Precedence of with statement versus njit kwargs

[hameerabbasi]

There is one side-effect of with statements: Their effects propagate down the call stack, even into library code. Consider the following code:

# In library.py
@njit(parallel=False)
def library_func(...):
    ...

# In user.py
with dppl.device_context(gpu):
    ...
    library_func(...)
    ...

In this case, if the with statement is preferred, the library loses control of what's executed inside library_func . Raising an error is okay, IMO, but for this reason, we shouldn't prefer what the with statement says.

Personally, I'd be okay with changing the default value of parallel to parallel=unspecified.

[diptorupd]

The design of numba-dpex has been changed to follow the Python Array API's compute-follows-data programming model. As such, the device_context is no longer used to specify the offload device. All offload decisions are handled at the typing phase based on whether the jit function is operating on dpnp arrays that have a queue attribute.

Closing as the issue is no longer relevant.