Preparing Reactant GPU backend on online CPU‑only machine for offline GPU deployment

[longhua8800w]

Summary

My setup:

• Machine A: internet, no GPU (build/prepare)
• Machine B: GPU, no internet (runtime)

I can:

• use Pkg on A to install CUDA.jl, Reactant, Enzyme, etc.
• copy the depot/envs from A to B
• run CUDA.jl + Enzyme directly on B (CuArray works)
• run Reactant on CPU on B

But I cannot reliably get Reactant's GPU backend working on B. Calls
like Reactant.set_default_backend("gpu") fail with errors such as
"no GPU client found / no functional client". It seems some PJRT/CUDA
pieces are only downloaded/initialized when running Reactant on a machine
that already has a GPU.

Feature request

It would be very helpful if Reactant provided either:

• a helper like

  Reactant.prepare_gpu_backend(target="cuda", cuda_version="12.4")

  that I can run on A (even without a GPU) to download/materialize
  everything the CUDA/PJRT backend needs, or

• a documented sequence of steps/env vars that let me:

  1. fully prepare the GPU backend on A
  2. copy depot/envs to B
  3. use Reactant's GPU backend on B completely offline

Even a clearly documented "you must run X once on a machine with a GPU,
then copy depot" workflow would already help a lot for offline / air‑gapped
environments.

[Pangoraw]

Once finished, this is one application for #2779

[wsmoses]

@longhua8800w does this section of the docs cover your issues, or are you hitting something else? https://enzymead.github.io/Reactant.jl/dev/introduction/configuration#Installing-Reactant-on-GPU-Servers-without-Internet

[longhua8800w]

Once finished, this is one application for #2779

Thanks for the clarification — this is really important work, and it should make Reactant much more robust for offline GPU deployment and environment migration. I appreciate your help!

[longhua8800w]

@longhua8800w does this section of the docs cover your issues, or are you hitting something else? https://enzymead.github.io/Reactant.jl/dev/introduction/configuration#Installing-Reactant-on-GPU-Servers-without-Internet

Thanks! I'll test it and report back.

[giordano]

If you don't have libcuda.so.1 on machine A, you'll need to fake it:

gcc -shared -Wl,-soname=libcuda.so.1 -o libcuda.so.1 /dev/null
LD_LIBRARY_PATH='.' julia ...

[longhua8800w]

After setting gpu = "cuda" and gpu_version = "12.9" in Project.toml and using a fake libcuda.so.1 to allow precompilation on a machine without GPU, Reactant itself now precompiles successfully. However, ReactantCUDAExt is always skipped due to a circular dependency:

text
┌ Reactant → ReactantSparseArraysExt
└─ Reactant → ReactantDatesExt
Environment:

Julia: 1.10.10

Reactant: v0.2.254

Reactant_jll: v0.0.371

Does this circular dependency prevent ReactantCUDAExt from being usable at runtime on the GPU machine? Or is it safe to copy the depot as-is and let it load dynamically?

[longhua8800w]

Update: After further investigation, the setup is now working end-to-end on the GPU machine.

The key steps that made it work:

Setting gpu = "cuda" and gpu_version = "12.9" in Project.toml

Using a fake libcuda.so.1 on the CPU-only machine to allow precompilation

After copying the depot to the GPU machine, fixing execute permissions on all files under the artifacts/ directory (find ... -type f -exec chmod +x {} ;) — this turned out to be the critical missing step

To answer my own question: the circular dependency warning for ReactantCUDAExt does not prevent runtime GPU usage — it loads dynamically and works correctly on the GPU machine.

Overall the offline deployment workflow is functional, though it requires some manual steps. I'm looking forward to PR #2779 and related work making this more robust and seamless in the future. Thanks again for the guidance!