Update special treatment for new PauliRot rule in graph decomp #2293
Conversation
| @qml.qnode(qml.device("lightning.qubit", wires=2)) | ||
| # SKIP-CHECK-DAG: %0 = transform.apply_registered_pass "decompose-lowering" | ||
| # SKIP-CHECK{LITERAL}: func.func public @circuit_17() -> tensor<f64> attributes {decompose_gatesets | ||
| def circuit_17(): |
There was a problem hiding this comment.
This test somehow is flaky, in the sense that the graph decomp solution is not deterministic. Sometimes it contains a ppr op (despite ppr not being in the target gate set).
| @qml.qnode(qml.device("lightning.qubit", wires=4)) | ||
| # SKIP-CHECK: %0 = transform.apply_registered_pass "decompose-lowering" | ||
| # SKIP-CHECK: func.func public @circuit_18(%arg0: tensor<3xf64>) -> tensor<f64> attributes {decompose_gatesets | ||
| def circuit_18(): |
|
Hello. You may have forgotten to update the changelog!
|
| resources = qml.specs(with_qjit, level="device")()["resources"].gate_types | ||
| assert resources == expected_resources | ||
|
|
||
| @pytest.mark.xfail( |
There was a problem hiding this comment.
Note that there are 5 other tests in this file skipped to push out jax update, due to the same paulirot issue.
There was a problem hiding this comment.
The remaining 5 pytests fail with the following error inside --decompose-lowering pass:
circuit_18_1/1_QuantumCompilationStage/1_SplitMultipleTapesPass.mlir:322:25: error: 'tensor.from_elements' op failed to verify that operand types match result element type
%out_qubits_9:2 = quantum.multirz(%extracted_8) %8, %9 : !quantum.bit, !quantum.bit
^
circuit_18_1/1_QuantumCompilationStage/1_SplitMultipleTapesPass.mlir:322:25: note: see current operation: %21 = "tensor.from_elements"(%17, %18) : (i64, i64) -> tensor<1xi64>
Apparantly multirz had some issues, and the new pauli rot rule is the first rule that contains a multirz in the body so this was never hit.
There was a problem hiding this comment.
Actually no, here's the list of their corresponding errors for the 5 remaining skipped pytests (all were passing as of the previous stable release):
test_ftqc_rotxzx: pennylane has no module namedftqctest_multi_qubits: aforementioned tensor.from_elements errortest_adjoint:ValueError: Could not capture _OperatorNode(op=TemporaryAND, num_work_wire_not_available=1, work_wire_dependent=False, min_work_wires=0) without the number of wires.. After adding"TemporaryAND", "ChangeOpBasis", "Prod"to the special list (@lazypanda10117 ), we discovered that the root cause is still the aforementioned tensor.from_elements errortest_ctrl: aforementioned tensor.from_elements errortest_template_qft: aforementioned tensor.from_elements error
lazypanda10117
changed the base branch from
main
to
lazypanda10117/two-stage-pauli-rot-lowering
lazypanda10117
changed the base branch from
lazypanda10117/two-stage-pauli-rot-lowering
to
main
|
closing in favor of #2326 |
Context:
Some decomposition test cases have been disabled previously due to incompatibility with the new
qml.PauliRotlowering rules. With the new implementation and fixes in #2284, we try re-enabling the test cases, but encounter various issues in the frontend pytest and lit tests.Currently, 5 pytests in
pytest/from_plxpr/test_decompose_transform.pyand 2 lit tests inlit/test_decomposition.pystill fail.Benefits: Reenabling tests that shouldn't have been skipped.
Related GitHub Issues:
#2284
PennyLaneAI/pennylane#8765