Attention Queue

Sorted by maintainer attention signal, still read-only and evidence-bound.

ItemScoreBucketNext ActionStale Days
issue #188812: [MPS] MetalShaderLibrary::getBundledLibrary singleton can crash at process exit151needs design decision
confirm whether existing evidence is enough; otherwise wait for a reproducer
4
issue #188900: `sparse.mul`: broadcasting a size-1 dimension of the sparse operand silently drops data139needs design decision
review early; this can change release or regression risk
2
issue #188323: [Inductor][CPU] dynamic=True convolution lowering crashes with ValueError: Exponent must be non-negative139needs design decision
review early; this can change release or regression risk
8
issue #186535: Windows, gloo: Access violation (0xC0000005) in ProcessGroupGloo::enqueue when calling allreduce on CUDA tensors — GlooAllreduceRegistry has no kCUDA creator139needs design decision
review early; this can change release or regression risk
31
issue #187912: [CPU] Concurrent `cpublas::brgemm` calls can crash in the AMX path when the underlying oneDNN ukernel is shared127needs design decision
confirm whether existing evidence is enough; otherwise wait for a reproducer
2
issue #157668: NCCL error caused due to use of NVLS in torch 2.7.1-cu128 on aarch64 gb200 cluster127needs design decision
confirm whether existing evidence is enough; otherwise wait for a reproducer
29
issue #150022: Dynamic Shapes with **kwargs121needs design decision
review early; this can change release or regression risk
216
issue #162422: Runtime failure when running torch.compile() & using GCC 11.5.0 on Neoverse V1121needs design decision
review early; this can change release or regression risk
220
issue #162476: heap-buffer-overflow in torch.quantized_max_pool2d via Python API121needs design decision
review early; this can change release or regression risk
300
issue #189194: [ROCm] torch 2.13 wheel: "Can't detect vectorized ISA for CPU" in torch.compile smoke test on non-ROCm image (regression vs 2.12.1)115needs design decision
review early; this can change release or regression risk
0
issue #189150: SIGSEGV / cudaErrorIllegalAddress when replaying a CUDA graph that captured multiple training iterations of two trainers over deeply unrolled recurrent modules115needs design decision
review early; this can change release or regression risk
0
issue #189282: [CPU] operator_benchmark: backward-pass (add, + batchnorm on aarch64) ~200-500x slower since ~Nov 2025 (x86 + aarch64; possible measurement artifact)115needs design decision
review early; this can change release or regression risk
0
issue #189281: [CPU] operator_benchmark: embedding / embeddingbag ~70-130x slower since ~Nov 2025 (x86 + aarch64; possible measurement issue)115needs design decision
review early; this can change release or regression risk
0
issue #189239: [Inductor]Unable to apply layout optimization on convolution backward causes training performance regression.115needs design decision
review early; this can change release or regression risk
0
issue #189144: [XPU][B580] `flex-attn-causal` performance drop with `last_level_cache_size` cache clear between runs115needs design decision
review early; this can change release or regression risk
1
issue #144965: RuntimeError "global alloc not supported yet" when using TorchScript optimization.115needs design decision
review early; this can change release or regression risk
1
issue #189126: Report a bug, as requested by the error message115needs design decision
review early; this can change release or regression risk
1
issue #189065: CI for distributed tests on B200 fully red since ~2026-07-01: rank 0 "CUDA-capable device(s) busy or unavailable" -> NCCL id timeout -> 1170-min job timeout115needs design decision
review early; this can change release or regression risk
2
issue #170003: batch_isend_irecv with nccl causes illegal memory access depending on P2P ordering115needs design decision
review early; this can change release or regression risk
23
issue #174288: [distributed] Batched isend/irecv with NCCL backend hangs on high load115needs design decision
review early; this can change release or regression risk
26
issue #162731: DTensor cached op propagation results can be mutated when propagating other ops115needs design decision
review early; this can change release or regression risk
27
issue #153517: [CI][CUDA][Distributed] test_non_blocking_with_eager_init timeout115needs design decision
review early; this can change release or regression risk
28
issue #167693: DISABLED test_side_stream_backward_overlap_cuda (__main__.TestAutogradStreamSynchronizationCUDA)115needs design decision
review early; this can change release or regression risk
28
issue #116423: PyTorch Distributed Elastic Launch Segmentation Fault with Python 3.12115needs design decision
review early; this can change release or regression risk
31
issue #106164: distributed.batch_isend_irecv() crash when send/recv refers to itself115needs design decision
review early; this can change release or regression risk
31
issue #85088: reentrant torch.utils.checkpoint does not work with NamedTuple outputs115needs design decision
review early; this can change release or regression risk
31
issue #71303: [RFC] Cross-Process Performance Analysis: Straggler Detection115needs design decision
review early; this can change release or regression risk
32
issue #75725: Warning originating in C10 backend does not get translated to Python warning if run from subprocess115needs design decision
review early; this can change release or regression risk
32
issue #76449: Enhance _verify_param_shape_across_processes115needs design decision
review early; this can change release or regression risk
32
issue #81684: Message exchange failure when perform alltoallv (cpus)115needs design decision
review early; this can change release or regression risk
33
issue #119845: Segmentation fault in dataloader worker sub-process115needs design decision
review early; this can change release or regression risk
33
issue #183459: Regression in torch.distributed._functional_collectives.all_to_all_single in 2.9.1 -> 2.11115needs design decision
review early; this can change release or regression risk
33
issue #180088: [DTensor] RNG tracker does not advance state for CPU tensors on CUDA mesh, causing trunc_normal_ infinite loop115needs design decision
review early; this can change release or regression risk
41
issue #173921: [Inductor] Significant numerical divergence (4.7% relative error) in Conv2d on CPU with torch.compile115needs design decision
review early; this can change release or regression risk
48
issue #179502: [DTensor] sharded view incorrectly passes when redistribution is needed115needs design decision
review early; this can change release or regression risk
93
issue #145610: mmap fails on 64k page aarch64 systems for AOTI model loading115needs design decision
review early; this can change release or regression risk
141
issue #66504: BatchNorm runtimeError: one of the variables needed for gradient computation has been modified by an inplace operation115needs design decision
review early; this can change release or regression risk
144
pr #188632: Fix Dynamo opaque object staticmethod guards112ready for maintainer decision
final maintainer merge/release decision
0
pr #186429: Fix ONNX export of pad_sequence with symbolic split lengths112ready for maintainer decision
final maintainer merge/release decision
0
pr #186461: Fix SymBool equality handling in symbolic shapes112ready for maintainer decision
final maintainer merge/release decision
0
pr #184716: Fix conv_transpose2d meta output padding validation112ready for maintainer decision
final maintainer merge/release decision
0
pr #185691: Avoid UB in float to signed integer casts112ready for maintainer decision
final maintainer merge/release decision
1
pr #184939: Support SymInt steps for linspace/logspace export112ready for maintainer decision
final maintainer merge/release decision
20
pr #189307: [CI/CD] Copy newer CUPTI headers into manywheel binary-build images108waiting on CI/check fix
final maintainer merge/release decision
0
pr #186927: [MPS] Gemv kernels108waiting on CI/check fix
final maintainer merge/release decision
0
pr #189312: [CUDA][cuBLAS] Change cuBLAS default workspace size for SM 11.0 to 32 MiB108waiting on CI/check fix
final maintainer merge/release decision
0
pr #188301: [inductor] Fix inductor dropping ordering dep between effectful ops with different kernel types108waiting on CI/check fix
final maintainer merge/release decision
0
pr #184648: [Test] Add has_sufficient_memory() hook to DeviceTypeTestBase108waiting on CI/check fix
final maintainer merge/release decision
1
pr #188237: Optimize CSR SpMM CPU grain size and inner accumulation108waiting on CI/check fix
final maintainer merge/release decision
1
pr #187501: [Stable C Shim] Use error message retrieval shim if available at runtime108waiting on CI/check fix
final maintainer merge/release decision
13
pr #185882: [PrivateUse1] Return `False` instead of `None` when PU1 backend is not available108waiting on CI/check fix
final maintainer merge/release decision
15
issue #153410: [Feature request] `torch.export` .save/.load could support `safetensors` and/or `weights_only=True`107needs design decision
review linked PR state before touching issue
171
issue #188892: nvidia-cudnn-cu13==9.20.0.48 bundled by torch is incomplete → CUDNN_STATUS_SUBLIBRARY_VERSION_MISMATCH106has linked PR
confirm whether existing evidence is enough; otherwise wait for a reproducer
1
issue #188970: [MPS] SIGSEGV in MPSStream::copy with triton.cudagraphs (reduce-overhead/max-autotune) on deep fp16 graphs106has linked PR
confirm whether existing evidence is enough; otherwise wait for a reproducer
1
pr #135631: [scan] Autograd100waiting on CI/check fix
final maintainer merge/release decision
521
pr #188115: [Inductor][TEST] Align `test_main_loop_scaling` with H100 support surface98waiting on CI/check fix
identify whether block is CI, merge conflict, or review gate
6
issue #149447: DTensor slicing on sharded dimension leads to replication97needs design decision
review early; this can change release or regression risk
200
issue #164929: get_optimizer_state_dict modifies optimizer state97needs design decision
review early; this can change release or regression risk
211
issue #156563: Segmentation fault (core dumped) in `torch.profiler.profile`97needs design decision
review early; this can change release or regression risk
231
issue #122019: `F.max_pool3d`: Segfault due to incorrect `dilation` check in kernel97needs design decision
review early; this can change release or regression risk
285
issue #159238: Performance regression: torch.jit.trace() significantly slower on RTX 5090 than RTX 4060 (cu128 nightly)97needs design decision
review early; this can change release or regression risk
333
issue #121219: Torch profiler corrupted names with Python 3.1197needs design decision
review early; this can change release or regression risk
335
issue #91165: [FSDP] FSDP with CPU offload consumes `1.65X` more GPU memory when training models with most of the params frozen97needs design decision
review early; this can change release or regression risk
339
issue #47260: DDP doesn't work with retain_graph = True97needs design decision
review early; this can change release or regression risk
394
issue #100253: profiler.export_stacks doesn't return stack trace unless experimental_config is provided97needs design decision
review early; this can change release or regression risk
395
issue #155147: torch.profiler raises Aborted (core dumped) failurer related with GIL (gilstate_tss_set)97needs design decision
review early; this can change release or regression risk
397
issue #155146: Segmentation fault when using torch.profiler97needs design decision
review early; this can change release or regression risk
398
issue #153941: TorchScript Fails Type Inference When Using .item()-Derived Scalar in Tensor Construction97needs design decision
review early; this can change release or regression risk
414
issue #153260: Export shouldn't run TS under the hood.97needs design decision
review early; this can change release or regression risk
421
issue #153145: Segmentation faults with CPU FlexAttention97needs design decision
review early; this can change release or regression risk
425
issue #152108: [inductor][cpu] AMP static shape default wrapper AOTInductor performance regression in 2025_04_20 nightly release97needs design decision
review early; this can change release or regression risk
440
issue #145088: [torchbench] torch._dynamo.exc.Unsupported: Graph break due to unsupported builtin None.morphologyEx97needs design decision
review early; this can change release or regression risk
490
issue #146788: Segmentation Fault in `torch.choose_qparams_optimized` with Invalid Parameters97needs design decision
review early; this can change release or regression risk
514
issue #124950: [distributed] First NCCL barrier does not respect timeout97needs design decision
review early; this can change release or regression risk
525
issue #71855: Regression in multi-node training speed with Transformers + PyTorch97needs design decision
review early; this can change release or regression risk
588
issue #136712: Segmentation fault (core dumped) in `torch.ao.nn.quantized.dynamic.LSTMCell/GRUCell`97needs design decision
review early; this can change release or regression risk
650
issue #136728: Aborted (core dumped) in `torch.package.package_exporter.PackageExporter`/`torch.package.PackageExporter`97needs design decision
review early; this can change release or regression risk
650
issue #135835: A segmentation fault will be raised when using the `torch._C._jit_to_backend`97needs design decision
review early; this can change release or regression risk
664
issue #81085: RuntimeError: required keyword attribute 'value' is undefined97needs design decision
review early; this can change release or regression risk
667
issue #128693: Segmentation fault (core dumped) in `torch.fused_moving_avg_obs_fake_quant`97needs design decision
review early; this can change release or regression risk
716