Memory segfaults with initializing 'lightning.kokkos' after 'lightning.qubit' on macOS (x86_64)

[maliasadi]

Running the following script on macOS x86_64 using PL-Lightning and PL-Lightning-Kokkos version 0.33.1 (installed via pip) will lead to memory segfaults. These segfaults are related to the initialization of the "lightning.kokkos" state vector.

import pennylane as qml
import pytest

def workflow1():
    @qml.qnode(qml.device("lightning.qubit", wires=1))
    def f(x):
        qml.RX(x, wires=0)
        return qml.expval(qml.PauliZ(wires=0))

    f(0.2)

def workflow2():
    @qml.qnode(qml.device("lightning.kokkos", wires=1))
    def g(x):
        qml.RX(x, wires=0)
        return qml.expval(qml.PauliZ(wires=0))

    g(0.2)

if __name__ == "__main__":
    pytest.main(["-x", __file__])

A few notes:

• Reordering workflow1 and workflow2 would resolve segfaults.
• Not a deterministic issue; you may not be able to reproduce it with running the test for the first time.
• Pre-initializing the 'lightning.kokkos' state-vector by calling qml.device("lightning.kokkos", ...) in workflow1 would resolve the issue.

I could reproduce this by building lightning.qubit and lightning.kokkos on the machine from src, and with using lldb, it seems that the issue is propagated from Kokkos vector kernels. I didn't dig more to see if this is related to the Lightning implementation or Kokkos kernels.

[dime10]

Does anyone know if this is still an issue?

[erick-xanadu]

I noticed that our wheels do not use the same kokkos version than the one used in check-catalyst, I'll keep a note of this and maybe we can investigate after StatePrep is merged. See here: PennyLaneAI/catalyst@507302c