Source code for pennylane.control_flow.for_loop

# Copyright 2025 Xanadu Quantum Technologies Inc.

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at

#     http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""For loop."""
import functools
from typing import Literal

from pennylane import capture
from pennylane.capture import FlatFn, enabled
from pennylane.capture.dynamic_shapes import register_custom_staging_rule
from pennylane.compiler.compiler import AvailableCompilers, active_compiler

from ._loop_abstract_axes import (
    add_abstract_shapes,
    get_dummy_arg,
    handle_jaxpr_error,
    loop_determine_abstracted_axes,
    validate_no_resizing_returns,
)


[docs] def for_loop( start, stop=None, step=1, *, allow_array_resizing: Literal["auto", True, False] = "auto" ): """for_loop([start, ]stop[, step]) A :func:`~.qjit` compatible for-loop for PennyLane programs. When used without :func:`~.qjit`, this function will fall back to a standard Python for loop. This decorator provides a functional version of the traditional for-loop, similar to `jax.cond.fori_loop <https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.fori_loop.html>`__. That is, any variables that are modified across iterations need to be provided as inputs/outputs to the loop body function: - Input arguments contain the value of a variable at the start of an iteration. - output arguments contain the value at the end of the iteration. The outputs are then fed back as inputs to the next iteration. The final iteration values are also returned from the transformed function. The semantics of ``for_loop`` are given by the following Python pseudo-code: .. code-block:: python def for_loop(start, stop, step, loop_fn, *args): for i in range(start, stop, step): args = loop_fn(i, *args) return args Unlike ``jax.cond.fori_loop``, the step can be negative if it is known at tracing time (i.e., constant). If a non-constant negative step is used, the loop will produce no iterations. .. note:: This function can be used in the following different ways: 1. ``for_loop(stop)``: Values are generated within the interval ``[0, stop)`` 2. ``for_loop(start, stop)``: Values are generated within the interval ``[start, stop)`` 3. ``for_loop(start, stop, step)``: Values are generated within the interval ``[start, stop)``, with spacing between the values given by ``step`` Args: start (int, optional): starting value of the iteration index. The default start value is ``0`` stop (int): upper bound of the iteration index step (int, optional): increment applied to the iteration index at the end of each iteration. The default step size is ``1`` Keyword Args: allow_array_resizing (Literal["auto", True, False]): How to handle arrays with dynamic shapes that change between iterations Returns: Callable[[int, ...], ...]: A wrapper around the loop body function. Note that the loop body function must always have the iteration index as its first argument, which can be used arbitrarily inside the loop body. As the value of the index across iterations is handled automatically by the provided loop bounds, it must not be returned from the function. .. seealso:: :func:`~.while_loop`, :func:`~.qjit` **Example** .. code-block:: python dev = qml.device("lightning.qubit", wires=1) @qml.qnode(dev) def circuit(n: int, x: float): @qml.for_loop(0, n, 1) def loop_rx(i, x): # perform some work and update (some of) the arguments qml.RX(x, wires=0) # update the value of x for the next iteration return jnp.sin(x) # apply the for loop final_x = loop_rx(x) return qml.expval(qml.Z(0)) >>> circuit(7, 1.6) array(0.97926626) ``for_loop`` is also :func:`~.qjit` compatible; when used with the :func:`~.qjit` decorator, the for loop will not be unrolled, and instead will be captured as-is during compilation and executed during runtime: >>> qml.qjit(circuit)(7, 1.6) Array(0.97926626, dtype=float64) .. note:: Please see the Catalyst :doc:`quickstart guide <catalyst:dev/quick_start>`, as well as the :doc:`sharp bits and debugging tips <catalyst:dev/sharp_bits>` page for an overview of using quantum just-in-time compilation. .. details:: :title: Usage Details .. note:: The following examples may yield different outputs depending on how the workflow function is executed. For instance, the function can be run directly as: >>> arg = 2 >>> workflow(arg) Alternatively, the function can be traced with ``jax.make_jaxpr`` to produce a JAXPR representation, which captures the abstract computational graph and generates the abstract shapes. The resulting JAXPR can then be evaluated using ``qml.capture.eval_jaxpr``: >>> jaxpr = jax.make_jaxpr(workflow)(arg) >>> qml.capture.eval_jaxpr(jaxpr.jaxpr, jaxpr.consts, arg) The following discussion applies to the experimental capture infrastructure, which can be turned on by ``qml.capture.enable()``. See the ``capture`` module for more information. A dynamically shaped array is an array whose shape depends on an abstract value. This is an experimental jax mode that can be turned on with: >>> import jax >>> import jax.numpy as jnp >>> jax.config.update("jax_dynamic_shapes", True) >>> qml.capture.enable() ``allow_array_resizing="auto"`` will try and choose between the following two possible modes. If the needed mode is ``allow_array_resizing=True``, then this will require re-capturing the loop, potentially taking more time. When working with dynamic shapes in a ``for_loop``, we have two possible options. ``allow_array_resizing=True`` treats every dynamic dimension as independent. .. code-block:: python @qml.for_loop(3, allow_array_resizing=True) def f(i, x, y): return jnp.hstack([x, y]), 2*y def workflow(i0): x0, y0 = jnp.ones(i0), jnp.ones(i0) return f(x0, y0) Even though ``x`` and ``y`` are initialized with the same shape, the shapes no longer match after one iteration. In this circumstance, ``x`` and ``y`` can no longer be combined with operations like ``x * y``, as they do not have matching shapes. With ``allow_array_resizing=False``, anything that starts with the same dynamic dimension must keep the same shape pattern throughout the loop. .. code-block:: python @qml.for_loop(3, allow_array_resizing=False) def f(i, x, y): return x * y, 2*y def workflow(i0): x0 = jnp.ones(i0) y0 = jnp.ones(i0) return f(x0, y0) Note that with ``allow_array_resizing=False``, all arrays can still be resized together, as long as the pattern still matches. For example, here both ``x`` and ``y`` start with the same shape, and keep the same shape as each other for each iteration. .. code-block:: python @qml.for_loop(3, allow_array_resizing=False) def f(i, x, y): x = jnp.hstack([x, y]) return x, 2*x def workflow(i0): x0 = jnp.ones(i0) y0 = jnp.ones(i0) return f(x0, y0) Note that new dynamic dimensions cannot yet be created inside a loop. Only things that already have a dynamic dimension can have that dynamic dimension change. For example, this is **not** a viable ``for_loop``, as ``x`` is initialized with an array with a concrete size. Note that while this example does not currently error out, similar code will likely cause XLA lowering errors. .. code-block:: python def w(): @qml.for_loop(3) def f(i, x): return jax.numpy.append(x, i) return f(jnp.array([])) """ if stop is None: start, stop = 0, start if active_jit := active_compiler(): compilers = AvailableCompilers.names_entrypoints ops_loader = compilers[active_jit]["ops"].load() return ops_loader.for_loop(start, stop, step) # if there is no active compiler, simply interpret the for loop # via the Python interpreter. def _decorator(body_fn): """Transform that will call the input ``body_fn`` within a for loop defined by the closure variables start, stop, and step. Args: body_fn (Callable): The function called within the for loop. Note that the loop body function must always have the iteration index as its first argument, which can be used arbitrarily inside the loop body. As the value of the index across iterations is handled automatically by the provided loop bounds, it must not be returned from the function. Closure Variables: start (int): starting value of the iteration index stop (int): (exclusive) upper bound of the iteration index step (int): increment applied to the iteration index at the end of each iteration allow_array_resizing (Literal["auto", True, False]) Returns: Callable: a callable with the same signature as ``body_fn`` """ return ForLoopCallable( start, stop, step, body_fn, allow_array_resizing=allow_array_resizing ) return _decorator
@functools.lru_cache def _get_for_loop_qfunc_prim(): """Get the loop_for primitive for quantum functions.""" # pylint: disable=import-outside-toplevel from pennylane.capture.custom_primitives import QmlPrimitive for_loop_prim = QmlPrimitive("for_loop") for_loop_prim.multiple_results = True for_loop_prim.prim_type = "higher_order" register_custom_staging_rule(for_loop_prim, lambda params: params["jaxpr_body_fn"].outvars) # pylint: disable=too-many-arguments @for_loop_prim.def_impl def _(start, stop, step, *args, jaxpr_body_fn, consts_slice, args_slice, abstract_shapes_slice): consts = args[consts_slice] init_state = args[args_slice] abstract_shapes = args[abstract_shapes_slice] # in case start >= stop, return the initial state fn_res = init_state for i in range(start, stop, step): fn_res = capture.eval_jaxpr(jaxpr_body_fn, consts, *abstract_shapes, i, *fn_res) return fn_res # pylint: disable=unused-argument @for_loop_prim.def_abstract_eval def _(start, stop, step, *args, args_slice, abstract_shapes_slice, **_): return args[abstract_shapes_slice] + args[args_slice] return for_loop_prim class ForLoopCallable: # pylint:disable=too-few-public-methods, too-many-arguments """Base class to represent a for loop. This class when called with an initial state will execute the while loop via the Python interpreter. Args: start (int): starting value of the iteration index stop (int): (exclusive) upper bound of the iteration index step (int): increment applied to the iteration index at the end of each iteration body_fn (Callable): The function called within the for loop. Note that the loop body function must always have the iteration index as its first argument, which can be used arbitrarily inside the loop body. As the value of the index across iterations is handled automatically by the provided loop bounds, it must not be returned from the function. Keyword Args: allow_array_resizing (Literal["auto", True, False]): How to handle arrays with dynamic shapes that change between iterations """ def __init__( self, start, stop, step, body_fn, *, allow_array_resizing: Literal["auto", True, False] = "auto", ): self.start = start self.stop = stop self.step = step self.body_fn = body_fn self.allow_array_resizing = allow_array_resizing def _call_capture_disabled(self, *init_state): args = init_state fn_res = args if len(args) > 1 else args[0] if len(args) == 1 else None for i in range(self.start, self.stop, self.step): fn_res = self.body_fn(i, *args) args = fn_res if len(args) > 1 else (fn_res,) if len(args) == 1 else () return fn_res def _get_jaxpr(self, init_state, allow_array_resizing): import jax # pylint: disable=import-outside-toplevel # need in_tree to include index so flat_fn will repack args correctly flat_args, in_tree = jax.tree_util.tree_flatten((0, *init_state)) # slice out the index so shape_locations indexes from non-index args/ results flat_args = flat_args[1:] tmp_array_resizing = False if allow_array_resizing == "auto" else allow_array_resizing abstracted_axes, abstract_shapes, shape_locations = loop_determine_abstracted_axes( tuple(flat_args), allow_array_resizing=tmp_array_resizing ) flat_fn = FlatFn(self.body_fn, in_tree=in_tree) if abstracted_axes: new_body_fn = add_abstract_shapes(flat_fn, shape_locations) dummy_init_state = [get_dummy_arg(arg) for arg in flat_args] abstracted_axes = ({},) + abstracted_axes # add in loop index else: new_body_fn = flat_fn dummy_init_state = flat_args try: jaxpr_body_fn = jax.make_jaxpr(new_body_fn, abstracted_axes=abstracted_axes)( 0, *dummy_init_state ) except ValueError as e: handle_jaxpr_error(e, (self.body_fn,), self.allow_array_resizing, "for_loop") error_msg = validate_no_resizing_returns(jaxpr_body_fn.jaxpr, shape_locations, "for_loop") if error_msg: if allow_array_resizing == "auto": # didn't work, so try with array resizing. return self._get_jaxpr(init_state, allow_array_resizing=True) raise ValueError(error_msg) assert flat_fn.out_tree return jaxpr_body_fn, abstract_shapes, flat_args, flat_fn.out_tree def _call_capture_enabled(self, *init_state): import jax # pylint: disable=import-outside-toplevel jaxpr_body_fn, abstract_shapes, flat_args, out_tree = self._get_jaxpr( init_state, allow_array_resizing=self.allow_array_resizing ) for_loop_prim = _get_for_loop_qfunc_prim() consts_slice = slice(0, len(jaxpr_body_fn.consts)) abstract_shapes_slice = slice(consts_slice.stop, consts_slice.stop + len(abstract_shapes)) args_slice = slice(abstract_shapes_slice.stop, None) results = for_loop_prim.bind( self.start, self.stop, self.step, *jaxpr_body_fn.consts, *abstract_shapes, *flat_args, jaxpr_body_fn=jaxpr_body_fn.jaxpr, consts_slice=consts_slice, args_slice=args_slice, abstract_shapes_slice=abstract_shapes_slice, ) results = results[-out_tree.num_leaves :] return jax.tree_util.tree_unflatten(out_tree, results) def __call__(self, *init_state): if enabled(): return self._call_capture_enabled(*init_state) return self._call_capture_disabled(*init_state)