Issues

• [quantized] Empty BatchNorm inputs ignore requested output quantization parameters

  bot-triagedoncall: quantization
  pytorch/pytorch#204796 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• Add torch.backends.cpu.matmul.allow_fp16_accumulation for fp16 linear/matmul on x86 CPUs

  bot-triagedenhancementmodule: cpumodule: halfmodule: linear algebramodule: mkldnnmodule: performancetriaged
  pytorch/pytorch#204794 · Filter by author mriganks-amd opened on Nov 13, 2026

• [quantized][oneDNN] linear_prepack produces incorrect results with non-contiguous bias

  bot-triagedoncall: quantization
  pytorch/pytorch#204793 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [torch.compile][Inductor] Forced cooperative reduction of a [15, 1048575] tensor over dim=0 fails CUDA kernel launch

  bot-triagedmodule: cudamodule: inductormodule: reductionsoncall: pt2triaged
  pytorch/pytorch#204790 · Filter by author pinpinpoo opened on Nov 13, 2026

• [CUDA] softmax/log_softmax along a non-last dim is up to ~180x slower than along the last dim

  bot-triagedenhancementmodule: cudamodule: nnmodule: performancetriaged
  pytorch/pytorch#204788 · Filter by author rynwang opened on Nov 13, 2026

• torch._numpy.pad silently crops input for negative pad_width

  bot-triagedmodule: error checkingmodule: numpymodule: paddingtriage review
  pytorch/pytorch#204781 · Filter by author Rajatendu1 opened on Nov 13, 2026

• [torch.compile][Inductor] Int8 weight-only linear returns incorrect values for a non-contiguous BF16 input

  bot-triagedmodule: correctness (silent)module: inductoroncall: pt2triage review
  pytorch/pytorch#204778 · Filter by author pinpinpoo opened on Nov 13, 2026
  3

• Quantized conv2d_add allows out-of-bounds writes with an undersized accumulator

  bot-triagedoncall: quantization
  pytorch/pytorch#204777 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [Bug][nn] BatchNorm counts empty batches in cumulative running statistics

  bot-triagedmodule: correctness (silent)module: empty tensormodule: norms and normalizationrelease triagetriage review
  pytorch/pytorch#204775 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [Bug][nn] MultiheadAttention fastpath skips the attn_mask check for is_causal=True

  bot-triagedmodule: correctness (silent)module: error checkingmodule: nnrelease triagetriage review
  pytorch/pytorch#204772 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• Tensor.to(1.5) silently casts to float64 while to(dtype=1.5) and every other dtype entry point raise TypeError

  bot-triagedmodule: error checkingmodule: pybindmodule: windowstriage review
  pytorch/pytorch#204771 · Filter by author cuiliaomei-beep opened on Nov 13, 2026

• BCEWithLogits accepts a negative pos_weight without validation, producing a negative loss

  bot-triagedmodule: error checkingmodule: lossmodule: windowstriage review
  pytorch/pytorch#204770 · Filter by author cuiliaomei-beep opened on Nov 13, 2026

• normalize, layer_norm, group_norm and instance_norm accept a non-positive eps without validation, silently producing NaN or zero on zero-variance input

  bot-triagedmodule: error checkingmodule: norms and normalizationmodule: windowstriage review
  pytorch/pytorch#204769 · Filter by author cuiliaomei-beep opened on Nov 13, 2026

• F.hardshrink accepts a negative lambd and returns the input unchanged, with no documented restriction (unlike F.softshrink)

  bot-triagedenhancementmodule: docsmodule: nnmodule: windowstriaged
  pytorch/pytorch#204768 · Filter by author cuiliaomei-beep opened on Nov 13, 2026

• [Bug][quantization] prepare_fx reuses input observers with different qconfig parameters

  bot-triagedoncall: quantization
  pytorch/pytorch#204766 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [Bug][sparse] torch.sparse.sum with an empty dim list skips reduction

  bot-triagedmodule: correctness (silent)module: reductionsmodule: sparsetriage review
  pytorch/pytorch#204757 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [Bug][export] strict=True drops registered buffer reassignments and returns stale state

  bot-triagedoncall: exportoncall: pt2
  pytorch/pytorch#204755 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [ROCm] Memory-efficient attention forward launches three zero-fill kernels per call with the CK backend

  bot-triagedenhancementmodule: performancemodule: rocmmodule: sdpatriaged
  pytorch/pytorch#204753 · Filter by author mgehre-amd opened on Nov 13, 2026

• DISABLED test_fake_export_cuda_conj_cpu_float32 (__main__.TestExportOpInfoCPU)

  module: flaky-testsoncall: exportoncall: pt2skipped
  pytorch/pytorch#204745 · Filter by author pytorch-bot opened on Nov 13, 2026
  2

• [Bug][export] Dynamic shape guard is dropped for size-dependent branches, causing incorrect outputs

  bot-triagedoncall: exportoncall: pt2
  pytorch/pytorch#204744 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [RFC] A CPU backend for torch._grouped_mm on oneDNN's grouped matmul

  bot-triagedenhancementmodule: cpumodule: linear algebramodule: mkldnnmodule: performancetriaged
  pytorch/pytorch#204742 · Filter by author mriganks-amd opened on Nov 13, 2026

• [torch.compile][Inductor] Forced cooperative reduction with a dynamic reduction dimension fails Triton constexpr compilation

  bot-triagedmodule: dynamic shapesmodule: inductormodule: reductionsoncall: pt2triaged
  pytorch/pytorch#204729 · Filter by author pinpinpoo opened on Nov 13, 2026
  1

• [Bug][FX] symbolic_trace mutates module buffers during tracing and silently drops in-place updates

  bot-triagedhas workaroundmodule: correctness (silent)module: fxtriage review
  pytorch/pytorch#204728 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• [Bug][quantization] oneDNN quantized Linear crashes when in_features is zero

  bot-triagedoncall: quantization
  pytorch/pytorch#204725 · Filter by author tangyuan499-commits opened on Nov 13, 2026

• AOTI stream affinity capability is hard-coded to USE_CUDA/USE_XPU, while the generated C ABI exports it unconditionally

  bot-triagedhas workaroundmodule: aotinductormodule: PrivateUse1oncall: exportoncall: pt2triaged
  pytorch/pytorch#204706 · Filter by author youlituao opened on Nov 13, 2026