Issues

• torch.median unsupported op

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#6338 · Filter by author chrishjones20 opened on Nov 6, 2026
  

• SDPA routing: 0.2.0 is 7x slower at 51 tokens (fixed on main, unreleased), and SHLO, which 9d8cd69a now picks at 128-256 tokens, measured slower than Flash

  HighFilter by Priority: High
  google-pytorch/torch_tpu#6304 · Filter by author Erlemar opened on Nov 6, 2026
  1

  

• [Ops][var/std] var, std and var_mean divide by the wrong element count with bounded dynamic shapes

  stat:awaiting-assignee
  google-pytorch/torch_tpu#6229 · Filter by author apullin opened on Nov 5, 2026
  2

  

• [feature] Some cuda-stream like API to control D2H/H2D

  stat:awaiting-requesterstat:awaiting-reviewers
  google-pytorch/torch_tpu#6216 · Filter by author zou3519 opened on Nov 4, 2026
  7

  

• [Ops][scaled_dot_product_attention] Backward with dropout_p > 0 uses a different dropout mask than the forward at some sequence lengths

  stat:awaiting-assigneestat:awaiting-reviewers
  google-pytorch/torch_tpu#6150 · Filter by author apullin opened on Nov 2, 2026
  5

  

• torch.logaddexp (aten::logaddexp.out) is not implemented for TPU

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#6118 · Filter by author joinalahmed opened on Nov 2, 2026
  2

  

• Versioned extension loader hides the torch/glue version mismatch behind "undefined symbol"

  stat:awaiting-requester

  LowFilter by Priority: Low
  Feature

  google-pytorch/torch_tpu#6113 · Filter by author joinalahmed opened on Nov 1, 2026
  4

  

• Bounded dynamism: 11 stale "untriaged" deny-list entries pass; clamp.Tensor returns wrong values with a bounded dim

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#6112 · Filter by author joinalahmed opened on Nov 1, 2026
  2

  

• [Ops][bincount] Silent data corruption for negative inputs and missing precondition validations in torch.bincount

  stat:awaiting-assignee

  HighFilter by Priority: High
  google-pytorch/torch_tpu#6107 · Filter by author Kaap10 opened on Oct 31, 2026
  2

  

• F.max_unpool2d (aten::max_unpool2d) is not implemented for TPU; blocks Captum DeepLift on CNNs

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#6103 · Filter by author joinalahmed opened on Oct 31, 2026
  

• torch.linalg.cross / torch.cross (aten::linalg_cross.out) is not implemented for TPU

  stat:awaiting-assignee
  google-pytorch/torch_tpu#6070 · Filter by author petkokp opened on Oct 30, 2026
  

• group_norm: bf16 input with f32 weight/bias returns f32 on TPU (CPU and the meta function return bf16)

  stat:awaiting-requester

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#6015 · Filter by author intexcor opened on Oct 29, 2026
  1

  

• [Ops][addmv] Missing shape and dimension validation for self in torch.addmv

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#6010 · Filter by author Kaap10 opened on Oct 29, 2026
  1

  

• F.cross_entropy over logits[:, :-1, :].flatten(0, 1) compiles for minutes when S-1 is not a multiple of 8

  stat:awaiting-assignee
  google-pytorch/torch_tpu#5829 · Filter by author ethannnnnn opened on Oct 24, 2026
  1

  

• Support aten::_int_mm (INT8 × INT8 → INT32) on TPU

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5826 · Filter by author chenmiaoming opened on Oct 24, 2026
  3

  

• torch.gcd (aten::gcd.out) is not supported

  stat:awaiting-requester

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5725 · Filter by author capemox opened on Oct 22, 2026
  1

  

• Numerical issue with gelu lowering at fp32

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5700 · Filter by author hundt98847 opened on Oct 21, 2026
  1

  

• Feature request: aten::mish (and mish_backward) not implemented for TPU

  stat:awaiting-assignee

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5679 · Filter by author intexcor opened on Oct 21, 2026
  1

  

• layer_norm under torch.autocast("tpu", bfloat16) fails to lower: stablehlo.broadcast_in_dim requires compatible element types

  stat:awaiting-requester

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5678 · Filter by author intexcor opened on Oct 21, 2026
  

• view_decomposition: NotImplementedError for input views on Dynamo dynamic recompile (independent storage-offset symbol)

  stat:awaiting-requester

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5674 · Filter by author gpu004 opened on Oct 21, 2026
  3

  

• BatchNorm never updates running_mean/running_var in training mode, in eager and with torch.compile

  stat:awaiting-assignee

  HighFilter by Priority: High
  google-pytorch/torch_tpu#5667 · Filter by author Kasties opened on Oct 20, 2026
  1

  

• In eager mode, an op on x[i] compiles one program per index i, and in an accumulating loop each one takes ~0.3 s

  stat:awaiting-requester

  MediumFilter by Priority: Medium
  google-pytorch/torch_tpu#5086 · Filter by author tengomucho opened on Oct 9, 2026
  3

  

• nonzero takes ~1.2 s to compile for every new input length, and again for every new number of nonzeros

  stat:awaiting-requester
  google-pytorch/torch_tpu#5085 · Filter by author tengomucho opened on Oct 9, 2026
  3

  

• torch.compile(backend="tpu") recompiles when an input is a view with a different storage offset

  stat:awaiting-requester
  google-pytorch/torch_tpu#5000 · Filter by author JingyaHuang opened on Oct 8, 2026
  2

  

• index_add_ (scatter-add) returns wrong rows for some shapes

  stat:awaiting-requester
  google-pytorch/torch_tpu#4784 · Filter by author tengomucho opened on Oct 5, 2026
  4