Issues

• [inductor] Odd-step linspace selects a different midpoint branch from eager CUDA

  bot-triagedmodule: correctness (silent)module: decompositionsoncall: pt2triage review
  pytorch/pytorch#202340 · Filter by author sw072 opened on Oct 28, 2026

• F.ctc_loss on CUDA with int32 targets and lengths (cuDNN): wrong, varying loss for non-contiguous log_probs; nonzero gradient past input_lengths

  bot-triagedmodule: correctness (silent)module: cudamodule: cudnnmodule: lossrelease triagetriage review
  pytorch/pytorch#202338 · Filter by author karanvora2599 opened on Oct 28, 2026

• torch.linspace midpoint differs between eager and compiled execution on ROCm

  bot-triagedmodule: decompositionsmodule: rocmoncall: pt2release triagetriage review
  pytorch/pytorch#202336 · Filter by author qtcdxxyuc opened on Oct 28, 2026

• [MPS] SiLU backward suffers cancellation for finite positive inputs

  bot-triagedmodule: correctness (silent)module: mpsmodule: numerical-stabilityrelease triagetriage review
  pytorch/pytorch#202335 · Filter by author quantumxiaol opened on Oct 28, 2026

• [docs] nn.init.xavier_*/kaiming_*: tensors with fewer than 2 dimensions are rejected, but the docs say "an n-dimensional torch.Tensor"

  bot-triagedmodule: docsmodule: initializationtriaged
  pytorch/pytorch#202333 · Filter by author Redempt1onzzZZ opened on Oct 28, 2026

• [docs] torch.nanmean: docs say "Input must be floating point or complex", but complex input is rejected on CPU

  bot-triagedmodule: complexmodule: cpumodule: docstriaged
  pytorch/pytorch#202332 · Filter by author Redempt1onzzZZ opened on Oct 28, 2026

• [docs] torch.median / nanmedian / prod: docs say dim=None reduces all dimensions, but passing dim=None raises

  bot-triagedmodule: docsmodule: python array apimodule: reductionstriaged
  pytorch/pytorch#202331 · Filter by author Redempt1onzzZZ opened on Oct 28, 2026

• DISABLED test_conv_bias_no_extra_cast_for_fp32 (__main__.CudaReproTests)

  bot-triagedmodule: inductormodule: xpuoncall: pt2skippedtriaged
  pytorch/pytorch#202330 · Filter by author chuanqi129 opened on Oct 28, 2026
  1

• B300 (sm_103): first call of torch.pow / index_put_ takes 5-12 s because libtorch_cuda.so has no sm_103 code

  bot-triagedenhancementmodule: binariesmodule: cudamodule: performancetriaged
  pytorch/pytorch#202323 · Filter by author henrylhtsang opened on Oct 28, 2026

• DISABLED test_cpp_custom_op_runtime_dispatch_stays_in_cpp (__main__.TestCppWrapperCustomOps)

  module: flaky-testsmodule: inductoroncall: pt2skippedtriaged
  pytorch/pytorch#202322 · Filter by author pytorch-bot opened on Oct 28, 2026
  1

• Allow selecting the varlen_attn backend

  bot-triagedfunction requestmodule: sdpatriaged
  pytorch/pytorch#202275 · Filter by author alanhdu opened on Oct 27, 2026

• Make varlen_attn visible to torch.compile

  bot-triagedenhancementmodule: custom-operatorsmodule: performancemodule: pt2-dispatcheroncall: pt2triaged
  pytorch/pytorch#202269 · Filter by author alanhdu opened on Oct 27, 2026
  1

• [Kineto] Timestamp issues in ApproximateClockToUnixTimeConverter

  bot-triagedoncall: profiler
  pytorch/pytorch#202265 · Filter by author ddelgadovargas-cyber opened on Oct 27, 2026

• DISABLED test_correctness_NAdam_use_closure_False_xpu_float32 (__main__.CompiledOptimizerParityTestsXPU)

  bot-triagedmodule: dynamomodule: pt2 optimizermodule: xpuoncall: pt2skippedtriaged
  pytorch/pytorch#202260 · Filter by author chuanqi129 opened on Oct 27, 2026
  1

• [torch.compile] Returned tensor loses storage aliasing after set_()

  bot-triagedmodule: correctness (silent)module: pt2-dispatcheroncall: pt2triaged
  pytorch/pytorch#202209 · Filter by author Guaiiii-a11y opened on Oct 27, 2026

• [torch.compile][PrivateUse1] _foreach_addcdiv_ meta/fake kernel rejects device-side scalar in capturable optimizer step

  bot-triagedmodule: dynamomodule: mtamodule: PrivateUse1module: pt2 optimizeroncall: pt2triaged
  pytorch/pytorch#202205 · Filter by author cheng1578361668 opened on Oct 27, 2026

• Use device-side assertions (CUDA_KERNEL_ASSERT2) in user-reachable index kernels

  bot-triagedenhancementmodule: cudamodule: rocmmodule: scatter & gather opstriaged
  pytorch/pytorch#202195 · Filter by author jeffdaily opened on Oct 27, 2026
  1

• torch.device("cpu:N") with N > 0 behaves differently in debug and release builds

  actionablebot-triagedlow prioritymodule: debug-buildmodule: error checkingtriaged
  pytorch/pytorch#202191 · Filter by author thisisatharva-rh opened on Oct 27, 2026
  1

• [FSDP2] set_gradient_divide_factor(f) divides by f twice when the reduce-scatter has one rank and the reduction forces sums or uses fp16/fp64

  bot-triagedmodule: fsdponcall: distributed parallelismsoncall: distributedtriage reviewtriaged
  pytorch/pytorch#202170 · Filter by author Robertboy18 opened on Oct 27, 2026

• torch.optim.Adam(capturable=True) fails CUDA graph capture when each beta is a 0-dim CUDA tensor

  actionablebot-triagedmodule: cuda graphsmodule: mtamodule: optimizertriaged
  pytorch/pytorch#202157 · Filter by author MogicianWu opened on Oct 26, 2026

• Windows cu128: single-token GQA SDPA stays on the math kernel

  bot-triagedenhancementmodule: performancemodule: sdpamodule: windowstriaged
  pytorch/pytorch#202156 · Filter by author AdventureBoi1 opened on Oct 26, 2026
  2

• CUDA sparse CSR matmul reads freed memory when values() is non-contiguous: wrong results from mv/mm/addmv/addmm

  bot-triagedmodule: correctness (silent)module: cudamodule: sparsetriage review
  pytorch/pytorch#202134 · Filter by author pearu opened on Oct 26, 2026

• CUDA addmm with sparse CSR input, mat1, mat2 gives wrong values when beta != 0 and the sparsity patterns differ

  bot-triagedmodule: correctness (silent)module: cudamodule: sparsetriage review
  pytorch/pytorch#202128 · Filter by author pearu opened on Oct 26, 2026
  

• [RFC] Input/weight backward split for torch.compile'd regions (zero-bubble schedules and DualPipeV)

  bot-triagedfeaturemodule: aotdispatchmodule: performancemodule: pipeliningmodule: pt2-dispatcheroncall: distributed parallelismsoncall: distributedoncall: pt2triaged
  pytorch/pytorch#202127 · Filter by author philliptay opened on Oct 26, 2026

• to_sparse_csr()/to_sparse_bsr()/mm give wrong results or corrupt the heap for a coalesced COO tensor with non-contiguous indices

  bot-triagedmodule: correctness (silent)module: crashmodule: sparsetriage review
  pytorch/pytorch#202122 · Filter by author pearu opened on Oct 26, 2026