Issue

This might be a simple oversight, but anyway, I stumbled across this while testing and benchmarking:

import numba as nb

@nb.vectorize('f8(f8)', target = 'cpu')
def foo(x):
    return x ** 2

@nb.vectorize('f8(f8)', target = 'cuda')
def bar(x):
    return x ** 2

assert foo.__name__ == 'foo'
assert bar.__name__ == 'bar'

The second assertion, i.e. the one for bar, fails, causing AttributeError: 'CUDAUFuncDispatcher' object has no attribute '__name__'. For foo, the assert passes as expected. The difference: foo is compiled for target cpu, bar is compiled for target cuda. Otherwise, everything is identical. It would be useful if numba was consistent here.

Context

• CPython == 3.10.5 (Linux, x86_64)
• numba == 0.55.2
• llvmlite == 0.38.1

Workaround

Not sure how reliable this is, but the following appears to work:

def _name(func):
    try:
        return func.__name__
    except AttributeError:
        return list(func.functions.values())[0][1].py_func.__name__[13:]

assert _name(bar) == 'bar'