Issues · pytorch/pytorch · GitHub

8000 Issues · pytorch/pytorch · GitHub

Labels Milestones New issue

`torch.library.custom.register_autograd` incompatible with `jacrev`

module: autograd

module: functorch

#179652

· PTNobel opened

on Apr 8, 2026

Call Hierarchy Metadata

#179643

· andrewd-aws opened

on Apr 7, 2026

skip_nested_compile doesn't get picked up in callback during checkpoint

module: activation checkpointing

#179614

· tugsbayasgalan opened

on Apr 7, 2026

[MPS] avg_pool1d produces negative values from non-negative input when zeros follow large floats

module: correctness (silent)

module: edge cases

module: pooling

#179608

· gaborv opened

on Apr 7, 2026

Inductor codecache raises "ValueError: can't pickle cyclic objects" when using flattened meshes

module: inductor

module: regression

module: serialization

oncall: distributed

#179597

· carmocca opened

on Apr 7, 2026

`torch.compile` fails on `torch.istft` while eager succeeds with warning only (`length` handling mismatch in fake/meta execution)

#179593

· tinywisdom opened

on Apr 7, 2026

`torch.compile` produces different output for single-head attention with `matmul → div → softmax → dropout → matmul` pattern (SFDP pattern 11) compared to eager mode

#179578

· prorrice opened

on Apr 7, 2026

`torch.compile` produces significantly different output for model with multiple `amax`/`amin` reductions on same tensor compared to eager mode

module: correctness (silent)

module: inductor

module: reductions

#179577

· prorrice opened

on Apr 7, 2026

`torch.compile` raises error for model with multiple `torch.randint` calls with different shapes and ranges while eager mode succeeds

#179576

· prorrice opened

on Apr 7, 2026

`torch.compile` produces different output for Transformer using `torch.einsum` projections compared to eager mode

module: correctness (silent)

module: inductor

module: numerical-stability

#179575

· prorrice opened

on Apr 7, 2026

`torch.compile` produces numerically different results for pre-grad batch linear fusion pattern (multiple `F.linear` with same input) compared to eager mode

#179574

· prorrice opened

on Apr 7, 2026

`torch.compile` raises error for redundant consecutive view operations pattern (`view → compute → view_back`) while eager mode succeeds

#179573

· himi1008 opened

on Apr 7, 2026

0