pytorch/pytorch

[RFC] Proposed Changes to Feature Tracking & Classification for PyTorch Releases starting Release 2.8
#152134 · atalman opened on Apr 24, 2025
8
[v.2.8.0] Release Tracker
#156745 · atalman opened on Jun 24, 2025
68

It seems that <code>torch.Tensor.addmv</code> and <code>torch.Tensor.addr</code> will check some inputs' dtype if and only if in <code>backward()</code>

actionable

complex_autograd

module: autograd

module: complex

module: linear algebra

module: type promotion

triaged

#95434

· ILoveSE opened

on Feb 24, 2023

jacfwd and jacrev are fundamentally broken for complex inputs

#94397

· lezcano opened

on Feb 8, 2023

[complex] Jacobian of a non-holomorphic complex valued function

complex_autograd

triaged

#92811

· kshitij12345 opened

on Jan 23, 2023

<code>cumprod, prod</code> will backward fail if <code>dtype</code> argument is different than the dtype of input tensor

#76786

· upupyy opened

on May 4, 2022

<code>addr, baddmm, dist, l1_loss</code> will backward fail when input tensors have different dtypes

#76785

· upupyy opened

on May 4, 2022

Discussion of TorchQuantum and QuantumNAS

complex_autograd

module: complex

triaged

#71446

· Hanrui-Wang opened

on Jan 18, 2022

Casting real parameter to complex during forward produces warning on backward

#68331

· lucgiffon opened

on Nov 15, 2021

Sign in slogdet is set to requires_grad = False even when using complex numbers.

complex_autograd

module: complex

module: linear algebra

triaged

#63138

· koji-inui opened

on Aug 12, 2021

Some ops throw runtime errors when called with complex tensors that require_grad but work when requires_grad=False

#60159

· mruberry opened

on Jun 17, 2021

complex128 autograd failures on PPC

#57617

· Flamefire opened

on May 5, 2021

Missing complex autograd support for some operators

complex_autograd

module: complex

triaged

#57358

· walterddr opened

on Apr 30, 2021

RuntimeError: polar does not support automatic differentiation for outputs with complex dtype.

complex_autograd

triaged

#54340

· YangangCao opened

on Mar 19, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

It seems that <code>torch.Tensor.addmv</code> and <code>torch.Tensor.addr</code> will check some inputs' dtype if and only if in <code>backward()</code>

jacfwd and jacrev are fundamentally broken for complex inputs

[complex] Jacobian of a non-holomorphic complex valued function

<code>cumprod, prod</code> will backward fail if <code>dtype</code> argument is different than the dtype of input tensor

<code>addr, baddmm, dist, l1_loss</code> will backward fail when input tensors have different dtypes

Discussion of TorchQuantum and QuantumNAS

Casting real parameter to complex during forward produces warning on backward

Sign in slogdet is set to requires_grad = False even when using complex numbers.

Some ops throw runtime errors when called with complex tensors that require_grad but work when requires_grad=False

complex128 autograd failures on PPC

Missing complex autograd support for some operators

RuntimeError: polar does not support automatic differentiation for outputs with complex dtype.

Issues

Search results