pytensor 2.28.2__tar.gz → 2.28.3__tar.gz

{pytensor-2.28.2/pytensor.egg-info → pytensor-2.28.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: pytensor
-Version: 2.28.2
+Version: 2.28.3
 Summary: Optimizing compiler for evaluating mathematical expressions on CPUs and GPUs.
 Author-email: pymc-devs <pymc.devs@gmail.com>
 License: .. _license:

{pytensor-2.28.2 → pytensor-2.28.3}/pytensor/_version.py

@@ -8,11 +8,11 @@ import json
 
 version_json = '''
 {
- "date": "2025-03-03T18:16:56+0100",
+ "date": "2025-03-05T13:02:56+0100",
  "dirty": false,
  "error": null,
- "full-revisionid": "757a10cd496ac24800f03491891ff6bba2d4807a",
- "version": "2.28.2"
+ "full-revisionid": "89d5366c888f430003aa6f16a13ad9cdd2826d95",
+ "version": "2.28.3"
 }
 '''  # END VERSION_JSON
 

{pytensor-2.28.2 → pytensor-2.28.3}/pytensor/link/jax/dispatch/slinalg.py

@@ -53,7 +53,6 @@ def jax_funcify_Solve(op, **kwargs):
 @jax_funcify.register(SolveTriangular)
 def jax_funcify_SolveTriangular(op, **kwargs):
     lower = op.lower
-    trans = op.trans
     unit_diagonal = op.unit_diagonal
     check_finite = op.check_finite
 
@@ -62,7 +61,7 @@ def jax_funcify_SolveTriangular(op, **kwargs):
             A,
             b,
             lower=lower,
-            trans=trans,
+            trans=0,  # this is handled by explicitly transposing A, so it will always be 0 when we get to here.
             unit_diagonal=unit_diagonal,
             check_finite=check_finite,
         )

{pytensor-2.28.2 → pytensor-2.28.3}/pytensor/link/numba/dispatch/slinalg.py

@@ -180,7 +180,6 @@ def solve_triangular_impl(A, B, trans, lower, unit_diagonal, b_ndim, overwrite_b
 
 @numba_funcify.register(SolveTriangular)
 def numba_funcify_SolveTriangular(op, node, **kwargs):
-    trans = bool(op.trans)
     lower = op.lower
     unit_diagonal = op.unit_diagonal
     check_finite = op.check_finite
@@ -208,7 +207,7 @@ def numba_funcify_SolveTriangular(op, node, **kwargs):
         res = _solve_triangular(
             a,
             b,
-            trans=trans,
+            trans=0,  # transposing is handled explicitly on the graph, so we never use this argument
             lower=lower,
             unit_diagonal=unit_diagonal,
             overwrite_b=overwrite_b,

{pytensor-2.28.2 → pytensor-2.28.3}/pytensor/tensor/c_code/dimshuffle.c

@@ -33,12 +33,17 @@ int APPLY_SPECIFIC(cpu_dimshuffle)(PyArrayObject *input, PyArrayObject **res, PA
     npy_intp original_size = PyArray_SIZE(input);
     npy_intp new_size = 1;
     for (npy_intp i = 0; i < nd_out; ++i) {
+        // We set the strides of length 1 dimensions to PyArray_ITEMSIZE(input).
+        // The value is arbitrary, because there is never a next element.
+        // np.expand_dims(x, 0) and x[None] do different things here.
+        // I would prefer zero, but there are some poorly implemented BLAS operations
+        // That don't handle zero strides correctly. At least they won't fail because of DimShuffle.
         if (new_order[i] != -1) {
             dimensions[i] = PyArray_DIMS(input)[new_order[i]];
-            strides[i] = PyArray_DIMS(input)[new_order[i]] == 1 ? 0 : PyArray_STRIDES(input)[new_order[i]];
+            strides[i] = PyArray_DIMS(input)[new_order[i]] == 1 ? PyArray_ITEMSIZE(input) : PyArray_STRIDES(input)[new_order[i]];
         } else {
             dimensions[i] = 1;
-            strides[i] = 0;
+            strides[i] = PyArray_ITEMSIZE(input);
         }
         new_size *= dimensions[i];
     }

{pytensor-2.28.2 → pytensor-2.28.3}/pytensor/tensor/slinalg.py

@@ -296,13 +296,12 @@ class SolveBase(Op):
         # We need to return (dC/d[inv(A)], dC/db)
         c_bar = output_gradients[0]
 
-        trans_solve_op = type(self)(
-            **{
-                k: (not getattr(self, k) if k == "lower" else getattr(self, k))
-                for k in self.__props__
-            }
-        )
-        b_bar = trans_solve_op(A.T, c_bar)
+        props_dict = self._props_dict()
+        props_dict["lower"] = not self.lower
+
+        solve_op = type(self)(**props_dict)
+
+        b_bar = solve_op(A.T, c_bar)
         # force outer product if vector second input
         A_bar = -ptm.outer(b_bar, c) if c.ndim == 1 else -b_bar.dot(c.T)
 
@@ -385,7 +384,6 @@ class SolveTriangular(SolveBase):
     """Solve a system of linear equations."""
 
     __props__ = (
-        "trans",
         "unit_diagonal",
         "lower",
         "check_finite",
@@ -393,11 +391,10 @@ class SolveTriangular(SolveBase):
         "overwrite_b",
     )
 
-    def __init__(self, *, trans=0, unit_diagonal=False, **kwargs):
+    def __init__(self, *, unit_diagonal=False, **kwargs):
         if kwargs.get("overwrite_a", False):
             raise ValueError("overwrite_a is not supported for SolverTriangulare")
         super().__init__(**kwargs)
-        self.trans = trans
         self.unit_diagonal = unit_diagonal
 
     def perform(self, node, inputs, outputs):
@@ -406,7 +403,7 @@ class SolveTriangular(SolveBase):
             A,
             b,
             lower=self.lower,
-            trans=self.trans,
+            trans=0,
             unit_diagonal=self.unit_diagonal,
             check_finite=self.check_finite,
             overwrite_b=self.overwrite_b,
@@ -445,9 +442,9 @@ def solve_triangular(
 
     Parameters
     ----------
-    a
+    a: TensorVariable
         Square input data
-    b
+    b: TensorVariable
         Input data for the right hand side.
     lower : bool, optional
         Use only data contained in the lower triangle of `a`. Default is to use upper triangle.
@@ -468,10 +465,17 @@ def solve_triangular(
         This will influence how batched dimensions are interpreted.
     """
     b_ndim = _default_b_ndim(b, b_ndim)
+
+    if trans in [1, "T", True]:
+        a = a.mT
+        lower = not lower
+    if trans in [2, "C"]:
+        a = a.conj().mT
+        lower = not lower
+
     ret = Blockwise(
         SolveTriangular(
             lower=lower,
-            trans=trans,
             unit_diagonal=unit_diagonal,
             check_finite=check_finite,
             b_ndim=b_ndim,
@@ -534,6 +538,7 @@ def solve(
     *,
     assume_a="gen",
     lower=False,
+    transposed=False,
     check_finite=True,
     b_ndim: int | None = None,
 ):
@@ -564,8 +569,10 @@ def solve(
     b : (..., N, NRHS) array_like
         Input data for the right hand side.
     lower : bool, optional
-        If True, only the data contained in the lower triangle of `a`. Default
+        If True, use only the data contained in the lower triangle of `a`. Default
         is to use upper triangle. (ignored for ``'gen'``)
+    transposed: bool, optional
+        If True, solves the system A^T x = b. Default is False.
     check_finite : bool, optional
         Whether to check that the input matrices contain only finite numbers.
         Disabling may give a performance gain, but may result in problems
@@ -577,6 +584,11 @@ def solve(
         This will influence how batched dimensions are interpreted.
     """
     b_ndim = _default_b_ndim(b, b_ndim)
+
+    if transposed:
+        a = a.mT
+        lower = not lower
+
     return Blockwise(
         Solve(
             lower=lower,

{pytensor-2.28.2 → pytensor-2.28.3/pytensor.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: pytensor
-Version: 2.28.2
+Version: 2.28.3
 Summary: Optimizing compiler for evaluating mathematical expressions on CPUs and GPUs.
 Author-email: pymc-devs <pymc.devs@gmail.com>
 License: .. _license: