Lift Subtensor over Join

Author: ricardoV94

pytensor/tensor/rewriting/subtensor_lift.py

@@ -4,16 +4,18 @@
 from numpy.core.numeric import normalize_axis_index, normalize_axis_tuple
 
 from pytensor import Variable
-from pytensor.graph import Constant, node_rewriter
-from pytensor.graph.rewriting.basic import copy_stack_trace
+from pytensor.graph import Constant, FunctionGraph, node_rewriter
+from pytensor.graph.rewriting.basic import NodeRewriter, copy_stack_trace
 from pytensor.scalar import basic as ps
 from pytensor.tensor.basic import (
     Alloc,
+    Join,
     MakeVector,
     alloc,
     as_tensor,
     expand_dims,
     get_underlying_scalar_constant_value,
+    join,
     register_infer_shape,
 )
 from pytensor.tensor.elemwise import CAReduce, DimShuffle, Elemwise
@@ -66,6 +68,40 @@ def _axis_is_indexed_by_basic_index(
     return any(ax < len(idxs) and not is_full_slice(idxs[ax]) for ax in axis)
 
 
+def _lift_subtensor_non_axis(
+    local_subtensor_lift_rewrite: NodeRewriter,
+    fgraph: FunctionGraph,
+    variable: Variable,
+    idx_tuple: tuple[int | slice],
+    axis: int,
+    old_subtensor_variable: Variable,
+) -> None | list[Variable]:
+    # Apply generic subtensor lift rewrite along "non-axis" dimensions
+    real_indices = [idx for idx in idx_tuple if not is_full_slice(idx)]
+    if len(real_indices) > 1 and variable.type.ndim > 1:
+        # Split the subtensor
+        idx_to_keep = idx_tuple[axis]
+        idxs_to_lift = (*idx_tuple[:axis], slice(None), *idx_tuple[axis + 1 :])
+
+        # Lift the non-axis indexes by calling the rewrite itself
+        indexed_variable = variable[idxs_to_lift]
+        [indexed_variable] = local_subtensor_lift_rewrite.transform(
+            fgraph, indexed_variable.owner
+        )
+        copy_stack_trace([old_subtensor_variable, indexed_variable], indexed_variable)
+
+        # Then reintroduce the axis index
+        ndim_reduced_left = _ndim_dropped_left_of_axis_by_basic_index(idx_tuple, axis)
+        new_axis = axis - ndim_reduced_left
+        idxs_to_keep = (*(slice(None),) * new_axis, idx_to_keep)
+        new_out = indexed_variable[idxs_to_keep]
+        copy_stack_trace(old_subtensor_variable, new_out)
+        return [new_out]
+
+    else:
+        return None
+
+
 @register_canonicalize
 @register_stabilize
 @register_specialize
@@ -297,29 +333,14 @@ def local_subtensor_of_softmax(fgraph, node):
     if _axis_is_indexed_by_basic_index(idx_tuple, axis):
         # If there are more dimensions being indexed, we can split them
         # And lift the non-axis indexes while keeping the axis index
-        real_indices = [idx for idx in idx_tuple if not is_full_slice(idx)]
-        if len(real_indices) > 1 and sm.type.ndim > 1:
-            # Split the subtensor
-            idx_to_keep = idx_tuple[axis]
-            idxs_to_lift = (*idx_tuple[:axis], slice(None), *idx_tuple[axis + 1 :])
-
-            # Lift the non-axis indexes by calling the rewrite itself
-            opt_sm = sm[idxs_to_lift]
-            [opt_sm] = local_subtensor_of_softmax.transform(fgraph, opt_sm.owner)
-            copy_stack_trace([old_out, sm], opt_sm)
-
-            # Then reintroduce the axis index
-            ndim_reduced_left = _ndim_dropped_left_of_axis_by_basic_index(
-                idx_tuple, axis
-            )
-            new_axis = axis - ndim_reduced_left
-            idxs_to_keep = (*(slice(None),) * new_axis, idx_to_keep)
-            new_out = opt_sm[idxs_to_keep]
-            copy_stack_trace(old_out, new_out)
-            return [new_out]
-
-        else:
-            return None
+        return _lift_subtensor_non_axis(
+            local_subtensor_lift_rewrite=local_subtensor_of_softmax,
+            fgraph=fgraph,
+            variable=sm,
+            idx_tuple=idx_tuple,
+            axis=axis,
+            old_subtensor_variable=old_out,
+        )
 
     # Index input to softmax
     x_sub = x[idx_tuple]
@@ -690,6 +711,52 @@ def local_subtensor_make_vector(fgraph, node):
             pass
 
 
+@register_canonicalize
+@register_specialize
+@node_rewriter([Subtensor])
+def local_subtensor_of_join(fgraph, node):
+    """Lift a Subtensor through a Join.
+
+    join(axis=1, x, y)[0] -> join(axis=0, x[0], y[0])
+    join(axis=1, x, y)[:, 0, -1] -> join(axis=1, x[:, :, -1], y[:, :, -1])[:, 0]
+
+    """
+    join_var, *idx = node.inputs
+
+    if not (join_var.owner and isinstance(join_var.owner.op, Join)):
+        return None
+
+    if len(fgraph.clients[join_var]) > 1:
+        # Join involves a full_copy, so we don't want to do it twice
+        return None
+
+    join_axis, *join_components = join_var.owner.inputs
+
+    # Rewrite only works when the join axis is a constant along a non-indexed dimension
+    if not isinstance(join_axis, Constant):
+        return None
+
+    [old_out] = node.outputs
+    axis = normalize_axis_index(join_axis.data, join_components[0].type.ndim)
+    idx_tuple = indices_from_subtensor(idx, node.op.idx_list)
+    if _axis_is_indexed_by_basic_index(idx_tuple, axis):
+        return _lift_subtensor_non_axis(
+            local_subtensor_lift_rewrite=local_subtensor_of_join,
+            fgraph=fgraph,
+            variable=join_var,
+            idx_tuple=idx_tuple,
+            axis=axis,
+            old_subtensor_variable=old_out,
+        )
+
+    # Lift index to the Join components
+    indexed_components = [component[idx_tuple] for component in join_components]
+    new_axis = axis - _ndim_dropped_left_of_axis_by_basic_index(idx_tuple, axis)
+    out = join(new_axis, *indexed_components)
+
+    return [out]
+
+
 @register_specialize
 @register_canonicalize
 @node_rewriter([Subtensor])

tests/tensor/rewriting/test_subtensor_lift.py

@@ -42,7 +42,7 @@
     tensor3,
     vector,
 )
-from pytensor.tensor.basic import MakeVector, expand_dims, make_vector
+from pytensor.tensor.basic import MakeVector, concatenate, expand_dims, make_vector
 from pytensor.tensor.elemwise import DimShuffle, Elemwise
 from pytensor.tensor.math import sum as pt_sum
 from pytensor.tensor.rewriting.subtensor_lift import (
@@ -252,6 +252,9 @@ def test_local_subtensor_of_softmax(original_fn, expected_fn):
     )
 
 
+shared_axis = shared(1, "axis")
+
+
 def test_local_subtensor_of_unbroadcast():
     # Test that Subtensor(Unbroadcast(x)) gets optimized into
     # Unbroadcast(Subtensor(x)).
@@ -661,6 +664,57 @@ def test_empty_subtensor(self):
         assert local_subtensor_make_vector.transform(fgraph, node) == [v]
 
 
+@pytest.mark.parametrize(
+    "original_fn, expected_fn",
+    [
+        (
+            lambda x, y: concatenate([x, y], axis=1)[1],
+            lambda x, y: concatenate([x[1], y[1]], axis=0),
+        ),
+        (
+            lambda x, y: concatenate([x, y], axis=-1)[1:],
+            lambda x, y: concatenate([x[1:], y[1:]], axis=1),
+        ),
+        # Indexing on both axis of concatenation and somewhere else:
+        (
+            lambda x, y: concatenate([x, y], axis=1)[0, 1:],
+            lambda x, y: concatenate([x[0], y[0]], axis=0)[1:],
+        ),
+        # Not supported, indexing on axis of concatenation
+        (
+            lambda x, y: concatenate([x, y], axis=0)[0],
+            lambda x, y: concatenate([x, y], axis=0)[0],
+        ),
+        (
+            lambda x, y: concatenate([x, y], axis=1)[:, 1:],
+            lambda x, y: concatenate([x, y], axis=1)[:, 1:],
+        ),
+        # Not supported, axis of concatenation is dynamically determined
+        (
+            lambda x, y: concatenate([x, y], axis=shared_axis)[1],
+            lambda x, y: concatenate([x, y], axis=shared_axis)[1],
+        ),
+    ],
+)
+def test_local_subtensor_of_join(original_fn, expected_fn):
+    rng = np.random.default_rng(257)
+    x = pt.matrix("x", shape=(5, 3))
+    y = pt.matrix("y", shape=(5, 3))
+    x_test = rng.normal(size=x.type.shape)
+    y_test = rng.normal(size=y.type.shape)
+
+    out = original_fn(x, y)
+    expected_opt_out = expected_fn(x, y)
+    opt_out = rewrite_graph(out)
+    assert equal_computations([opt_out], [expected_opt_out]), debugprint(
+        [expected_opt_out, opt_out], print_type=True
+    )
+    np.testing.assert_allclose(
+        opt_out.eval({x: x_test, y: y_test}, mode=NO_OPTIMIZATION_MODE),
+        out.eval({x: x_test, y: y_test}, mode=NO_OPTIMIZATION_MODE),
+    )
+
+
 def test_local_subtensor_shape_constant():
     x = tensor(dtype=np.float64, shape=(1, None)).shape[0]
     (res,) = local_subtensor_shape_constant.transform(None, x.owner)