Reactant prototype

[gdalle]

First trial for #265

@wsmoses I run into a weird world age error in the test, any clue?

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 0% with 34 lines in your changes missing coverage. Please review.

Project coverage is 5.06%. Comparing base (c47e26c) to head (a5260d8).

Files with missing lines	Patch %	Lines
.../ext/DifferentiationInterfaceReactantExt/onearg.jl	0.00%	31 Missing :warning:
...ReactantExt/DifferentiationInterfaceReactantExt.jl	0.00%	3 Missing :warning:

:exclamation: There is a different number of reports uploaded between BASE (c47e26c) and HEAD (a5260d8). Click for more details.

HEAD has 98 uploads less than BASE

Flag	BASE (c47e26c)	HEAD (a5260d8)
DIT	20	1
DI	80	1

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #325       +/-   ##
==========================================
- Coverage   98.57%   5.06%   -93.51%     
==========================================
  Files         107      93       -14     
  Lines        4620    4465      -155     
==========================================
- Hits         4554     226     -4328     
- Misses         66    4239     +4173     

Flag	Coverage Δ
DI	7.40% <0.00%> (-91.27%)	:arrow_down:
DIT	0.13% <ø> (-98.22%)	:arrow_down:

Flags with carried forward coverage won't be shown. Click here to find out more.

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[wsmoses]

Hm that seems to imply that you're not using Reactant.compile

[gdalle]

You can look at the code, but I really think I am

https://github.com/gdalle/DifferentiationInterface.jl/blob/ed381bcdb5c4eb2a83239a9cbaf3044a1accdf56/DifferentiationInterface/ext/DifferentiationInterfaceReactantExt/DifferentiationInterfaceReactantExt.jl#L8

https://github.com/gdalle/DifferentiationInterface.jl/blob/ed381bcdb5c4eb2a83239a9cbaf3044a1accdf56/DifferentiationInterface/ext/DifferentiationInterfaceReactantExt/onearg.jl#L1-L17

[wsmoses]

I get you, but also the error log clearly indicates it wasn't. There should be some sort of casette compile in the logs if so.

  Closest candidates are:
    (::Reactant.var"#109#110")(::Any) (method too new to be called from this world context.)
     @ Reactant ~/.julia/packages/Reactant/DFtbF/src/Reactant.jl:850
  
  Stacktrace:
    [1] macro expansion
      @ ~/.julia/packages/Enzyme/G8o86/src/utils.jl:0 [inlined]
    [2] codegen_world_age(ft::Type{Reactant.var"#109#110"}, tt::Type{Tuple{Reactant.ConcreteRArray{Float64, (3,), 1}}})
      @ Enzyme ~/.julia/packages/Enzyme/G8o86/src/utils.jl:168
    [3] autodiff
      @ ~/.julia/packages/Enzyme/G8o86/src/Enzyme.jl:242 [inlined]
    [4] autodiff
      @ ~/.julia/packages/Enzyme/G8o86/src/Enzyme.jl:321 [inlined]
    [5] gradient(rm::EnzymeCore.ReverseMode{false, EnzymeCore.FFIABI, false}, f::Reactant.var"#109#110", x::Reactant.ConcreteRArray{Float64, (3,), 1})
      @ Enzyme ~/.julia/packages/Enzyme/G8o86/src/Enzyme.jl:1005
    [6] gradient(f::Function, backend::AutoEnzyme{Nothing}, x::Reactant.ConcreteRArray{Float64, (3,), 1}, ::DifferentiationInterface.NoGradientExtras)
      @ DifferentiationInterfaceEnzymeExt ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/ext/DifferentiationInterfaceEnzymeExt/reverse_onearg.jl:124
    [7] gradient(f::Function, rebackend::DifferentiationInterfaceReactantExt.ReactantBackend{AutoEnzyme{Nothing}}, x::Vector{Float64}, extras::DifferentiationInterfaceReactantExt.ReactantGradientExtras{Reactant.var"#109#110", DifferentiationInterface.NoGradientExtras})
      @ DifferentiationInterfaceReactantExt ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/ext/DifferentiationInterfaceReactantExt/onearg.jl:16
    [8] gradient(f::typeof(sum), backend::DifferentiationInterfaceReactantExt.ReactantBackend{AutoEnzyme{Nothing}}, x::Vector{Float64})
      @ DifferentiationInterface ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/src/first_order/gradient.jl:74
    [9] macro expansion
      @ /opt/hostedtoolcache/julia/1.10.4/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:669 [inlined]
   [10] top-level scope
      @ ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/test/Double/Enzyme-Reactant/test.jl:508

[wsmoses]

Are you trying to autodiff a reactant compiled function by chance? Reactant needs to be on the outside of all the gradient calls/etc?

[gdalle]

Are you trying to autodiff a reactant compiled function by chance?

Yes, that's what the code above shows. I'm doing

DI.gradient(f_compiled, backend, x_reactant)

Reactant needs to be on the outside of all the gradient calls/etc?

I assume the question mark at the end is unintended? And I should be doing the following?

compile(x -> DI.gradient(f, backend, x_reactant))

[wsmoses]

yeah sorry ignore the question mark. Indeed reactant needs to compile the outermost function [e.g. compile the gradient call]

[gdalle]

Still getting a world age error, even though now I'm compiling the gradient closure

[wsmoses]

@gdalle hm the log still thinks that you're trying to pass a compiled fn into autodiff rather than the other way round:

  MethodError: no method matching (::Reactant.var"#109#110")(::Reactant.ConcreteRArray{Float64, (3,), 1})
  The applicable method may be too new: running in world age 31490, while current world is 31491.
  
  Closest candidates are:
    (::Reactant.var"#109#110")(::Any) (method too new to be called from this world context.)
     @ Reactant ~/.julia/packages/Reactant/DFtbF/src/Reactant.jl:850
  
  Stacktrace:
   [1] gradient(f::Function, rebackend::DifferentiationInterfaceReactantExt.ReactantBackend{AutoEnzyme{Nothing}}, x::Vector{Float64}, extras::DifferentiationInterfaceReactantExt.ReactantGradientExtras{Reactant.var"#109#110"})
     @ DifferentiationInterfaceReactantExt ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/ext/DifferentiationInterfaceReactantExt/onearg.jl:16
   [2] gradient(f::typeof(sum), backend::DifferentiationInterfaceReactantExt.ReactantBackend{AutoEnzyme{Nothing}}, x::Vector{Float64})
     @ DifferentiationInterface ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/src/first_order/gradient.jl:74
   [3] macro expansion
     @ /opt/hostedtoolcache/julia/1.10.4/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:669 [inlined]
   [4] top-level scope
     @ ~/work/DifferentiationInterface.jl/DifferentiationInterface.jl/DifferentiationInterface/test/Double/Enzyme-Reactant/test.jl:508

[gdalle]

The entire code is here (gradient is automatically preceded by preparation) and I really don't see where I'm doing that:

https://github.com/gdalle/DifferentiationInterface.jl/blob/gd/reactant/DifferentiationInterface/ext/DifferentiationInterfaceReactantExt/onearg.jl

[wsmoses]

@gdalle okay I just released a reactant bump which fixes this

[gdalle]

New kind of error:

conversion to pointer not defined for Reactant.TracedRArray{Float64, (6,), 1}

[wsmoses]

huh weird, open an issue with an MWE?

[gdalle]

Tests are passing locally :partying_face: now onto implementing more operators