Infiltrating package functions.

[fpartl]

Hello,

I am trying to debug one of the function from my VAEs package. I have Infiltrator.jl and Revise.jl added to my main environment @v1.7. For debugging purposes I am running this code in VAEs environment (default active environment in VS Code). What does the message "Cannot infiltrate while the REPL is idle. This typically..." means?

(VAEs) pkg> st
     Project VAEs v0.1.0
      Status `~/.julia/dev/VAEs/Project.toml`
  [052768ef] CUDA v3.12.0
  [587475ba] Flux v0.13.4

julia> using Revise

julia> using Infiltrator

julia> using VAEs
[ Info: Precompiling VAEs [986d28bc-f63f-40d7-9091-9de846b66fef]
┌ Warning: Package VAEs does not have Infiltrator in its dependencies:
│ - If you have VAEs checked out for development and have
│   added Infiltrator as a dependency but haven't updated your primary
│   environment's manifest file, try `Pkg.resolve()`.
│ - Otherwise you may need to report an issue with VAEs
└ Loading Infiltrator into VAEs from project dependency, future warnings for VAEs are suppressed.

julia> using Flux

julia> xsize, ysize, seqsize, batch = 10, 20, 3, 6
(10, 20, 3, 6)

julia> bilstm = BiRecur(Flux.LSTM, xsize, ysize)
WARNING: both Losses and NNlib export "ctc_loss"; uses of it in module Flux must be qualified
BiRecur(Flux.Recur{Flux.LSTMCell{Matrix{Float32}, Vector{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Tuple{Matrix{Float32}, Matrix{Float32}}}, 10, 20)

julia> x₁ = rand(Float32, xsize, seqsize, batch); size(x₁)
(10, 3, 6)

julia> ŷ₁ = bilstm(x₁);
Running Union version of BiRecur(Array{Float32, 3}) with size(x) = (10, 3, 6)
Running Vector version of BiRecur(Vector{SubArray{Float32, 2, Array{Float32, 3}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}}) with size(x[begin]) = (10, 6)
Cannot infiltrate while the REPL is idle. This typically happens when an async task is running in the background.
Disabling infiltration point at (::BiRecur{Flux.Recur{Flux.LSTMCell{Matrix{Float32}, Vector{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Tuple{Matrix{Float32}, Matrix{Float32}}}})(x::Vector{SubArray{Float32, 2, Array{Float32, 3}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}}) at BiRecur.jl:56.

Here is the status of my @v1.7 environment.

(VAEs) pkg> activate
  Activating project at `~/.julia/environments/v1.7`

(@v1.7) pkg> st
      Status `~/.julia/environments/v1.7/Project.toml`
  [6e4b80f9] BenchmarkTools v1.3.1
  [31a5f54b] Debugger v0.7.6
  [5903a43b] Infiltrator v1.5.0
  [295af30f] Revise v3.3.4
  [986d28bc] VAEs v0.1.0 `~/.julia/dev/VAEs`

And here is the debug function.

using Infiltrator

...

function (b::BiRecur)(x::AbstractVector{T}) where {T <: AbstractVecOrMat{Float32}}
    println("Running Vector version of BiRecur($(typeof(x))) with size(x[begin]) = $(size(x[begin]))")

    fw = [b.fw(s) for s ∈ x]
    bw = Flux.flip(b.bw, x)
    result = cat.(fw, bw, dims=1)

    @infiltrate # Cannot infiltrate while the REPL is idle.

    return result
end

Basically I do not know the right workflow for package development (debugging). I can use debugger built in the VS Code but in my case I need the only sizes of my tensors and debugger gives me too much irrelevant and confusing informations.

[pfitzseb]

Impossible to say what's going on without a MWE. This does look like a false positive though.

Technically, I'm only checking whether Base.active_repl_backend.in_eval is true or not; you can circumvent that check by setting Infiltrator.CHECK_TASK[] = false and see if that works.

Aside: I'd recommend not loading Infiltrator in your package; instead load it at the REPL and use Main.@infiltrate in your package via Revise.

[fpartl]

It works with Infiltrator.CHECK_TASK[] = false. I'll try to create some MWE. ;-)

Aside: When I use just Main.@infiltrate without using Infiltrator my package can't be precompiled. With this solution it works but it is a little bit chatty for my taste. :-)

if isdefined(Main, :Infiltrator)
    Main.infiltrate(@__MODULE__, Base.@locals, @__FILE__, @__LINE__)
end

[pfitzseb]

Right, that's why I suggested only adding @infiltrate after the package is already loaded :)

[pfitzseb]

Re MWE: If you can share VAEs.jl then I can work with that :)

[fpartl]

VAEs.jl is total mess... rather try this. Run the ./test/mwe.jl file. ;-)

[pfitzseb]

julia> using Revise

julia> using Infiltrator

julia> using InfiltratorMWE

julia> using Flux

julia> # Infiltrator.CHECK_TASK[] = false

       xsize, ysize, batch = 10, 20, 6
(10, 20, 6)

julia> cell = CrazyRNNCell(Flux.LSTM, xsize, ysize)
WARNING: both Losses and NNlib export "ctc_loss"; uses of it in module Flux must be qualified
CrazyRNNCell(Flux.Recur{Flux.LSTMCell{Matrix{Float32}, Vector{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Tuple{Matrix{Float32}, Matrix{Float32}}}, 10, 40)

julia> x = rand(Float32, xsize, batch)
10×6 Matrix{Float32}:
 0.528503   0.00164235  0.0748651  0.0936595  0.482486  0.000504434
 0.339793   0.768868    0.658447   0.585099   0.136515  0.73854
 0.0926465  0.869227    0.870997   0.675454   0.147139  0.164868
 0.507805   0.950567    0.935592   0.675728   0.967071  0.576838
 0.47404    0.298015    0.784961   0.479873   0.443843  0.421045
 0.102862   0.196318    0.700858   0.908191   0.323899  0.230969
 0.167544   0.41923     0.261201   0.649286   0.79906   0.243012
 0.282973   0.282155    0.306425   0.797724   0.366297  0.112101
 0.512463   0.466117    0.0114826  0.289244   0.402586  0.0258904
 0.759036   0.627821    0.601763   0.883049   0.605143  0.870425

julia> y = cell(x); # And voilà!
Infiltrating (::CrazyRNNCell{Flux.Recur{Flux.LSTMCell{Matrix{Float32}, Vector{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Tuple{Matrix{Float32}, Matrix{Float32}}}})(x::Matrix{Float32})
  at /home/pfitzseb/.julia/dev/InfiltratorMWE/src/CrazyRNNCell.jl:31

infil> @trace
[1] (::CrazyRNNCell{Flux.Recur{Flux.LSTMCell{Matrix{Float32}, Vector{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Tuple{Matrix{Float32}, Matrix{Float32}}}})(x::Matrix{Float32})
    at /home/pfitzseb/.julia/dev/InfiltratorMWE/src/CrazyRNNCell.jl:31
[2] top-level scope

infil>

works fine for me. What Julia version/OS are you on?

[fpartl]

Oh... come on! :-D

julia> versioninfo()
Julia Version 1.7.3
Commit 742b9abb4d (2022-05-06 12:58 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)
Environment:
  JULIA_PKG_USE_CLI_GIT = true
  JULIA_NUM_THREADS = 8

[pfitzseb]

Do you have anything in your startup.jl (like e.g. OhMyREPL)?

[fpartl]

No, it is empty.

[pfitzseb]

Just to confirm. Did you use inline evaluation (Ctrl+Enter) for that or copy-paste it into the REPL?

[fpartl]

Inline, is it a problem?

Oh I see... copy-paste works.

[pfitzseb]

Yes, as per the readme:

Running code that ends up triggering the @infiltrate REPL mode via inline evaluation in VSCode or Juno can cause issues, so it's recommended to always use the REPL directly.

I'll add some more info to the error message :)

[fpartl]

My bad. Sorry and thank you! 👍

[pfitzseb]

1.6 will have an updated error message: