Skip to content

Catalyst v0.9.0

Latest
Latest
Compare
Choose a tag to compare
@joeycarter joeycarter released this 05 Nov 16:58
v0.9.0
d809155
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Learn about vigilant mode.

New features

  • Catalyst now supports the specification of shot-vectors when used with qml.sample measurements on the lightning.qubit device. (#1051)

    Shot-vectors allow shots to be specified as a list of shots, [20, 1, 100], or as a tuple of the form ((num_shots, repetitions), ...) such that ((20, 3), (1, 100)) is equivalent to shots=[20, 20, 20, 1, 1, ..., 1].

    This can result in more efficient quantum execution, as a single job representing the total number of shots is executed on the quantum device, with the measurement post-processing then coarse-grained with respect to the shot-vector.

    For example,

    dev = qml.device("lightning.qubit", wires=1, shots=((5, 2), 7))

@qjit
@qml.qnode(dev)
def circuit():
    qml.Hadamard(0)
    return qml.sample()
    >>> circuit()
(Array([[0], [1], [0], [1], [1]], dtype=int64),
Array([[0], [1], [1], [0], [1]], dtype=int64),
Array([[1], [0], [1], [1], [0], [1], [0]], dtype=int64))

    Note that other measurement types, such as expval and probs, currently do not support shot-vectors.

  • A new function catalyst.pipeline allows the quantum-circuit-transformation pass pipeline for QNodes within a qjit-compiled workflow to be configured. (#1131) (#1240)

    import pennylane as qml
from catalyst import pipeline, qjit

my_passes = {
    "cancel_inverses": {},
    "my_circuit_transformation_pass": {"my-option" : "my-option-value"},
}

dev = qml.device("lightning.qubit", wires=2)

@pipeline(my_passes)
@qml.qnode(dev)
def circuit(x):
    qml.RX(x, wires=0)
    return qml.expval(qml.PauliZ(0))

@qjit
def fn(x):
    return jnp.sin(circuit(x ** 2))

    pipeline can also be used to specify different pass pipelines for different parts of the same qjit-compiled workflow:

    my_pipeline = {
    "cancel_inverses": {},
    "my_circuit_transformation_pass": {"my-option" : "my-option-value"},
}

my_other_pipeline = {"cancel_inverses": {}}

@qjit
def fn(x):
    circuit_pipeline = pipeline(my_pipeline)(circuit)
    circuit_other = pipeline(my_other_pipeline)(circuit)
    return jnp.abs(circuit_pipeline(x) - circuit_other(x))

    The pass pipeline order and options can be configured globally for a qjit-compiled function, by using the circuit_transform_pipeline argument of the qjit decorator.

    my_passes = {
    "cancel_inverses": {},
    "my_circuit_transformation_pass": {"my-option" : "my-option-value"},
}

@qjit(circuit_transform_pipeline=my_passes)
def fn(x):
    return jnp.sin(circuit(x ** 2))

    Global and local (via @pipeline) configurations can coexist, however local pass pipelines will always take precedence over global pass pipelines.

    The available MLIR passes are listed and documented in the passes module documentation.

  • A peephole merge rotations pass, which acts similarly to the Python-based PennyLane merge rotations transform, is now available in MLIR and can be applied to QNodes within a qjit-compiled function. (#1162) (#1205) (#1206)

    The merge_rotations pass can be provided to the catalyst.pipeline decorator:

    from catalyst import pipeline, qjit

my_passes = {
    "merge_rotations": {}
}

dev = qml.device("lightning.qubit", wires=1)

@qjit
@pipeline(my_passes)
@qml.qnode(dev)
def g(x: float):
    qml.RX(x, wires=0)
    qml.RX(x, wires=0)
    qml.Hadamard(wires=0)
    return qml.expval(qml.PauliX(0))

    It can also be applied directly to qjit-compiled QNodes via the catalyst.passes.merge_rotations Python decorator:

    from catalyst.passes import merge_rotations

@qjit
@merge_rotations
@qml.qnode(dev)
def g(x: float):
    qml.RX(x, wires=0)
    qml.RX(x, wires=0)
    qml.Hadamard(wires=0)
    return qml.expval(qml.PauliX(0))

  • Static arguments of a qjit-compiled function can now be indicated by name via a static_argnames argument to the qjit decorator. (#1158)

    Specified static argument names will be treated as compile-time static values, allowing any hashable Python object to be passed to this function argument during compilation.

    >>> @qjit(static_argnames="y")
... def f(x, y):
...     print(f"Compiling with y={y}")
...     return x + y
>>> f(0.5, 0.3)
Compiling with y=0.3

    The function will only be re-compiled if the hash values of the static arguments change. Otherwise, re-using previous static argument values will result in no re-compilation:

    Array(0.8, dtype=float64)
>>> f(0.1, 0.3)  # no re-compilation occurs
Array(0.4, dtype=float64)
>>> f(0.1, 0.4)  # y changes, re-compilation
Compiling with y=0.4
Array(0.5, dtype=float64)

  • Catalyst Autograph now supports updating a single index or a slice of JAX arrays using Python's array assignment operator syntax. (#769) (#1143)

    Using operator assignment syntax in favor of at...op expressions is now possible for the following operations:

    • x[i] += y in favor of x.at[i].add(y)
    • x[i] -= y in favor of x.at[i].add(-y)
    • x[i] *= y in favor of x.at[i].multiply(y)
    • x[i] /= y in favor of x.at[i].divide(y)
    • x[i] **= y in favor of x.at[i].power(y)
    @qjit(autograph=True)
def f(x):
    first_dim = x.shape[0]
    result = jnp.copy(x)

    for i in range(first_dim):
      result[i] *= 2  # This is now supported

    return result
    >>> f(jnp.array([1, 2, 3]))
Array([2, 4, 6], dtype=int64)

  • Catalyst now has a standalone compiler tool called catalyst-cli that quantum-compiles MLIR input files into an object file independent of the Python frontend. (#1208) (#1255)

    This compiler tool combines three stages of compilation:

    1. quantum-opt: Performs the MLIR-level optimizations and lowers the input dialect to the LLVM dialect.
    2. mlir-translate: Translates the input in the LLVM dialect into LLVM IR.
    3. llc: Performs lower-level optimizations and creates the object file.

    catalyst-cli runs all three stages under the hood by default, but it also has the ability to run each stage individually. For example:

    # Creates both the optimized IR and an object file
catalyst-cli input.mlir -o output.o

# Only performs MLIR optimizations
catalyst-cli --tool=opt input.mlir -o llvm-dialect.mlir

# Only lowers LLVM dialect MLIR input to LLVM IR
catalyst-cli --tool=translate llvm-dialect.mlir -o llvm-ir.ll

# Only performs lower-level optimizations and creates object file
catalyst-cli --tool=llc llvm-ir.ll -o output.o

    Note that catalyst-cli is only available when Catalyst is built from source, and is not included when installing Catalyst via pip or from wheels.

  • Experimental integration of the PennyLane capture module is available. It currently only supports quantum gates, without control flow. (#1109)

    To trigger the PennyLane pipeline for capturing the program as a Jaxpr, simply set experimental_capture=True in the qjit decorator.

    import pennylane as qml
from catalyst import qjit

dev = qml.device("lightning.qubit", wires=1)

@qjit(experimental_capture=True)
@qml.qnode(dev)
def circuit():
    qml.Hadamard(0)
    qml.CNOT([0, 1])
    return qml.expval(qml.Z(0))

Improvements

  • Multiple qml.sample calls can now be returned from the same program, and can be structured using Python containers. For example, a program can return a dictionary of the form return {"first": qml.sample(), "second": qml.sample()}. (#1051)

  • Catalyst now ships with null.qubit, a Catalyst runtime plugin that mocks out all functions in the QuantumDevice interface. This device is provided as a convenience for testing and benchmarking purposes. (#1179)

    qml.device("null.qubit", wires=1)

@qml.qjit
@qml.qnode(dev)
def g(x):
    qml.RX(x, wires=0)
    return qml.probs(wires=[0])

  • Setting the seed argument in the qjit decorator will now seed sampled results, in addition to mid-circuit measurement results. (#1164)

    dev = qml.device("lightning.qubit", wires=1, shots=10)

@qml.qnode(dev)
def circuit(x):
    qml.RX(x, wires=0)
    m = catalyst.measure(0)

    if m:
        qml.Hadamard(0)

    return qml.sample()

@qml.qjit(seed=37, autograph=True)
def workflow(x):
    return jnp.squeeze(jnp.stack([circuit(x) for i in range(4)]))
    >>> workflow(1.8)
Array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
       [1, 1, 0, 0, 1, 1, 0, 0, 1, 0],
       [0, 0, 1, 0, 1, 1, 0, 0, 1, 1],
       [1, 1, 1, 0, 0, 1, 1, 0, 1, 1]], dtype=int64)
>>> workflow(1.8)
Array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
       [1, 1, 0, 0, 1, 1, 0, 0, 1, 0],
       [0, 0, 1, 0, 1, 1, 0, 0, 1, 1],
       [1, 1, 1, 0, 0, 1, 1, 0, 1, 1]], dtype=int64)

    Note that statistical measurement processes such as expval, var, and probs are currently not affected by seeding when shot noise is present.

  • The cancel_inverses MLIR compilation pass (-remove-chained-self-inverse) now supports cancelling all Hermitian gates, as well as adjoints of arbitrary unitary operations. (#1136) (#1186) (#1211)

    For the full list of supported Hermitian gates please see the cancel_inverses documentation in catalyst.passes.

  • Support is expanded for backend devices that exclusively return samples in the measurement basis. Pre- and post-processing now allows qjit to be used on these devices with qml.expval, qml.var and qml.probs measurements in addition to qml.sample, using the measurements_from_samples transform. (#1106)

  • Scalar tensors are eliminated from control flow operations in the program, and are replaced with bare scalars instead. This improves compilation time and memory usage at runtime by avoiding heap allocations and reducing the amount of instructions. (#1075)

  • Catalyst now supports NumPy 2.0. (#1119) (#1182)

  • Compiling QNodes to asynchronous functions will no longer print to stderr in case of an error. (#645)

  • Gradient computations have been made more efficient, as calling gradients twice (with the same gradient parameters) will now only lower to a single MLIR function. (#1172)

  • qml.sample() and qml.counts() on lightning.qubit/kokkos can now be seeded with qjit(seed=...). (#1164) (#1248)

  • The compiler pass -remove-chained-self-inverse can now also cancel adjoints of arbitrary unitary operations (in addition to the named Hermitian gates). (#1186) (#1211)

  • Add Lightning-GPU support to Catalyst docs and update tests. (#1254)

Breaking changes

  • The static_size field in the AbstractQreg class has been removed. (#1113)

    This reverts a previous breaking change.

  • Nesting QNodes within one another now raises an error. (#1176)

  • The debug.compile_from_mlir function has been removed; please use debug.replace_ir instead. (#1181)

  • The compiler.last_compiler_output function has been removed; please use compiler.get_output_of("last", workspace) instead. (#1208)

Bug fixes

  • Fixes a bug where the second execution of a function with abstracted axes is failing. (#1247)

  • Fixes a bug in catalyst.mitigate_with_zne that would lead to incorrectly extrapolated results. (#1213)

  • Fixes a bug preventing the target of qml.adjoint and qml.ctrl calls from being transformed by AutoGraph. (#1212)

  • Resolves a bug where mitigate_with_zne does not work properly with shots and devices supporting only counts and samples (e.g., Qrack). (#1165)

  • Resolves a bug in the vmap function when passing shapeless values to the target. (#1150)

  • Fixes a bug that resulted in an error message when using qml.cond on callables with arguments. (#1151)

  • Fixes a bug that prevented taking the gradient of nested accelerate callbacks. (#1156)

  • Fixes some small issues with scatter lowering: (#1216) (#1217)

    • Registers the func dialect as a requirement for running the scatter lowering pass.
    • Emits error if %input, %update and %result are not of length 1 instead of segfaulting.

  • Fixes a performance issue with catalyst.vmap, where the root cause was in the lowering of the scatter operation. (#1214)

  • Fixes a bug where conditional-ed single gates cannot be used in qjit, e.g. qml.cond(x > 1, qml.Hadamard)(wires=0). (#1232)

Internal changes

  • Removes deprecated PennyLane code across the frontend. (#1168)

  • Updates Enzyme to version v0.0.149. (#1142)

  • Adjoint canonicalization is now available in MLIR for CustomOp and MultiRZOp. It can be used with the --canonicalize pass in quantum-opt. (#1205)

  • Removes the MemMemCpyOptPass in llvm O2 (applied for Enzyme), which reduces bugs when running gradient-like functions. (#1063)

  • Bufferization of gradient.ForwardOp and gradient.ReverseOp now requires three steps: gradient-preprocessing, gradient-bufferize, and gradient-postprocessing. gradient-bufferize has a new rewrite for gradient.ReturnOp. (#1139)

  • A new MLIR pass detensorize-scf is added that works in conjunction with the existing linalg-detensorize pass to detensorize input programs. The IR generated by JAX wraps all values in the program in tensors, including scalars, leading to unnecessary memory allocations for programs compiled to CPU via the MLIR-to-LLVM pipeline. (#1075)

  • Importing Catalyst will now pollute less of JAX's global variables by using LoweringParameters. (#1152)

  • Cached primitive lowerings is used instead of a custom cache structure. (#1159)

  • Functions with multiple tapes are now split with a new mlir pass --split-multiple-tapes, with one tape per function. The reset routine that makes a measurement between tapes and inserts an X gate if measured one is no longer used. (#1017) (#1130)

  • Prefer creating new qml.devices.ExecutionConfig objects over using the global qml.devices.DefaultExecutionConfig. Doing so helps avoid unexpected bugs and test failures in case the DefaultExecutionConfig object becomes modified from its original state. (#1137)

  • Remove the old QJITDevice API. (#1138)

  • The device-capability loading mechanism has been moved into the QJITDevice constructor. (#1141)

  • Several functions related to device capabilities have been refactored. (#1149)

    In particular, the signatures of get_device_capability, catalyst_decompose, catalyst_acceptance, and QJITDevice.__init__ have changed, and the pennylane_operation_set function has been removed entirely.

  • Catalyst now generates nested modules denoting quantum programs. (#1144)

    Similar to MLIR's gpu.launch_kernel function, Catalyst, now supports a call_function_in_module. This allows Catalyst to call functions in modules and have modules denote a quantum kernel. This will allow for device-specific optimizations and compilation pipelines.

    At the moment, no one is using this. This is just the necessary scaffolding to support device-specific transformations. As such, the module will be inlined to preserve current semantics. However, in the future, we will explore lowering this nested module into other IRs/binary formats and lowering call_function_in_module to something that can dispatch calls to another runtime/VM.

Contributors

This release contains contributions from (in alphabetical order):

Joey Carter,
Spencer Comin,
Amintor Dusko,
Lillian M.A. Frederiksen,
Sengthai Heng,
David Ittah,
Mehrdad Malekmohammadi,
Vincent Michaud-Rioux,
Romain Moyard,
Erick Ochoa Lopez,
Daniel Strano,
Raul Torres,
Paul Haochen Wang.

Assets 14
Loading