512_N512_cpu old 11...

Reason MI300 cluster is having outages, hence capacity is reduced, leading to queueing on 4-GPU MI300 runners. cc @jeffdaily @sunway513 @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @malfet @pytorch/pytorch-dev-infra

New Feature for Release Composable Kernel (CK) plans to upgrade its language standard from C++17 → C++20 in the next few months. This will change our default CMake from C++17 to C++20. We’d like to confirm there are no blockers for PyTorch builds and downstreams. Point(s) of contact ThomasNing Release Mode (pytorch/...

🐛 Describe the bug On MPS, F.embedding / nn.Embedding / torch.index_select silently return all-zero rows (no error, no NaN) for every index i >= ceil(2**31 / dim) when the source weight tensor has more than 2**31 elements. Indices below that boundary return bit-exact rows, so the corruption is partial and easy to mi...

This test was disabled because it is failing on main branch (recent examples). cc @chauhang @penguinwu

Platforms: linux This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 0 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THI...

🐛 Describe the bug The nn.functional.threshold OpInfo reference in torch/testing/_internal/common_methods_invocations.py fails for unsigned dtypes under NumPy >= 2.5. test_reference_numerics_{small,normal}_nn_functional_threshold_{cpu,cuda}_uint8 (and the _refs variants) fail with: OverflowError: Python integer -9 o...

🐛 Describe the bug If the process holding the lock is forcefully terminated (e.g., by OOM Killer, SIGKILL, Ctrl+C, or Slurm job cancellation), the .lock file is never cleaned up. On subsequent runs, all processes detect the stale .lock file and enter an infinite wait loop, causing a permanent deadlock. To Reproduce...

🐛 Describe the bug torch.compile fails with InductorError: KeyError in Scheduler.compute_ancestors when a custom op mutates freshly-allocated (torch.empty) output buffers in place and those buffers are then reshaped in a way Inductor cannot prove is a free view (under dynamic shapes). This is the shape of a flash_at...

Summary kernel_information.json (#160540, extended in #183952) and the profiler-timeline provenance (#186230, #187328) map each generated kernel to a set of pre/post-grad node names, stack traces, and now extern I/O shapes. What is still missing is the physical, per-kernel op structure: the ordered list of ops that...

🐛 Describe the bug cd test/onnx python test_pytorch_onnx_onnxruntime_cuda.py -k test_arithmetic_bfp16 onnxruntime.capi.onnxruntime_pybind11_state.NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for Add(14) node with name '/Add' Versions torch 2.7.1 onnx 1.17.0 onnxruntime...

🐛 Describe the bug torch.fft.irfft2 ONNX export now succeeds on nightly (fixed in #164698), but the exported graph cannot be run by onnxruntime in practice. PyTorch's dynamo exporter currently produces ONNX IR version 10 graphs. The newest onnxruntime release that supports CUDA 11.8 is onnxruntime-gpu==1.16.3, which...

Reason MI300 cluster is having outages, hence capacity is reduced, leading to queueing on 4-GPU MI300 runners. cc @jeffdaily @sunway513 @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @malfet @pytorch/pytorch-dev-infra

🐛 Describe the bug When trying to compile a simple function that uses cpu tensors, torch inductor initialises a context on cuda:0. If used in a multiprocessing context (E.G.) a torch dataloader, this quickly results in OOM (simple repro example allocates 500MB of VRAM per process). CUDA_VISIBLE_DEVICES="" works as a...

Platforms: rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 0 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THIN...

Platforms: asan, linux, mac, macos, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 14 workflow(s) with 42 failures and 14 successes. Debugging instructions (after clicking on the recent...

Platforms: linux, slow, rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 6 workflow(s) with 12 failures and 6 successes. Debugging instructions (after clicking on the recent samples link): DO N...

🕵️‍♂️ Detected with FACTO 🐛 Describe the bug As the title says: import torch def test_argmax(x): device = x.device try: output = torch.argmax(x) print(f" argmax test succeeds for device: {device}. output: {output}") except Exception as e: print(f" argmax test fails for device: {device}: {e}") x_cpu_contiguous = torc...

🐛 Describe the bug We tried to leverage per_channel quantization in QAT and exported the trained model in onnx format. model = dummy pytorch model export_model = torch.export.export_for_training( model, example_inputs).module() quantizer = XNNPACKQuantizer().set_global(get_symmetric_quantization_config( is_per_chann...

Several MPS ops still dispatch through MPSGraph (LookUpOrCreateCachedGraph), which compiles and caches a graph per input shape with no eviction. On shape-varying workloads -- GNNs, variable-length attention, hyperparameter sweeps -- the cache grows without bound and the process eventually OOMs on Apple Silicon. (tor...

🐛 Describe the bug Using the latest pytorch nightly, torch.compile throws an error with torchao mxfp8 MoE training code. Repro Use machine with access to local B200 GPUs Clone https://github.com/pytorch/ao Install from source: USE_CUDA=1 pip install -e . --no-build-isolation Run benchmark script: rm -rf /tmp/torchin...

Platforms: rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 4 workflow(s) with 4 failures and 4 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THIN...

Platforms: linux, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 4 workflow(s) with 8 failures and 4 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSU...

Platforms: dynamo, asan, linux, mac, macos, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 4 workflow(s) with 12 failures and 4 successes. Debugging instructions (after clicking on the...

Platforms: mac, macos This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 4 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUM...

Platforms: asan, linux, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NO...

Platforms: dynamo This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 12 workflow(s) with 12 failures and 12 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME...

🐛 Describe the bug decide_layout_opt() only counts aten.convolution.default nodes when deciding whether to enable channels-last layout optimization. When AOTAutograd splits a model into separate forward and backward graphs, the backward graph may contain only convolution_backward without a corresponding forward conv...

🚀 The feature, motivation and pitch PyTorch's current 'reflect' padding_mode uses half-sample reflection, i.e. it does not repeat the boundary sample: input: ABCDEF reflect: DCB|ABCDEF|EDC (boundaries not repeated) I'd like to add a 'symmetric' padding_mode that does repeat the boundary sample: input: ABCDEF symmetr...

🐛 Describe the bug When compiling an attention-like function containing complex numbers (torch.polar, torch.view_as_complex) and batched matrix multiplication (torch.matmul) on CPU with torch.compile(mode="max-autotune"), the compiler crashes with an AssertionError. The crash occurs during the C++ backend's BMM/GEMM...

🚀 The feature, motivation and pitch Summary When autotuning is enabled for GEMM kernels, both the TMA (Tensor Memory Accelerator) path and the non-TMA path are included as candidates and benchmarked against each other. This ensures the fastest path is always selected based on actual measured performance. However, fo...

🐛 Describe the bug symm_mem::two_shot_all_reduce_ (and the one-shot / multimem variants) synchronize the group with barriers that require the device execution order of the ops to match their group issue order. CUDA does not order kernels across streams, so issuing these ops from more than one stream — combined with...

🐛 Describe the bug transformers.cache_utils.DynamicCache inherits from torch.nn.Module. It seems to confuse torch.export.export and gives the following error: File ".../site-packages/torch/export/_trace.py", line 1697, in _export_to_aten_ir_make_fx raise UserError(UserErrorType.CONSTRAINT_VIOLATION, str(e)) # noqa:...

This PR #184567 introduced regression in a few cases for flex-attn-causal with triton, for example: H_q=16, H_kv=2, N_CTX_q=1, N_CTX_kv=1024, D_HEAD_qk=128, D_HEAD_v=128 case dropped from 0.81 -> 0.41 TFLOPS More details: intel/intel-xpu-backend-for-triton#7302 TL;DR; 18MB of cache clear between runs seems like not...

Platforms: xpu This test was disabled because it is failing on main branch (recent examples). cc @gujinghui @EikanWang @fengyuan14 @guangyey cc @gujinghui @EikanWang @fengyuan14 @guangyey

Platforms: xpu This test was disabled because it is failing on main branch (recent examples). cc @gujinghui @EikanWang @fengyuan14 @guangyey

This PR can be subscribed by anyone who is interested. Autorevert bot should post here, as comments, all its Autorevert decisions while running in shadow mode. This is, anytime Autorevert decided to act to revert a PR (this means faster than a human). The goal of this PR is to enable the team to quickly troubleshoot...

Platforms: rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 0 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THIN...

Platforms: asan, linux, mac, macos, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 26 workflow(s) with 78 failures and 26 successes. Debugging instructions (after clicking on the recent...

Platforms: linux, slow, rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 9 workflow(s) with 18 failures and 9 successes. Debugging instructions (after clicking on the recent samples link): DO N...

Platforms: dynamo This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 7 workflow(s) with 7 failures and 7 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME TH...

Problem statement Today, if intermediary Tensors are saved during the forward and used during the backward, the double backward will be silently wrong because their contribution will be ignored. A small repro is: import torch from torch.autograd import Function, gradcheck, gradgradcheck class MyFn(Function): @static...

torch.compile recompiles when input alternates between AsyncCollectiveTensor and Tensor Summary When a compiled region's input alternates between AsyncCollectiveTensor (ACT) and a resolved Tensor across steps, Dynamo guards on the tensor class and recompiles each time the class changes. With FSDP + TP this happens d...

🐛 Describe the bug I am trying to export an nn.Module that involves manipulation of complex-valued tensors with torch.export.export and then compiling with torch._inductor.aoti_compile_and_package. Compilation succeeds but I run into errors when loading the package and running. In the following minimal reproducing e...

🐛 Describe the bug I thought #180566 would fix it but I was wrong. A live torch.compile'd module keeps weakrefs on its parameters (Dynamo guards). swap_tensors() refuses to operate on tensors that have weakrefs. The follow should reproduce the issue: """Minimal reproducer for why group offloading cannot use torch.ut...

🚀 Feature I would like to have a MultiScheduler that allows me to use multiple schedulers after each other, with the same optimizer. Motivation It is not uncommon to have a learning rate schedule that looks like the following example: First have a burn in learning rate for 1000-2000 batches (eg. quadratic or exponen...

🐛 Describe the bug Consider this code: import torch from torch._inductor.utils import fresh_cache def my_addmm(x, a, b, alpha, beta): return torch.addmm(x, a, b, alpha=beta, beta=alpha) x = torch.randn((128, 128)).cuda().half() a = torch.randn(128, 128).cuda().half() b = torch.randn(128, 128).cuda().half().t() with...

Hi PyTorch team, I am working on a research prototype called TopologyAwareJointOptimizer. I would like feedback on whether topology-aware memory optimization is a direction that should be explored around torch.compile, torch.nn, torchao, or an out-of-tree PyTorch backend. The problem I am trying to address is traini...

🚀 The feature, motivation and pitch Redundant NaN Propagation in max op in non-onlined Softmax Issue Current Inductor softmax can do redundant NaN propagation in the softmax-internal max reduction. For the fallback/persistent softmax form, the computation is effectively: xmax = max(x, dim=-1, keepdim=True) xexp = ex...

🐛 Describe the bug tools.stats.upload_stats_lib._download_artifact() is intended to skip GitHub-hosted fallback artifacts whose embedded runattemptN token does not match the requested workflow_run_attempt. Today the mismatch branch only prints a Skipping ... message. After that, the function continues to download th...

Platforms: rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 4 workflow(s) with 4 failures and 4 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THIN...

Platforms: linux, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 12 workflow(s) with 24 failures and 12 successes. Debugging instructions (after clicking on the recent samples link): DO...

Platforms: inductor, rocm This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT A...

Platforms: asan, dynamo, linux, mac, macos, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 10 workflow(s) with 30 failures and 10 successes. Debugging instructions (after clicking on th...

Platforms: linux This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME THI...

Platforms: dynamo This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME TH...

TL;DR This sub-RFC addresses the third slice of the umbrella RFC: providing registration interfaces for Inductor codegen so that a PrivateUse1 backend can be auto-discovered, configured, compiled, and run without monkeypatching upstream Inductor code. Four things are addressed: (1) backend discovery via entry-point...

TL;DR This sub-RFC addresses the second slice of the umbrella RFC: making Dynamo's six hardcoded lookup tables recognize PrivateUse1 device types and runtime functions by splicing them from the device interface registry. We introduce get_interface_for_device__PrivateUse1() — an iterable that returns only PrivateUse1...

TL;DR This sub-RFC addresses the first slice of the umbrella RFC: making torch.compile query the DeviceInterface contract for device identity (is_gpu) and capabilities (is_triton_capable, get_multi_processor_count, exposes_streams) instead of hardcoded cuda/xpu/mps lists. We introduce register_interface_for_device_P...

TL;DR This RFC document aims to propose and discuss making PrivateUse1 (out-of-tree) backends first-class citizens in torch.compile. Our focus is on eliminating the hardcoded cuda/xpu/mps device lists and if-branches that permeate the Dynamo and Inductor codebase, replacing them with a unified registry-based contrac...

🐛 Describe the bug openai/CLIP#227 torch.jit.load fails when path contains non-ascii characters import torch torch.jit.load('C:\\Users\\流星暴雨/.cache/clip\\ViT-B-16.pt', "cpu") Traceback (most recent call last): File "<string>", line 1, in <module> File "[C:\Users\]()流星暴雨\AppData\Local\Programs\Python\Python39\lib\sit...

🐛 Describe the bug torch==2.12.1+cu130 pins nvidia-cudnn-cu13==9.20.0.48, whose libcudnn_graph.so.9 dlopens libcudnn_engines_tensor_ir.so by bare soname - but that wheel doesn't ship it (it's only in ≥9.23). With no other cuDNN around this fails silently and falls back. But if a different system cuDNN 9.x is on the...

Platforms: mac, macos This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 19 workflow(s) with 38 failures and 19 successes. Debugging instructions (after clicking on the recent samples link): DO NOT AS...

Platforms: linux, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 3 workflow(s) with 6 failures and 3 successes. Debugging instructions (after clicking on the recent samples link): DO NO...

Platforms: linux, rocm, asan, mac, macos, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 11 workflow(s) with 33 failures and 11 successes. Debugging instructions (after clicking on the recent...

Platforms: dynamo This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 6 hours, it has been determined flaky in 9 workflow(s) with 9 failures and 9 successes. Debugging instructions (after clicking on the recent samples link): DO NOT ASSUME TH...

Platforms: asan, linux, rocm, slow This test was disabled because it is failing in CI. See recent examples and the most recent trunk workflow logs. Over the past 3 hours, it has been determined flaky in 2 workflow(s) with 2 failures and 2 successes. Debugging instructions (after clicking on the recent samples link):...

🐛 Describe the bug Minimal repro: Inductor FX graph cache reuses rank-specific code. Run twice with two GPUs. The first run cold-compiles and should pass; the second run reloads the shared warm FX graph cache and should fail because one rank can load the other rank's device/rank-specific generated wrapper. rm -rf /t...

🐛 Describe the bug Calling the forward method on TorchScript models can, under some specific conditions, raise a RuntimeError with message: "Global alloc not supported yet.". I think this is linked to an old issue: #69078, however, I managed to consistently reproduce this error. The code that reproduces the bug is q...

Platforms: xpu This test was disabled because it is failing on main branch (recent examples). cc @gujinghui @EikanWang @fengyuan14 @guangyey cc @mruberry @chauhang @penguinwu @gujinghui @EikanWang @fengyuan14 @guangyey

🐛 Describe the bug For a family of deep fp16 inference models, torch.compile on MPS deterministically crashes the process (SIGSEGV, exit 139) whenever the compile mode enables triton.cudagraphs, and works (bit-exact correct outputs) whenever it does not: configuration triton.cudagraphs result eager — ✅ OK torch.comp...

🐛 Describe the bug terminate called after throwing an instance of 'c10::Error' what(): inserted INTERNAL ASSERT FAILED at "/pytorch/c10/cuda/CUDACachingAllocator.cpp":3252, ... Versions curl -sL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py | python command returns: Collecting env...

🐛 Describe the bug For baddbmm with a broadcast bias (bias[B, 1, N], stride [N, 0, 1] after broadcasting), Inductor's max-autotune chooses between (a) Triton templates and (b) the extern ATen kernel — but never considers the decomposition bmm(x, w) + bias, which is faster than both of its existing choices whenever t...

🐛 Describe the bug Follow-up to #181789 (related: #172819 / #173391): one more case where _dynamic_scale_rblock mutates the autotune candidate set after the cache contract is frozen, defeating autotune_local_cache=True. Verified on current nightly. check_autotune_cache only creates an AutotuneCache when there is mor...

🐛 Describe the bug Calling torch.linalg.qr on a float64 matrix repeatedly in a loop causes the process to be killed by SIGSEGV (exit code 139) after ~31 iterations on ROCm 7.x (AMD Instinct MI300A). No Python exception is raised — the process dies silently with no traceback. The crash is specific to torch.linalg.qr...

🚀 The feature, motivation and pitch Allow CUDA-like PrivateUse1 backends to opt into existing CUDA optimization paths in shared at::native operators via an environment variable, avoiding code duplication and performance loss. 🚀 Feature Both this issue and #176715 share the same goal: third-party privateuse1 backends...