🐛 Cannot train on IQM's crystal device

[lmarti-dev]

System and Environment Information

OS: Windows 10
MQT Version: 2.3.0 on Python 3.11, installed with uv
Qiskit 2.1.2

Bug Description

The following code (straight from the documentation) fails:

from mqt.predictor.rl import Predictor as RL_Predictor
from mqt.bench.targets import get_device

device = get_device("iqm_crystal_20")
rl_pred = RL_Predictor(device=device, figure_of_merit="expected_fidelity")
rl_pred.train_model(timesteps=100000)

with error (I shortened the path a bit)

File ".viqm\Lib\site-packages\mqt\predictor\rl\parsing.py", line 153, in get_bqskit_native_gates
    raise ValueError(msg)
ValueError: The 'r' gate of device 'iqm_crystal_20' is not supported in BQSKIT.       

It happens immediately after the training starts. The "r" gate is indeed defined in IQM's definition of Crystal (aka Garnet). I assume that the resolution of this error depends on BQSKIT.

If you look at Qiskit's definition of the RGate (the "r" gate in question), it seems to be equivalent to the BQSKIT U1qGate. Then, the fix is a simple addition to the gate_map in parsing.py.

If one goes ahead and adds the gate to the dict, this allows the training to proceed to the next error: U1qGate._qasm_name is an unallowed value by qasm, because it starts with a capital U. This, again, can be solved by pretending that gate is the r gate to qiskit, qasm, and bqskit, though the fix is rather ugly.

from bqskit.ir.gates.parameterized.u1q import U1qGate
from qiskit.qasm2 import CustomInstruction
from qiskit.circuit.library import RGate

import qiskit.qasm2

U1qGate._qasm_name = "r"

qiskit.qasm2.LEGACY_CUSTOM_INSTRUCTIONS = qiskit.qasm2.LEGACY_CUSTOM_INSTRUCTIONS + (
    CustomInstruction("r", 2, 1, RGate,builtin=True),
)

This fixes the gate issue. The training eventually fails seemingly because the IQM device is not bidirectional, causing a KeyError in reward.py.

I guess the issue doesn't lie in this repo. I'd be glad if there was a pretrained model for the IQM device.

Steps to Reproduce

Run the code from this page for iqm_crystal_20.

from mqt.predictor.rl import Predictor as RL_Predictor
from mqt.bench.targets import get_device

device = get_device("iqm_crystal_20")
rl_pred = RL_Predictor(device=device, figure_of_merit="expected_fidelity")
rl_pred.train_model(timesteps=100000)

[burgholzer]

Thanks for raising this issue and poking around the library trying to fix this.
While not all of it might be the Predictor's fault per-se, I still believe this is something that we should fix.
I'll take a quick look to better assess the situation and how large of a fix this would need from our side.

[lmarti-dev]

Thanks for investigating. I could also create an issue related to the bidirectionality, though I'm not sure if it's a bug in itself. The program indeed warns that results might not be optimal, but it ends with a KeyError, which is unexpected. The fix is actually fairly trivial, but I'm not sure it's something you want to fix at all.