Unrecognized options: {'target'} when using @jit(target="cuda")

user_17391404 0
Viewed 1205

I have optimized some python code using the decorator @jit from numba library. However, I want to indicate to @jit to use my GPU device explicitly. From: Difference between @cuda.jit and @jit(target='gpu'), I understand that I need to use @jit(target="cuda") to do it.

I tried to do it by doing something like this:

from numba import jit, cuda

@jit(target='cuda')  # The code runs normally without (target='cuda')
def function(args):
    # some code

And I got the following error:

KeyError: "Unrecognized options: {'target'}. Known options are dict_keys(['_nrt', 'boundscheck', 'debug', 'error_model', 'fastmath', 'forceinline', 'forceobj', 'inline', 'looplift', 'no_cfunc_wrapper', 'no_cpython_wrapper', 'no_rewrites', 'nogil', 'nopython', 'parallel', 'target_backend'])"

I have read this: How to run numba.jit decorated function on GPU? but the solution did not work.

I would appreciate some help to make @jit(target='cuda') work without rewriting the code using @cuda.jit as this last one is for writing CUDA kernel in Python and compile and run it.

Many thanks in advance!

numba cuda python
2 Answers

I faced a similar issue and got through by changing target to target_backend

So, my decorator goes like @jit(target_backend="cuda")

It's just a quick-fix that worked for me, I didn't dig further

user_65592860

AFAIK, CUDA targets are not supported anymore for jit andnjit (it was supported few years ago as a wrapper to cuda.jit). It is not documented. There is however such parameter for numba.vectorize and numba.guvectorize. In the Numba code, one can see that there is a parameter called target_backend which is obviously not used (anymore?). There is a parameter call _target which is read but not meant to be used directly by end-users. Additionally, it calls cuda.jit in the end anyway. Such part of the code seems to be dead.

If you want to write a GPU-based code, then please use numba.vectorize and numba.guvectorize or cuda.jit (the last is quite low-level compared to the two first).

user_129395570
Related