Create JIT compiled device

[mlxd]

Feature details

As PennyLane has support for multiple plugins to provide simulation of quantum circuits using a variety of different software and hardware devices, there can be many ways to reduce call overheads for circuits with many repeated gate-calls. One such method is the use of just-in-time compilation techniques. This task requires the use of the Numba library to develop a PennyLane device which allows JIT compilation of gate-calls. The device is to be called "numba.qubit", and will be compared against the existing implementations for "default.qubit" and the C++ backend "lightning.qubit".

Implementation

• Develop a new PennyLane plugin device named "numba.qubit". This can be modelled on the existing devices (e.g. see TensorFlow, or JAX).
• Add testing to ensure the gates perform as expected (see here for existing interface tests).

How important would you say this feature is?

1: Not important. Would be nice to have.

Additional information

To understand the impact of JIT compilation on running circuits, the new device should also be benchmarked against the existing "default.qubit" and "lightning.qubit" devices (see here for info on installing and using lightning).

[arshpreetsingh]

Hi @mlxd
Quick Question, Will there be new interface as well something like pennylane/interface/numba.py ?

[josh146]

@arshpreetsingh nope, Numba does not support autodifferentiability 🙂

Instead, the Numba device can be used with existing interfaces. E.g., using the Torch interface with the Numba device will allow for users to use this new device with torch.tensor objects.