Define all batched dot operations as matmul

New rewrite is added to convert unpaired batched row/column matvec or vec products as equivalent matmul products.

Author: ricardoV94

pytensor/tensor/math.py

@@ -3916,23 +3916,7 @@ def logsumexp(x, axis=None, keepdims=False):
     return log(sum(exp(x), axis=axis, keepdims=keepdims))
 
 
-# Predefine all batched variations of Dot
-_inner_prod = Blockwise(
-    _dot,
-    signature="(n),(n)->()",
-)
-
-_matrix_vec_prod = Blockwise(
-    _dot,
-    signature="(m,k),(k)->(m)",
-)
-
-_vec_matrix_prod = Blockwise(
-    _dot,
-    signature="(k),(k,n)->(n)",
-)
-
-_matrix_matrix_matmul = Blockwise(
+_matmul = Blockwise(
     _dot,
     signature="(m,k),(k,n)->(m,n)",
     gufunc_spec=("numpy.matmul", 2, 1),
@@ -3988,11 +3972,11 @@ def matmul(x1: "ArrayLike", x2: "ArrayLike", dtype: Optional["DTypeLike"] = None
     if x1.type.ndim == 1 and x2.type.ndim == 1:
         out = _dot(x1, x2)
     elif x1.type.ndim == 1:
-        out = _matrix_matrix_matmul(x1[None], x2).squeeze(-2)
+        out = vecmat(x1, x2)
     elif x2.type.ndim == 1:
-        out = _matrix_matrix_matmul(x1, x2[:, None]).squeeze(-1)
+        out = matvec(x1, x2)
     else:
-        out = _matrix_matrix_matmul(x1, x2)
+        out = _matmul(x1, x2)
 
     if dtype is not None:
         out = out.astype(dtype)
@@ -4042,7 +4026,7 @@ def vecdot(
     >>> z_batch = pt.vecdot(x_batch, y_batch)  # shape (3,)
     >>> # Equivalent to numpy.vecdot(x_batch, y_batch)
     """
-    out = _inner_prod(x1, x2)
+    out = matmul(x1[..., None, :], x2[..., :, None]).squeeze((-2, -1))
 
     if dtype is not None:
         out = out.astype(dtype)
@@ -4091,7 +4075,7 @@ def matvec(
     >>> result = pt.matvec(batched_A, batched_v)  # shape (2, 3)
     >>> # Equivalent to numpy.matvec(batched_A, batched_v)
     """
-    out = _matrix_vec_prod(x1, x2)
+    out = matmul(x1, x2[..., None]).squeeze(-1)
 
     if dtype is not None:
         out = out.astype(dtype)
@@ -4129,18 +4113,18 @@ def vecmat(
     --------
     >>> import pytensor.tensor as pt
     >>> # Vector-matrix product
-    >>> v = pt.vector("v", shape=(3,))  # shape (3,)
-    >>> A = pt.matrix("A", shape=(3, 4))  # shape (3, 4)
+    >>> v = pt.vector("v", shape=(3,))
+    >>> A = pt.matrix("A", shape=(3, 4))
     >>> result = pt.vecmat(v, A)  # shape (4,)
     >>> # Equivalent to numpy.vecmat(v, A)
     >>>
     >>> # Batched vector-matrix product
-    >>> batched_v = pt.matrix("v", shape=(2, 3))  # shape (2, 3)
-    >>> batched_A = pt.tensor3("A", shape=(2, 3, 4))  # shape (2, 3, 4)
+    >>> batched_v = pt.matrix("v", shape=(2, 3))
+    >>> batched_A = pt.tensor3("A", shape=(2, 3, 4))
     >>> result = pt.vecmat(batched_v, batched_A)  # shape (2, 4)
     >>> # Equivalent to numpy.vecmat(batched_v, batched_A)
     """
-    out = _vec_matrix_prod(x1, x2)
+    out = matmul(x2.mT, x1[..., None]).squeeze(-1)
 
     if dtype is not None:
         out = out.astype(dtype)
@@ -4155,18 +4139,18 @@ def vectorize_node_dot(op, node, batched_x, batched_y):
     old_y_ndim = old_y.type.ndim
     match (old_x_ndim, old_y_ndim):
         case (1, 1):
-            batch_op = _inner_prod
+            batch_fn = vecdot
         case (2, 1):
-            batch_op = _matrix_vec_prod
+            batch_fn = matvec
         case (1, 2):
-            batch_op = _vec_matrix_prod
+            batch_fn = vecmat
         case (2, 2):
-            batch_op = _matrix_matrix_matmul
+            batch_fn = matmul
         case _:
             raise ValueError(
                 f"Core dot Op should have 1D or 2D inputs, got {old_x_ndim}D and {old_y_ndim}D."
             )
-    return batch_op(batched_x, batched_y).owner
+    return batch_fn(batched_x, batched_y).owner
 
 
 def nan_to_num(x, nan=0.0, posinf=None, neginf=None):

pytensor/tensor/rewriting/blas.py

@@ -98,7 +98,7 @@
 from pytensor.tensor.exceptions import NotScalarConstantError
 from pytensor.tensor.math import (
     Dot,
-    _matrix_matrix_matmul,
+    _matmul,
     add,
     mul,
     neg,
@@ -908,7 +908,7 @@ def local_dot22_to_dot22scalar(fgraph, node):
 
 
 @register_specialize
-@node_rewriter([_matrix_matrix_matmul])
+@node_rewriter([_matmul])
 def specialize_matmul_to_batched_dot(fgraph, node):
     """Rewrite Matmul (Blockwise matrix-matrix) without implicit broadcasted batched dimension as BatchedDot.
 

pytensor/tensor/rewriting/elemwise.py

@@ -37,6 +37,7 @@
     broadcasted_by,
     register_canonicalize,
     register_specialize,
+    register_stabilize,
 )
 from pytensor.tensor.variable import TensorConstant
 
@@ -390,6 +391,7 @@ def is_dimshuffle_useless(new_order, input):
 
 
 @register_canonicalize
+@register_stabilize
 @register_specialize
 @node_rewriter([DimShuffle])
 def local_dimshuffle_lift(fgraph, node):

pytensor/tensor/rewriting/linalg.py

@@ -26,7 +26,7 @@
 from pytensor.tensor.blas import Dot22
 from pytensor.tensor.blockwise import Blockwise
 from pytensor.tensor.elemwise import DimShuffle, Elemwise
-from pytensor.tensor.math import Dot, Prod, _matrix_matrix_matmul, log, outer, prod
+from pytensor.tensor.math import Dot, Prod, _matmul, log, outer, prod
 from pytensor.tensor.nlinalg import (
     SVD,
     KroneckerProduct,
@@ -282,7 +282,7 @@ def cholesky_ldotlt(fgraph, node):
                 # This rewrite only applies to matrix Dot
                 and A.owner.inputs[0].type.ndim == 2
             )
-            or (A.owner.op == _matrix_matrix_matmul)
+            or (A.owner.op == _matmul)
         )
     ):
         return

pytensor/tensor/rewriting/math.py

@@ -28,14 +28,14 @@
     as_tensor_variable,
     cast,
     constant,
+    expand_dims,
     get_underlying_scalar_constant_value,
     moveaxis,
     ones_like,
     register_infer_shape,
     switch,
     zeros_like,
 )
-from pytensor.tensor.blockwise import Blockwise
 from pytensor.tensor.elemwise import CAReduce, DimShuffle, Elemwise
 from pytensor.tensor.exceptions import NotScalarConstantError
 from pytensor.tensor.extra_ops import broadcast_arrays
@@ -45,10 +45,7 @@
     Sum,
     _conj,
     _dot,
-    _inner_prod,
-    _matrix_matrix_matmul,
-    _matrix_vec_prod,
-    _vec_matrix_prod,
+    _matmul,
     add,
     digamma,
     dot,
@@ -196,60 +193,134 @@ def local_lift_transpose_through_dot(fgraph, node):
         return ret
 
 
-@register_stabilize
+@register_canonicalize
 @register_specialize
-@node_rewriter(tracks=[Blockwise])
+@node_rewriter(tracks=[_matmul])
 def local_batched_matmul_to_core_matmul(fgraph, node):
-    """Rewrite matmul where only one of the inputs has batch dimensions to a reshaped core matmul.
+    """Move batch dimensions of matmul operands to core matmul
 
-    Example, if x has batch dimensions, but y not:
+    Example, if x has batch dimensions that don't overlap with batch dimensions of y
     x @ y -> (x.reshape(-1, x.shape[-1]) @ y).reshape(*x.shape[:-1], y.shape[-1])
 
-    It also works when y has batch dimensions, but x not.
+    It also works for batch dimensions of y that don't overlap with batch dimensions of x
     """
 
-    # Check whether we have a matmul operation in this node
-    if not (
-        isinstance(node.op.core_op, Dot)
-        and len(node.op.inputs_sig[0]) == 2
-        and len(node.op.inputs_sig[1]) == 2
-    ):
-        return None
-
     x, y = node.inputs
     batch_ndim = node.op.batch_ndim(node)
 
-    # Check if x has batch dimensions, but y not (or only broadcastable dimensions)
-    if any(not b_dim for b_dim in x.type.broadcastable[:-2]) and all(
-        y.type.broadcastable[:-2]
-    ):
-        x_stacked = x.reshape((-1, x.shape[-1]))
-        out_stacked = x_stacked @ y.squeeze(tuple(range(batch_ndim)))
-        out = out_stacked.reshape((*x.shape[:-1], y.shape[-1]))
-        return [out]
-
-    # Otherwise, check if y has batch dimension, but x not
-    elif any(not b_dim for b_dim in y.type.broadcastable[:-2]) and all(
-        x.type.broadcastable[:-2]
-    ):
-        # For the y batch case we need to first move the batch axes and then reshape
-        # y.shape == (*b, k, n)
-        y_tr = moveaxis(y, -2, 0)  # (k, *b, n)
-        y_stacked = y_tr.reshape((y.shape[-2], -1))  # (k, *b * n)
-        out_stacked = x.squeeze(tuple(range(batch_ndim))) @ y_stacked  # (m, *b * n)
-        out_stacked_tr = out_stacked.reshape(
-            (x.shape[-2], *y.shape[:-2], y.shape[-1])
-        )  # (m, *b, n)
-        out = moveaxis(out_stacked_tr, 0, -2)  # (*b, m, n)
-        return [out]
-
-    # Both x and y have batch dimensions, nothing to do here
-    return None
+    x_axis_to_merge = [
+        i
+        for i, (bcast_x, bcast_y) in enumerate(
+            zip(x.type.broadcastable[:-2], y.type.broadcastable[:-2])
+        )
+        if bcast_y and not bcast_x
+    ]
+
+    y_axis_to_merge = [
+        i
+        for i, (bcast_x, bcast_y) in enumerate(
+            zip(x.type.broadcastable[:-2], y.type.broadcastable[:-2])
+        )
+        if bcast_x and not bcast_y
+    ]
+
+    if not (x_axis_to_merge or y_axis_to_merge):
+        return None
+
+    x_shape = tuple(x.shape)
+    y_shape = tuple(y.shape)
+    x_is_row = x.type.broadcastable[-2]
+    y_is_col = y.type.broadcastable[-1]
+    n_x_axis_to_merge = len(x_axis_to_merge)
+    n_y_axis_to_merge = len(y_axis_to_merge)
+    n_axis_to_merge = n_x_axis_to_merge + n_y_axis_to_merge
+
+    x_stacked, y_stacked = x, y
+    dims_were_merged = False
+
+    if n_x_axis_to_merge:
+        # ravel batch dimensions of x on the core (m) axis
+        x_axis_destination = tuple(range(-n_x_axis_to_merge - 2, -2))
+        x_stacked = moveaxis(x, x_axis_to_merge, x_axis_destination)
+        if x_is_row:
+            # x was a row matrix, squeeze it to clean up the graph
+            x_stacked = x_stacked.squeeze(-2)
+        if n_x_axis_to_merge > 1 or not x_is_row:
+            # Ravel moved batch dims together with (m) if needed
+            x_stacked_shape = tuple(x_stacked.shape)
+            x_stacked = x_stacked.reshape(
+                (*x_stacked_shape[: batch_ndim - n_x_axis_to_merge], -1, x_shape[-1])
+            )
+            dims_were_merged = True
+
+    if n_y_axis_to_merge:
+        # ravel batch dimensions of y on the core (n) axis
+        y_axis_destination = tuple(range(-n_y_axis_to_merge - 1, -1))
+        y_stacked = moveaxis(y, y_axis_to_merge, y_axis_destination)
+        if y_is_col:
+            # y was a column matrix, squeeze it to clean up the graph
+            y_stacked = y_stacked.squeeze(-1)
+        if n_y_axis_to_merge > 1 or not y_is_col:
+            # Ravel moved batch dims together with (n) if needed
+            y_stacked_shape = tuple(y_stacked.shape)
+            y_stacked = y_stacked.reshape(
+                (*y_stacked_shape[: batch_ndim - n_y_axis_to_merge], y_shape[-2], -1)
+            )
+            dims_were_merged = True
+
+    # Squeeze x_dims corresponding to merged dimensions of y
+    x_axis_to_squeeze = np.array(y_axis_to_merge)
+    for i in reversed(x_axis_to_merge):
+        # The corresponding dimensions of y may have shifted when we merged dimensions of x
+        x_axis_to_squeeze[x_axis_to_squeeze > i] -= 1
+    x_stacked = x_stacked.squeeze(tuple(x_axis_to_squeeze))
+
+    # Same for y
+    y_axis_to_squeeze = np.array(x_axis_to_merge)
+    for i in reversed(y_axis_to_merge):
+        y_axis_to_squeeze[y_axis_to_squeeze > i] -= 1
+    y_stacked = y_stacked.squeeze(tuple(y_axis_to_squeeze))
+
+    out_stacked = x_stacked @ y_stacked
+
+    # Split back any merged dimensions
+    if dims_were_merged:
+        x_merged_shapes = [x_shape[i] for i in x_axis_to_merge]
+        if not x_is_row:
+            # Otherwise we handle that later with expand_dims, which is cleaner
+            x_merged_shapes.append(x_shape[-2])
+        y_merged_shapes = [y_shape[i] for i in y_axis_to_merge]
+        if not y_is_col:
+            # Otherwise we handle that later with expand_dims, which is cleaner
+            y_merged_shapes.append(y_shape[-1])
+        out_stacked_shape = tuple(out_stacked.shape)
+        out_unstacked = out_stacked.reshape(
+            (
+                *out_stacked_shape[: batch_ndim - n_axis_to_merge],
+                *x_merged_shapes,
+                *y_merged_shapes,
+            )
+        )
+    else:
+        out_unstacked = out_stacked
+
+    # Add back dummy row, col axis
+    # We do this separately to avoid the reshape as much as we can
+    if y_is_col and (n_y_axis_to_merge or dims_were_merged):
+        out_unstacked = expand_dims(out_unstacked, -1)
+    if x_is_row and (n_x_axis_to_merge or dims_were_merged):
+        out_unstacked = expand_dims(out_unstacked, -n_y_axis_to_merge - 2)
+
+    # Move batch axis back to their original location
+    source = range(-n_axis_to_merge - 2, 0)
+    destination = (*x_axis_to_merge, -2, *y_axis_to_merge, -1)
+    out = moveaxis(out_unstacked, source, destination)
+    return [out]
 
 
 @register_canonicalize
 @register_specialize
-@node_rewriter([_inner_prod, _matrix_vec_prod, _vec_matrix_prod, _matrix_matrix_matmul])
+@node_rewriter([_matmul])
 def local_blockwise_dot_to_mul(fgraph, node):
     """Rewrite blockwise dots that correspond to multiplication without summation.