[Program Capture] Add pytree support to captured qml.grad and qml.jacobian

doc/releases/changelog-dev.md

@@ -26,8 +26,10 @@
   into plxpr. When evaluating a captured `qml.grad` (`qml.jacobian`) instruction, it will
   dispatch to `jax.grad` (`jax.jacobian`), which differs from the Autograd implementation
   without capture.
+  Pytree inputs and outputs are supported.
   [(#6120)](https://github.com/PennyLaneAI/pennylane/pull/6120)
   [(#6127)](https://github.com/PennyLaneAI/pennylane/pull/6127)
+  [(#6134)](https://github.com/PennyLaneAI/pennylane/pull/6134)
 
 * Improve unit testing for capturing of nested control flows.
   [(#6111)](https://github.com/PennyLaneAI/pennylane/pull/6111)

pennylane/_grad.py

@@ -25,6 +25,7 @@
 
 from pennylane.capture import enabled
 from pennylane.capture.capture_diff import _get_grad_prim, _get_jacobian_prim
+from pennylane.capture.flatfn import FlatFn
 from pennylane.compiler import compiler
 from pennylane.compiler.compiler import CompileError
 
@@ -33,7 +34,9 @@
 
 def _capture_diff(func, argnum=None, diff_prim=None, method=None, h=None):
     """Capture-compatible gradient computation."""
-    import jax  # pylint: disable=import-outside-toplevel
+    # pylint: disable=import-outside-toplevel
+    import jax
+    from jax.tree_util import tree_flatten, tree_unflatten, treedef_tuple
 
     if isinstance(argnum, int):
         argnum = [argnum]
@@ -42,9 +45,40 @@ def _capture_diff(func, argnum=None, diff_prim=None, method=None, h=None):
 
     @wraps(func)
     def new_func(*args, **kwargs):
-        jaxpr = jax.make_jaxpr(partial(func, **kwargs))(*args)
-        prim_kwargs = {"argnum": argnum, "jaxpr": jaxpr.jaxpr, "n_consts": len(jaxpr.consts)}
-        return diff_prim.bind(*jaxpr.consts, *args, **prim_kwargs, method=method, h=h)
+        flat_args, in_trees = zip(*(tree_flatten(arg) for arg in args))
+        full_in_tree = treedef_tuple(in_trees)
+
+        # Create a new input tree that only takes inputs marked by argnum into account
+        trainable_in_trees = (in_tree for i, in_tree in enumerate(in_trees) if i in argnum)
+        trainable_in_tree = treedef_tuple(trainable_in_trees)
+
+        # Create argnum for the flat list of input arrays. For each flattened argument,
+        # add a list of flat argnums if the argument is trainable and an empty list otherwise.
+        start = 0
+        flat_argnum = sum(
+            (
+                (
+                    list(range(start, (start := start + len(flat_arg))))
+                    if i in argnum
+                    else list(range((start := start + len(flat_arg)), start))
+                )
+                for i, flat_arg in enumerate(flat_args)
+            ),
+            start=[],
+        )
+
+        # Create fully flattened function (flat inputs & outputs)
+        flat_fn = FlatFn(partial(func, **kwargs) if kwargs else func, full_in_tree)
+        flat_args = sum(flat_args, start=[])
+        jaxpr = jax.make_jaxpr(flat_fn)(*flat_args)
+        prim_kwargs = {"argnum": flat_argnum, "jaxpr": jaxpr.jaxpr, "n_consts": len(jaxpr.consts)}
+        out_flat = diff_prim.bind(*jaxpr.consts, *flat_args, **prim_kwargs, method=method, h=h)
+        # flatten once more to go from 2D derivative structure (outputs, args) to flat structure
+        out_flat, _ = tree_flatten(out_flat)
+        assert flat_fn.out_tree is not None, "out_tree should be set after executing flat_fn"
+        # The derivative output tree is the composition of output tree and trainable input trees
+        combined_tree = flat_fn.out_tree.compose(trainable_in_tree)
+        return tree_unflatten(combined_tree, out_flat)
 
     return new_func
 

pennylane/capture/capture_diff.py

@@ -103,7 +103,7 @@ def _(*args, argnum, jaxpr, n_consts, method, h):
         def func(*inner_args):
             return jax.core.eval_jaxpr(jaxpr, consts, *inner_args)
 
-        return jax.jacobian(func, argnums=argnum)(*args)
+        return jax.tree_util.tree_flatten(jax.jacobian(func, argnums=argnum)(*args))[0]
 
     # pylint: disable=unused-argument
     @jacobian_prim.def_abstract_eval

pennylane/capture/explanations.md

@@ -159,12 +159,11 @@ You can also see the const variable `a` as argument `e:i32[]` to the inner neste
 ### Pytree handling
 
 Evaluating a jaxpr requires accepting and returning a flat list of tensor-like inputs and outputs.
-list of tensor-like outputs.  These long lists can be hard to manage and are very
-restrictive on the allowed functions, but we can take advantage of pytrees to allow handling
-arbitrary functions.
+These long lists can be hard to manage and are very restrictive on the allowed functions, but we
+can take advantage of pytrees to allow handling arbitrary functions.
 
 To start, we import the `FlatFn` helper. This class converts a function to one that caches
-the resulting result pytree into `flat_fn.out_tree` when executed. This can be used to repack the
+the result pytree into `flat_fn.out_tree` when executed. This can be used to repack the
 results into the correct shape. It also returns flattened results. This does not particularly
 matter for program capture, as we will only be producing jaxpr from the function, not calling
 it directly.

pennylane/capture/flatfn.py

@@ -29,23 +29,48 @@ class FlatFn:
     property, so that the results can be repacked later. It also returns flattened results
     instead of the original result object.
 
+    If an ``in_tree`` is provided, the function accepts flattened inputs instead of the
+    original inputs with tree structure given by ``in_tree``.
+
+    **Example**
+
+    >>> import jax
+    >>> from pennylane.capture.flatfn import FlatFn
     >>> def f(x):
     ...     return {"y": 2+x["x"]}
     >>> flat_f = FlatFn(f)
-    >>> res = flat_f({"x": 0})
+    >>> arg = {"x": 0.5}
+    >>> res = flat_f(arg)
+    >>> res
+    [2.5]
+    >>> jax.tree_util.tree_unflatten(flat_f.out_tree, res)
+    {'y': 2.5}
+
+    If we want to use a fully flattened function that also takes flat inputs instead of
+    the original inputs with tree structure, we can provide the treedef for this input
+    structure:
+
+    >>> flat_args, in_tree = jax.tree_util.tree_flatten((arg,))
+    >>> flat_f = FlatFn(f, in_tree)
+    >>> res = flat_f(*flat_args)
     >>> res
-    [2]
+    [2.5]
     >>> jax.tree_util.tree_unflatten(flat_f.out_tree, res)
     {'y': 2.5}
 
+    Note that the ``in_tree`` has to be  created by flattening a tuple of all input
+    arguments, even if there is only a single argument.
     """
 
-    def __init__(self, f):
+    def __init__(self, f, in_tree=None):
         self.f = f
+        self.in_tree = in_tree
         self.out_tree = None
         update_wrapper(self, f)
 
     def __call__(self, *args):
+        if self.in_tree is not None:
+            args = jax.tree_util.tree_unflatten(self.in_tree, args)
         out = self.f(*args)
         out_flat, out_tree = jax.tree_util.tree_flatten(out)
         self.out_tree = out_tree

tests/capture/test_capture_diff.py

@@ -260,6 +260,55 @@ def circuit(x):
 
         jax.config.update("jax_enable_x64", initial_mode)
 
+    @pytest.mark.parametrize("argnum", ([0, 1], [0], [1]))
+    def test_grad_pytree_input(self, x64_mode, argnum):
+        """Test that the qml.grad primitive can be captured with pytree inputs."""
+
+        initial_mode = jax.config.jax_enable_x64
+        jax.config.update("jax_enable_x64", x64_mode)
+        fdtype = jax.numpy.float64 if x64_mode else jax.numpy.float32
+
+        def inner_func(x, y):
+            return jnp.prod(jnp.sin(x["a"]) * jnp.cos(y[0]["b"][1]) ** 2)
+
+        def func_qml(x):
+            return qml.grad(inner_func, argnum=argnum)(
+                {"a": x}, ({"b": [None, 0.4 * jnp.sqrt(x)]},)
+            )
+
+        def func_jax(x):
+            return jax.grad(inner_func, argnums=argnum)(
+                {"a": x}, ({"b": [None, 0.4 * jnp.sqrt(x)]},)
+            )
+
+        x = 0.7
+        jax_out = func_jax(x)
+        jax_out_flat, jax_out_tree = jax.tree_util.tree_flatten(jax_out)
+        qml_out_flat, qml_out_tree = jax.tree_util.tree_flatten(func_qml(x))
+        assert jax_out_tree == qml_out_tree
+        assert qml.math.allclose(jax_out_flat, qml_out_flat)
+
+        # Check overall jaxpr properties
+        if isinstance(argnum, int):
+            argnum = [argnum]
+        jaxpr = jax.make_jaxpr(func_qml)(x)
+        assert jaxpr.in_avals == [jax.core.ShapedArray((), fdtype, weak_type=True)]
+        assert len(jaxpr.eqns) == 3
+        assert jaxpr.out_avals == [jax.core.ShapedArray((), fdtype, weak_type=True)] * len(argnum)
+
+        grad_eqn = jaxpr.eqns[2]
+        diff_eqn_assertions(grad_eqn, grad_prim, argnum=argnum)
+        assert [var.aval for var in grad_eqn.outvars] == jaxpr.out_avals
+        assert len(grad_eqn.params["jaxpr"].eqns) == 6  # 5 numeric eqns, 1 conversion eqn
+
+        manual_out = jax.core.eval_jaxpr(jaxpr.jaxpr, jaxpr.consts, x)
+        manual_out_flat, manual_out_tree = jax.tree_util.tree_flatten(manual_out)
+        # Assert that the output from the manual evaluation is flat
+        assert manual_out_tree == jax.tree_util.tree_flatten(manual_out_flat)[1]
+        assert qml.math.allclose(jax_out_flat, manual_out_flat)
+
+        jax.config.update("jax_enable_x64", initial_mode)
+
 
 def _jac_allclose(jac1, jac2, num_axes, atol=1e-8):
     """Test that two Jacobians, given as nested sequences of arrays, are equal."""
@@ -279,10 +328,6 @@ class TestJacobian:
     @pytest.mark.parametrize("argnum", ([0, 1], [0], [1], 0, 1))
     def test_classical_jacobian(self, x64_mode, argnum):
         """Test that the qml.jacobian primitive can be captured with classical nodes."""
-        if isinstance(argnum, list) and len(argnum) > 1:
-            # These cases will only be unlocked with Pytree support
-            pytest.xfail()
-
         initial_mode = jax.config.jax_enable_x64
         jax.config.update("jax_enable_x64", x64_mode)
         fdtype = jnp.float64 if x64_mode else jnp.float32
@@ -307,14 +352,14 @@ def inner_func(x, y):
         func_jax = jax.jacobian(inner_func, argnums=argnum)
 
         jax_out = func_jax(x, y)
-        num_axes = 1 if isinstance(argnum, int) else 2
+        num_axes = 1 if (int_argnum := isinstance(argnum, int)) else 2
         assert _jac_allclose(func_qml(x, y), jax_out, num_axes)
 
         # Check overall jaxpr properties
         jaxpr = jax.make_jaxpr(func_jax)(x, y)
         jaxpr = jax.make_jaxpr(func_qml)(x, y)
 
-        if isinstance(argnum, int):
+        if int_argnum:
             argnum = [argnum]
 
         exp_in_avals = [shaped_array(shape) for shape in [(4,), (2, 3)]]
@@ -331,6 +376,9 @@ def inner_func(x, y):
         diff_eqn_assertions(jac_eqn, jacobian_prim, argnum=argnum)
 
         manual_eval = jax.core.eval_jaxpr(jaxpr.jaxpr, jaxpr.consts, x, y)
+        # Evaluating jaxpr gives flat list results. Need to adapt the JAX output to that
+        if not int_argnum:
+            jax_out = sum(jax_out, start=())
         assert _jac_allclose(manual_eval, jax_out, num_axes)
 
         jax.config.update("jax_enable_x64", initial_mode)
@@ -473,3 +521,63 @@ def circuit(x):
         assert _jac_allclose(manual_res, expected_res, 1)
 
         jax.config.update("jax_enable_x64", initial_mode)
+
+    @pytest.mark.parametrize("argnum", ([0, 1], [0], [1]))
+    def test_jacobian_pytrees(self, x64_mode, argnum):
+        """Test that the qml.jacobian primitive can be captured with
+        pytree inputs and outputs."""
+
+        initial_mode = jax.config.jax_enable_x64
+        jax.config.update("jax_enable_x64", x64_mode)
+        fdtype = jax.numpy.float64 if x64_mode else jax.numpy.float32
+
+        def inner_func(x, y):
+            return {
+                "prod_cos": jnp.prod(jnp.sin(x["a"]) * jnp.cos(y[0]["b"][1]) ** 2),
+                "sum_sin": jnp.sum(jnp.sin(x["a"]) * jnp.sin(y[1]["c"]) ** 2),
+            }
+
+        def func_qml(x):
+            return qml.jacobian(inner_func, argnum=argnum)(
+                {"a": x}, ({"b": [None, 0.4 * jnp.sqrt(x)]}, {"c": 0.5})
+            )
+
+        def func_jax(x):
+            return jax.jacobian(inner_func, argnums=argnum)(
+                {"a": x}, ({"b": [None, 0.4 * jnp.sqrt(x)]}, {"c": 0.5})
+            )
+
+        x = 0.7
+        jax_out = func_jax(x)
+        jax_out_flat, jax_out_tree = jax.tree_util.tree_flatten(jax_out)
+        qml_out_flat, qml_out_tree = jax.tree_util.tree_flatten(func_qml(x))
+        assert jax_out_tree == qml_out_tree
+        assert qml.math.allclose(jax_out_flat, qml_out_flat)
+
+        # Check overall jaxpr properties
+        if isinstance(argnum, int):
+            argnum = [argnum]
+        jaxpr = jax.make_jaxpr(func_qml)(x)
+        assert jaxpr.in_avals == [jax.core.ShapedArray((), fdtype, weak_type=True)]
+        assert len(jaxpr.eqns) == 3
+
+        # Compute the flat argnum in order to determine the expected number of out tracers
+        flat_argnum = []
+        if 0 in argnum:
+            flat_argnum.append(0)
+        if 1 in argnum:
+            flat_argnum.extend([1, 2])
+        assert jaxpr.out_avals == [jax.core.ShapedArray((), fdtype)] * (2 * len(flat_argnum))
+
+        jac_eqn = jaxpr.eqns[2]
+
+        diff_eqn_assertions(jac_eqn, jacobian_prim, argnum=flat_argnum)
+        assert [var.aval for var in jac_eqn.outvars] == jaxpr.out_avals
+
+        manual_out = jax.core.eval_jaxpr(jaxpr.jaxpr, jaxpr.consts, x)
+        manual_out_flat, manual_out_tree = jax.tree_util.tree_flatten(manual_out)
+        # Assert that the output from the manual evaluation is flat
+        assert manual_out_tree == jax.tree_util.tree_flatten(manual_out_flat)[1]
+        assert qml.math.allclose(jax_out_flat, manual_out_flat)
+
+        jax.config.update("jax_enable_x64", initial_mode)