World Age Issues with TensorCast calls from pyJulia

[andrewkhardy]

I am getting a world age bug when running TensorCast in JuliaPy, but not when using it normally.

If I run an example function such as

function testfunc()
    V = [10,20,30]
    M = collect(reshape(1:12, 3,4))
    @cast A[j,i] := M[i,j] + 10 * V[i]
    return A
end

it works when I call it in Julia via

A = testfunc()

However, when I run the Python script

from julia.api import Julia
jl = Julia(compiled_modules=False)
from julia import Main
Main.include("testfunction.jl")
from julia.Main import testfunc

A = testfunc()

I get the error

<PyCall.jlwrap (in a Julia function called from Python)
JULIA: MethodError: no method matching _trex(::Symbol, ::Type{Matrix{Int64}}, ::Tuple{Int64, Int64})
The applicable method may be too new: running in world age 32486, while current world is 38140.
Closest candidates are:
  _trex(::Any, ::Any, ::Any) at /home/andrewhardy/.julia/packages/TransmuteDims/mGkQB/src/TransmuteDims.jl:372 (method too new to be called from this world context.)

which only occurs for the @cast call. What can I do?

[mcabbott]

I have never used pyJulia, but did briefly manage to get this error directly from the repl:

julia> TensorCast.transmute(rand(2,2), Val((2, 1)))
ERROR: MethodError: no method matching _trex(::Symbol, ::Type{Matrix{Float64}}, ::Tuple{Int64, Int64})
The applicable method may be too new: running in world age 45668, while current world is 47824.

What TensorCast.jl is doing here is very simple, the macro expands to this:

julia> @pretty @cast A[j,i] := M[i,j] + 10 * V[i]
begin
    @boundscheck ndims(M) == 2 || throw(ArgumentError("expected a 2-tensor M[i, j]"))
    @boundscheck axes(M, 1) == axes(V, 1) || throw(DimensionMismatch("range of index i must agree"))
    @boundscheck V isa Tuple || (ndims(V) == 1 || throw(ArgumentError("expected a vector or tuple V[i]")))
    local guineafowl = transmute(M, Val((2, 1)))
    local giraffe = transmute(V, Val((nothing, 1)))
    A = @__dot__(guineafowl + 10giraffe)
end

But transmute is a generalised version of permutedims, which has some @generated methods, and these are (I presume) what's causing the world age problem. But why, I'm not sure. I think such things can happen if code is updated after loading, or something... am not so sure how to track down what's gone wrong, nor to fix it.

[andrewkhardy]

For what it is worth @reduce does not throw these errors. Does that have any generated methods?

[mcabbott]

Depends on the expression, for both. Any transmute(M, ::Val) call is generated, I think, and @pretty will show you what the macro calls.