CUDA runtime_call test fails: stub handler invoked despite CUDA_ENABLED build

[blasphemetheus]

The test runtime_call with CUDA client (device↔host copies) at exla/test/exla/defn/runtime_call_test.exs:34 fails with:

** (RuntimeError) EXLA was not compiled with CUDA support.
This error means your EXLA compilation is out of sync with your libexla.so NIF.

Even after a clean make clean && mix compile with nvcc available and -DCUDA_ENABLED correctly set in CFLAGS.

Analysis

The .o file is compiled correctly — _impl is present, _stub is absent. But the final libexla.so contains both symbols, and the stub wins the XLA FFI handler registration.

nm cache/libexla.so | grep cuda_stub
00000000000648d0 t ...exla_runtime_callback_cuda_stub  # should not exist

Reproduction

cd exla
make clean && mix compile
mix test test/exla/defn/runtime_call_test.exs:34

Environment

• GPU: NVIDIA GeForce RTX 5090 (SM 12.0a)
• CUDA 12.8 / cuDNN 9.13.0
• XLA 0.10.0 / EXLA 0.11.0

Introduced in PR #1677. Fails consistently on main.

[polvalente]

Can you check if setting EXLA_FORCE_REBUILD=true or partial fixes this?

[blasphemetheus]

yeah, was working on multiple versions of nx in a git worktree when I came across this so very plausible user error due to doing that (I remember there was some weird stuff happening with make seeing new versions of files and not updating them when I assumed it should happening in this same session). I'll double check this is an actual issue and try the flags when I get back to this same setup on Sunday or Monday

[blasphemetheus]

yeah, couldn't reproduce after I got rid of the git worktree i had going, just worktree build cache contamination

[blasphemetheus]

oops close as not an issue if thats possible