Issues

• [docs] RecordingObserver docstring documents nonexistent qscheme and reduce_range args

  bot-triagedoncall: quantization
  pytorch/pytorch#200566 · Filter by author alwaysprince05 opened on Oct 11, 2026

• linalg.solve_triangular raises "Invalid transpose type" when A has conj and neg bits and out has conj bit

  bot-triagedmodule: complexmodule: linear algebrarelease triagetriage review
  pytorch/pytorch#200565 · Filter by author kostasrigatos opened on Oct 11, 2026

• torch.baddbmm has inconsistent NaN propagation between CPU and CUDA when alpha=0

  bot-triagedmodule: edge casesmodule: linear algebratriaged
  pytorch/pytorch#200563 · Filter by author tangyuan499-commits opened on Oct 11, 2026
  5

• nn.ModuleList.insert with a negative or out-of-range index raises and leaves the list corrupted

  bot-triagedmodule: nntriaged
  pytorch/pytorch#200558 · Filter by author superdiaodiao opened on Oct 11, 2026
  2

• [dynamo] torch.*.synchronize(device=None) crashes under torch.compile: torch.device(None) TypeError

  bot-triagedmodule: dynamooncall: pt2triaged
  pytorch/pytorch#200554 · Filter by author jinyouzhi opened on Oct 11, 2026

• [Inductor] torch.mm and torch.bmm do not respect out_dtype when using native_matmul

  bot-triagedmodule: correctness (silent)module: inductoroncall: pt2release triagetriage review
  pytorch/pytorch#200547 · Filter by author nathan-az opened on Oct 11, 2026

• torch.compile fails during fake-tensor propagation on deprecated uint8 mask indexing that eager still accepts

  bot-triagedmodule: pt2-dispatcheroncall: pt2triaged
  pytorch/pytorch#200546 · Filter by author tinywisdom opened on Oct 11, 2026
  1

• [MPS] bfloat16 lerp fails with ambiguous Metal abs call

  bot-triagedmodule: inductormodule: mpsoncall: pt2release triagetriaged
  pytorch/pytorch#200545 · Filter by author art-test-stack opened on Oct 11, 2026
  2

  

• float8_e8m0fnu: float32 -> E8M0 cast rounds every exact tie up, never to the even code

  bot-triagedmodule: correctness (silent)module: floatx (formerly float8)release triagetriage review
  pytorch/pytorch#200544 · Filter by author sw23 opened on Oct 10, 2026

• torch.compile: complex sgn decomposition overflows on tiny inputs which causes NaN gradients in abs backward

  bot-triagedmodule: decompositionsoncall: pt2triage review
  pytorch/pytorch#200543 · Filter by author Shakil-Muhit opened on Oct 10, 2026

• torch.triu/torch.tril on CPU write out of bounds when out= has non-contiguous batch dims

  bot-triagedmodule: correctness (silent)module: cpumodule: crashrelease triagetriage review
  pytorch/pytorch#200542 · Filter by author ethanclawsie opened on Oct 10, 2026

• [cuDNN][Blackwell] F.conv1d produces even/odd batch-position differences for identical FP32 inputs

  bot-triagedmodule: convolutionmodule: correctness (silent)module: cudnntriage review
  pytorch/pytorch#200541 · Filter by author huster-wgm opened on Oct 10, 2026

• [inductor] M=1 mm decomposition writes a (1, K, N) fp32 product to memory when the fused producers cross realize_opcount_threshold (3-6x slower GEMV)

  bot-triagedhas workaroundmodule: inductormodule: performanceoncall: pt2release triagetriaged
  pytorch/pytorch#200539 · Filter by author LoveManna opened on Oct 10, 2026

• torch.autocast(dtype=torch.bfloat16) runs cuDNN LSTM/GRU in float16

  bot-triagedmodule: amp (automated mixed precision)module: correctness (silent)module: cudnnmodule: rnnmodule: rocmtriage review
  pytorch/pytorch#200537 · Filter by author KeithG33 opened on Oct 10, 2026

• Numerical accuracy docs: bf16 matmul on Intel AMX CPUs treats subnormal inputs as zero above 16x16x16, regardless of set_flush_denormal

  bot-triagedmodule: bfloat16module: cpumodule: docsmodule: edge casesmodule: mkldnntriaged
  pytorch/pytorch#200531 · Filter by author tech4biz-yasha opened on Oct 10, 2026

• Official torch 2.14.1+cpu wheel GET returns HTTP 403 from a GitHub-hosted Ubuntu runner

  bot-triagedhigh prioritymodule: binariesmodule: third_partytriage review
  pytorch/pytorch#200527 · Filter by author VibhavTiwari opened on Oct 10, 2026
  2

• [dynamo] torch.full with a fill tensor that requires grad compiles (and backprops into it) where eager raises TypeError

  bot-triagedmodule: correctness (silent)module: dynamooncall: pt2triage review
  pytorch/pytorch#200524 · Filter by author Roy-Kid opened on Oct 10, 2026

• _record_memory_history stalls the allocating thread for 100–300ms at every 2^k entries (unbounded RingBuffer realloc under the default max_entries)

  bot-triagedenhancementmodule: cudamodule: CUDACachingAllocatormodule: performancetriaged
  pytorch/pytorch#200514 · Filter by author HOOLoLo opened on Oct 10, 2026
  1

• CI: torchtitan_features_integration fails at collection with "pyarrow requires NumPy 2.0 or newer, found 1.26.4"

  bot-triagedmodule: citriaged
  pytorch/pytorch#200511 · Filter by author kingdomhit opened on Oct 10, 2026

• [CUDA] torch.linalg.svd / svdvals / pinv: batched performance cliff at n = 32 -> 33 (default gesvdj loops per matrix), up to 130x slower than driver="gesvda"

  bot-triagedenhancementmodule: cudamodule: linear algebramodule: performancetriaged
  pytorch/pytorch#200486 · Filter by author shakilahmedemon opened on Oct 10, 2026
  1

• [CUDA] torch.mode is ~80-100x slower when the reduced dimension exceeds 2048 (per-slice host loop fallback)

  bot-triagedenhancementmodule: cudamodule: performancemodule: sorting and selectiontriaged
  pytorch/pytorch#200485 · Filter by author shakilahmedemon opened on Oct 10, 2026

• Torch MaxPooling1D rejects int8/int16/int64 accepted by TensorFlow

  bot-triagedmodule: poolingtriaged
  pytorch/pytorch#200482 · Filter by author L4AIsS opened on Oct 10, 2026
  1

• [xpu] torch.compile(mode="reduce-overhead") silently no-ops on XPU

  bot-triagedenhancementmodule: cuda graphsmodule: inductormodule: performancemodule: xpuoncall: pt2triaged
  pytorch/pytorch#200478 · Filter by author etaf opened on Oct 10, 2026
  

• PyTorch/Helion project creating enormous amount of B200 requests that are affecting PyTorch/PyTorch

  bot-triagedmodule: cimodule: infratriage review
  pytorch/pytorch#200457 · Filter by author nWEIdia opened on Oct 9, 2026
  5

• [functorch] vmap has no batching rule and no fallback for the _foreach_* ops

  bot-triagedfeaturemodule: functorchmodule: mtamodule: vmaptriaged
  pytorch/pytorch#200456 · Filter by author peterdsharpe opened on Oct 9, 2026