Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Add rollout information #514

Merged
merged 2 commits into from
Nov 11, 2021
Merged

Add rollout information #514

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dee7971
add rollout information
Nov 10, 2021
d2e1c53
Apply suggestions from code review
mzgubic Nov 11, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 22 additions & 0 deletions docs/src/index.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,28 @@ They should also have a test-only dependency on [ChainRulesTestUtils.jl](https:/

Note that the packages with rules do not have to depend on AD systems, and neither do the AD systems have to depend on individual packages.

## ChainRules roll-out status

Numerous [packages](https://juliahub.com/ui/Packages/ChainRulesCore/G6ax7/?page=2) depend on ChainRulesCore to define rules for their functions.

6 AD engines currently use ChainRules to get access to rules:

[Zygote.jl](https://github.com/FluxML/Zygote.jl) is a reverse-mode AD that supports using `rrule`s, calling back into AD, and opting out of rules.
However, its own [ZygoteRules.jl](https://github.com/FluxML/ZygoteRules.jl/) primitives (`@adjoint`s) take precedence before `rrule`s when both are defined -- even if the `@adjoint` is less specific than the `rrule`.
Internally it uses its own set of tangent types, e.g. `nothing` instead of `NoTangent`/`ZeroTangent`.
It also `unthunk`s every tangent.

[Diffractor.jl](https://github.com/JuliaDiff/Diffractor.jl) is a forward- and reverse-mode AD that fully supports ChainRules, including calling back into AD, opting out of rules, and uses tangent types internally.

[Yota](https://github.com/dfdx/Yota.jl) is a reverse-mode AD that fully supports ChainRules, including calling back into AD, opting out of rules, and uses tangent types internally.

[ReverseDiff](https://github.com/JuliaDiff/ReverseDiff.jl) is a reverse-mode AD that supports using `rrule`s, but not calling back into AD and opting out of rules.

[Nabla.jl](https://github.com/invenia/Nabla.jl) is a reverse-mode AD that supports using `rrule`s, but not opting out of rules, nor calling back into AD.

[ReversePropagation.jl](https://github.com/dpsanders/ReversePropagation.jl) is a reverse-mode AD that supports using `rrule`s for scalar functions, but not calling back into AD and opting out of rules.


## Key functionality

Consider a relationship $y = f(x)$, where $f$ is some function.
Expand Down