Failure to qjit two workflows transformed by `clifford_t_decomposition` and gridsynth when using AOT capture

[albi3ro]

@qml.qjit()
@partial(qml.transforms.clifford_t_decomposition, method="gridsynth")
@qml.qnode(qml.device('lightning.qubit', wires=1))
def c():
    qml.H(0)
    #qml.RZ(x, 0)
    qml.RZ(0.5, 0)
    return qml.expval(qml.X(0))

@qml.qjit()
@partial(qml.transforms.clifford_t_decomposition, method="gridsynth")
@qml.qnode(qml.device('lightning.qubit', wires=1))
def c2(x):
    qml.RZ(x, 0)
    return qml.expval(qml.X(0))

c2(0.5)

If one qnode with AOT capture transformed by clifford_t_decomposition with grid_synth has already been defined, another new qnode transformed by the same transform cannot be compiled.

This does not happen if the first qjit object is captured on the first call. It does not matter if the second workflow is AOT or lazily captured.

While it looks like method="sk" is causing concretization problems with parametrized inputs, it does not cause failures with AOT capture.

This is mostly just really weird.

[dime10]

Agreed, that is weird... I wonder if some of the caching used in the ross decomposition is responsible for this 🤔 (cc @sengthai)

[sengthai]

Yes, @dime10 you are right. I disabled the cache in the Ross decomposition, and this problem was solved. 👀

[josephleekl]

I'm investigating this a bit more, disabling the cache indeed fixed the behaviour above, but turns out even just this fails:

@qml.qjit()
@partial(qml.transforms.clifford_t_decomposition, method="gridsynth")
@qml.qnode(qml.device('lightning.qubit', wires=1))
def c():
    qml.RZ(0.5, 0)   
    qml.T(0)
    qml.RZ(0.5, 0)
    return qml.expval(qml.X(0))


print(c())

which doesn't seem to be solved by disabling the cache.

update: turns out there are more than one cached functions, I need to disable them together.