Reactant: No CUDA Driver v13 support on aarch64

[oschulz]

I got a warning on a Grace-Hopper system that the NVIDIA driver was too old (Reactant worked, though). I updated it, but now it's seems to be too new :-) :

julia> using Reactant_jll
Precompiling Reactant_jll...
Info Given Reactant_jll was explicitly requested, output will be shown live 
┌ Debug: Detected CUDA Driver version 13.0.0
└ @ Reactant_jll ~/.julia-aarch64/packages/Reactant_jll/yJPxA/.pkg/platform_augmentation.jl:60
┌ Warning: CUDA version 13.0.0 in /lib/aarch64-linux-gnu/libcuda.so.1 not supported with this version of Reactant (min supported: 12.6)
└ @ Reactant_jll ~/.julia-aarch64/packages/Reactant_jll/yJPxA/.pkg/platform_augmentation.jl:103
  1 dependency successfully precompiled in 2 seconds. 26 already precompiled.
  1 dependency had output during precompilation:
┌ Reactant_jll
│  [Output was shown above]
└  

julia> using Reactant
Precompiling Reactant...
  2 dependencies successfully precompiled in 37 seconds. 79 already precompiled.

julia> Reactant.devices()
1-element Vector{Reactant.XLA.PJRT.Device}:
 Reactant.XLA.PJRT.Device(Ptr{Nothing} @0x000000003d1a5f60, "CPU:0 cpu")

julia> versioninfo()
Julia Version 1.11.7
Commit f2b3dbda30a (2025-09-08 12:10 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (aarch64-linux-gnu)
  CPU: 72 × unknown
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, neoverse-v2)
Threads: 36 default, 0 interactive, 18 GC (on 72 virtual cores)
Environment:
  JULIA_DEBUG = Reactant_jll
  JULIA_PKG_SERVER = pkg.julialang.org
  JULIA_NUM_THREADS = 36

[wsmoses]

Cc @giordano 13 should be supported? We at least build that jll

[oschulz]

nvidia-smi says that I have driver version 580.82.07 on that system. On my laptop I have 580.65.06 and Reactant is happy. What confuses me is the v13 above - that looks like a CUDA toolkit version, not like a CUDA driver version, right?

[giordano]

I'm confused. Is there anything wrong in https://github.com/JuliaBinaryWrappers/Reactant_jll.jl/blob/1e44790a83dc5e646cd5dfa5c8e5b6788d908326/.pkg/platform_augmentation.jl#L95-L105? v13.0 should enter this branch, right?

[oschulz]

Are there any check I can run to find out more, @giordano ?

[wsmoses]

can you in any julia session basically paste the equivalent of that linked code (maybe add more prints to figure out why it dislikes the branch)?

[giordano]

]dev Reactant_jll and debug the file I showed above. The path taken on your computer doesn't make sense to me.

[oschulz]

"/lib/aarch64-linux-gnu/libcuda.so.1" is a symlink to "/usr/lib/aarch64-linux-gnu/libcuda.so.580.82.07" which in installed via APT package "libnvidia-compute-580". CUDA.jl works, btw.

[oschulz]

can you in any julia session basically paste the equivalent of that linked code

I ran it manually, copy/paste:

# [...] code that results in
cuname = "libcuda.so.1"

julia> handle = Libdl.dlopen(cuname)
Ptr{Nothing} @0x00000000044446e0

julia> current_cuda_version = cuDriverGetVersion(handle)
        
v"13.0.0"

julia> path = Libdl.dlpath(handle)
"/lib/aarch64-linux-gnu/libcuda.so.1"

[wsmoses]

what is:
v"13.0" <= current_cuda_version < v"14" && arch(platform) == "x86_64", and/or specifically why does it fail

wait I see why -- we never built an aarch64 cuda 13 jll, just x86

[giordano]

Ah, yes, we don't have the aarch64 build.

[oschulz]

It that a big thing to do, or something just needs to be triggered?

[wsmoses]

we hit a bazel problem early on, but that likely is now fixed (or fixable by posting an issue upstream): https://github.com/JuliaPackaging/Yggdrasil/blob/4efe2797fba117ce8da2fe497c7c409dc2bec53c/R/Reactant/build_tarballs.jl#L499

Remove the continue and see how it fails? (and pending the results we can file issues to get things fixed)

[oschulz]

Remove the continue and see how it fails?

Uh, me or @giordano ?