fix: NCCL import of @checked + support for CUDA 13

[avik-pal]

The current release fails with

WARNING: could not import APIUtils.@checked into LibNCCL
ERROR: LoadError: UndefVarError: `@checked` not defined in `NCCL.LibNCCL`
Stacktrace:
 [1] top-level scope
   @ :0
 [2] include(mod::Module, _path::String)
   @ Base ./Base.jl:562
 [3] include(x::String)
   @ NCCL /mnt/.julia/packages/NCCL/CwvFb/src/NCCL.jl:1
 [4] top-level scope
   @ /mnt/.julia/packages/NCCL/CwvFb/src/NCCL.jl:5
 [5] include
   @ ./Base.jl:562 [inlined]
 [6] include_package_for_output(pkg::Base.PkgId, input::String, depot_path::Vector{String}, dl_load_path::Vector{String}, load_path::Vector{String}, concrete_deps::Vector{Pair{Base.PkgId, UInt128}}, source::Nothing)
   @ Base ./loading.jl:2881
 [7] top-level scope
   @ stdin:6
in expression starting at /mnt/.julia/packages/NCCL/CwvFb/src/libnccl.jl:56
in expression starting at /mnt/.julia/packages/NCCL/CwvFb/src/libnccl.jl:1
in expression starting at /mnt/.julia/packages/NCCL/CwvFb/src/NCCL.jl:1
in expression starting at stdin:6
  ✗ NCCL
  0 dependencies successfully precompiled in 4 seconds. 103 already precompiled.

ERROR: The following 1 direct dependency failed to precompile:

NCCL 

Failed to precompile NCCL [3fe64909-d7a1-4096-9b7d-7a0f12cf0f6b] to "/mnt/.julia/compiled/v1.11/NCCL/jl_t0Ow3O".
WARNING: could not import APIUtils.@checked into LibNCCL
ERROR: LoadError: UndefVarError: `@checked` not defined in `NCCL.LibNCCL`
Stacktrace:
 [1] top-level scope
   @ :0
 [2] include(mod::Module, _path::String)
   @ Base ./Base.jl:562
 [3] include(x::String)
   @ NCCL /mnt/.julia/packages/NCCL/CwvFb/src/NCCL.jl:1
 [4] top-level scope
   @ /mnt/.julia/packages/NCCL/CwvFb/src/NCCL.jl:5
 [5] include
   @ ./Base.jl:562 [inlined]
 [6] include_package_for_output(pkg::Base.PkgId, input::String, depot_path::Vector{String}, dl_load_path::Vector{String}, load_path::Vector{String}, concrete_deps::Vector{Pair{Base.PkgId, UInt128}}, source::Nothing)
   @ Base ./loading.jl:2881
 [7] top-level scope
   @ stdin:6
in expression starting at /mnt/.julia/packages/NCCL/CwvFb/src/libnccl.jl:56
in expression starting at /mnt/.julia/packages/NCCL/CwvFb/src/libnccl.jl:1
in expression starting at /mnt/.julia/packages/NCCL/CwvFb/src/NCCL.jl:1
in expression starting at stdin:

https://buildkite.com/julialang/lux-dot-jl/builds/6534/steps/canvas?jid=019906f7-2982-40b8-8ff9-97e4675f0e15#019906f7-2982-40b8-8ff9-97e4675f0e15/356-890

cc @simonbyrne @kshyatt

[simonbyrne]

Does this depend on a specific version of CUDA.jl?

[avik-pal]

Honestly, I am not sure where this macro is exactly defined. @maleadt / @kshyatt might know the recommended way to import this

[simonbyrne]

Wasn't this fixed by #60? Do we just need to make a new release?

[simonbyrne]

@maleadt any chance you can make a new release? (I don't have write access to this repo)

[avik-pal]

Wasn't this fixed by #60? Do we just need to make a new release?

Ah yes, I missed that (though we still want the changes to prologue since the changes will be overwritten once we regenerate the bindings)

[simonbyrne]

Good point. Can you rebase?

[maleadt]

I don't have write access to this repo

I invited you to the JuliaGPU/CUDA team, which includes NCCL.jl.

[avik-pal]

tests are failing because we are missing the cuda 13 builds for nccl. triggering one at JuliaPackaging/Yggdrasil#12061

[avik-pal]

Once JuliaRegistries/General#138616 lands we should have the proper JLLs