GradType: a subset of dtype that is differentiable, containing all float and complex dtypes

[dvorst]

🚀 The feature, motivation and pitch

In the documentation, dtype is often used as a type hint where in actuality only differentiable dtypes are allowed (float and complex). Therefore the request to add GradType, a type that can have a gradient / is differentiable, which encompasses all different float and complex types. This allows type hints to more precisely indicate what is allowed.

I would like to implement this feature myself if accepted.

Alternatives

No response

Additional context

No response

cc @svekars @brycebortree @ezyang @albanD @gqchen @pearu @nikitaved @soulitzer @Varal7

[mikaylagawarecki]

Do you have specific examples of functions that are type-hinted with dtype that would benefit from this

[dvorst]

I suppose everything for which a gradient is calculated during backprop, such as torch.nn.Linear