[BUG] Graph decomposition raises extraneous warnings

[dime10]

Problem description

When the graph decomposition cannot generate a solution to the given gate set, it raises a warning. This is sensible, however it will raise additional warnings for gates that the user hasn't even specified. In the provided example, a total of five warnings are raised:

"... unable to find a decomposition for {'Hadamard'} ..."
"Falling back to the legacy decomposition system."
"Operator RZ does not define a decomposition ..."
"Operator RX does not define a decomposition ..."
"Operator RZ does not define a decomposition ..."

The first two seem sensible while the latter three seem excessive since they pertain to operations within the graph (neither entry nor exit nodes), for which the user has no context.

Source code

from functools import partial

import pennylane as qml

qml.capture.enable()
qml.decomposition.enable_graph()

@qml.qjit
@partial(qml.transforms.decompose, gate_set={qml.GlobalPhase})
@qml.qnode(qml.device("lightning.qubit", wires=2))
def circuit(x):
    qml.Hadamard(x)
    return qml.state()

circuit(0)

Taken from https://github.com/PennyLaneAI/catalyst/blame/6a19ecf7f23c0f3f22bd5e43a0783394bf9d285d/frontend/test/pytest/from_plxpr/test_decompose_transform.py#L54

[astralcai]

Here's what happened:

• The graph does not find a path from Hadamard to the target gate set, triggering the first warning.
• The system falls back to using Hadamard.decomposition() which produced the RX and the RZ gates.
• But the graph also never found a decomposition rule for RX or RZ to the target gate set, so a warning is raised for each

[dime10]

Thanks @astralcai! Do you think the second set of warnings are useful to a user then?

[astralcai]

Thanks @astralcai! Do you think the second set of warnings are useful to a user then?

Now that I think about it, probably not, since the first warning would've already notified the user about something not going as planned...