Precompile fails with LoadError: Non-differentiable function mul_int

[1advaita1]

The following result is obtained by using Zygote and subsequently Zygote.refresh():

julia> using Zygote
ERROR: InitError: LoadError: Non-differentiable function mul_int
Stacktrace:
 [1] error(::String) at .\error.jl:33
 [2] (::typeof(∂(λ)))(::Int64) at .\none:0
 [3] * at .\int.jl:54 [inlined]
 [4] (::typeof(∂(*)))(::Int64) at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\compiler\interface2.jl:0
 [5] pow at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\compiler\interface.jl:135 [inlined]
 [6] (::typeof(∂(Zygote.pow)))(::Int64) at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\compiler\interface2.jl:0
 [7] (::getfield(Zygote, Symbol("##32#33")){typeof(∂(Zygote.pow))})(::Int64) at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\compiler\interface.jl:38
 [8] gradient(::Function, ::Int64, ::Vararg{Int64,N} where N) at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\compiler\interface.jl:47
 [9] top-level scope at none:0
 [10] include at .\boot.jl:317 [inlined]
 [11] include_relative(::Module, ::String) at .\loading.jl:1044
 [12] include at .\sysimg.jl:29 [inlined]
 [13] include at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\Zygote.jl:1 [inlined]
 [14] precompile at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\Zygote.jl:52 [inlined]
 [15] (::getfield(Zygote, Symbol("##1386#1387")))() at C:\Users\advaita\.julia\packages\Requires\9Jse8\src\init.jl:11
 [16] __init__() at C:\Users\advaita\.julia\packages\Requires\9Jse8\src\init.jl:18
 [17] _include_from_serialized(::String, ::Array{Any,1}) at .\loading.jl:633
 [18] _require_search_from_serialized(::Base.PkgId, ::String) at .\loading.jl:713
 [19] _require(::Base.PkgId) at .\loading.jl:937
 [20] require(::Base.PkgId) at .\loading.jl:858
 [21] require(::Module, ::Symbol) at .\loading.jl:853
in expression starting at C:\Users\advaita\.julia\packages\Zygote\HErRu\src\precompile.jl:10
during initialization of module Zygote

julia> using Zygote

julia>

julia> Zygote.refresh()
ERROR: LoadError: Non-differentiable function mul_int

The first time I ran it, it showed:

WARNING: could not import ZygoteRules._pullback into Zygote
WARNING: could not import ZygoteRules.pullback into Zygote
ERROR: LoadError: InitError: LoadError: Non-differentiable function mul_int

Status in Julia v1.0.4

(v1.0) pkg> st
    Status `C:\Users\advaita\.julia\environments\v1.0\Project.toml`
  [7869d1d1] IRTools v0.2.3 #master (https://github.com/MikeInnes/IRTools.jl.git)
  [e88e6eb3] Zygote v0.3.4 #master (https://github.com/FluxML/Zygote.jl.git)

The same issue exists in Julia v1.2.0

[DoktorMike]

I get the same thing using Zygote#master in julia 1.3 RC2

[simonmandlik]

Same here, in julia 1.1

[janEbert]

Fix by DoktorMike: ]add ZygoteRules#master

[zenna]

Thanks for the fix. Currently a fresh install of both Zygote and Flux are both broken on a fresh install of Julia. Maybe some tighter compat bounds would help?

[mileslucas]

I am still unable to precompile even using ZygoteRules#master. Julia 1.2, Mac OS X Catalina, fresh Julia project.

[zenna]

Are you on Zygote#master too?

[mileslucas]

@zenna It works in a completely clean install

(test) pkg> add Zygote#master ZygoteRules#master

So I suppose that fix works. Unfortunately I don't want to use base Zygote, I want to use Turing, which encounters a variety of messy semvar restrictions. As of right now, I cannot get any version of Turing working, but that is beyond the scope of this issue.

[zenna]

Yeah. I have the same issue with Flux

[racinmat]

I just upgraded ZygoteRules to v0.2.0, and the issue happened. With ZygoteRules v0.1.0 it works fine though.

[robertfeldt]

Same problems here: Installing ZygoteRules#master is not enough and I cannot go with Zygote#master since there are dependency conflicts with SpecialFunctions:

(v1.2) pkg> add Zygote#master
   Cloning git-repo `https://github.com/FluxML/Zygote.jl.git`
  Updating git-repo `https://github.com/FluxML/Zygote.jl.git`
 Resolving package versions...
ERROR: Unsatisfiable requirements detected for package SpecialFunctions [276daf66]:
 SpecialFunctions [276daf66] log:
 ├─possible versions are: [0.7.0-0.7.2, 0.8.0] or uninstalled
 ├─restricted to versions 0.8.0-0.8 by Zygote [e88e6eb3], leaving only versions 0.8.0
 │ └─Zygote [e88e6eb3] log:
 │   ├─possible versions are: 0.4.0 or uninstalled
 │   └─Zygote [e88e6eb3] is fixed to version 0.4.0
 └─restricted to versions 0.7.2 by an explicit requirement — no versions left

(v1.2) pkg> add Zygote
 Resolving package versions...
  Updating `~/.julia/environments/v1.2/Project.toml`
  [e88e6eb3] + Zygote v0.3.4
  Updating `~/.julia/environments/v1.2/Manifest.toml`
 [no changes]

julia> using ZygoteRules

julia> using Zygote
[ Info: Precompiling Zygote [e88e6eb3-aa80-5325-afca-941959d7151f]
WARNING: could not import ZygoteRules._forward into Zygote
ERROR: LoadError: LoadError: Non-differentiable function mul_int
Stacktrace:
 [1] error(::String) at ./error.jl:33

[Red-Portal]

Find a way around. Just downgrade ZygoteRules.
At least for me, going back didn't break anything.
Maybe this will introduce an issue if you're actually using Zygote for something.

add ZygoteRules@0.1.0

I wonder if this kind of critical dependency issue will occur again in the future.
This is really critical since you simply neither roll back or go bleeding edge.

[MikeInnes]

There's a PR open to fix this compat bound: JuliaRegistries/General#5065

It's less likely to happen in future because we have upper bounds on compatibility now.