running julia's CUDA through pyjulia

[ArturPrzybysz]

Hello! I am trying to use julia's CUDA package in Python, however I run into troubles.

The code that I run is following:

from julia.api import Julia
import juliapkg as pkg

executable = pkg.executable()
jl = Julia(runtime=executable)  # , compiled_modules=False)

jl.eval('''
import Pkg
Pkg.add("CUDA")
using CUDA

function gpu_compute()
    return CUDA.rand(Float16, 2, 2) * CUDA.rand(Float16, 2, 2)
end
''')

result = jl.eval('gpu_compute()')
print(result)

and I run it using python-jl.

The output that I get is a SEGFAULT:

signal (11): Segmentation fault
in expression starting at none:4
unknown function (ip: (nil))
Allocations: 2904 (Pool: 2894; Big: 10); GC: 0
Segmentation fault (core dumped)

When I replace line 5: jl = Julia(runtime=executable) # , compiled_modules=False) with jl = Julia(runtime=executable , compiled_modules=False) and run it simply with Python I get:

Reason: unsupported dynamic function invocation (call to pointerset(ptr::Core.LLVMPtr{T, A}, x::T, i::I, ::Val{align}) where {T, A, I, align} @ LLVM.Interop ~/.julia/packages/LLVM/RFlwq/src/interop/pointer.jl:46)
Stacktrace:
 [1] unsafe_store!
   @ ~/.julia/packages/LLVM/RFlwq/src/interop/pointer.jl:88
 [2] arrayset_bits
   @ ~/.julia/packages/CUDA/tVtYo/src/device/array.jl:134
 [3] #arrayset
   @ ~/.julia/packages/CUDA/tVtYo/src/device/array.jl:127
 [4] setindex!
   @ ~/.julia/packages/CUDA/tVtYo/src/device/array.jl:166
 [5] kernel
   @ ~/.julia/packages/CUDA/tVtYo/src/random.jl:58
Hint: catch this exception as `err` and call `code_typed(err; interactive = true)` to introspect the erronous code with Cthulhu.jl' occurred while calling julia code:
gpu_compute()

How can I get CUDA to work here?

[ArturPrzybysz]

Pkg.status()
  [052768ef] CUDA v4.4.0
  [438e738f] PyCall v1.96.0

> pip show julia
Name: julia
Version: 0.6.1
Summary: Julia/Python bridge with IPython support.

nvidia-smi outputs that the driver version is 530.41.03, CUDA Version: 12.1 and I have GeForce RTX 4090.

Julia version is 1.9.1

The julia code using CUDA runs fine on the same julia binary as the one used by pyjulia.