Systematic use of mockable numba_basic.numba_jit

Direct import is not properly mocked by tests when trying to run `compare_numba_and_py` with `eval_obj_mode=True`

Author: ricardoV94

pytensor/link/numba/dispatch/blockwise.py

@@ -4,7 +4,8 @@
 from numba.core.extending import overload
 from numba.np.unsafe.ndarray import to_fixed_tuple
 
-from pytensor.link.numba.dispatch.basic import numba_funcify, numba_njit
+from pytensor.link.numba.dispatch import basic as numba_basic
+from pytensor.link.numba.dispatch.basic import numba_funcify
 from pytensor.link.numba.dispatch.vectorize_codegen import (
     _jit_options,
     _vectorized,
@@ -56,7 +57,7 @@ def numba_funcify_Blockwise(op: BlockwiseWithCoreShape, node, **kwargs):
         src += f"to_fixed_tuple(core_shapes[{i}], {core_shapes_len[i]}),"
     src += ")"
 
-    to_tuple = numba_njit(
+    to_tuple = numba_basic.numba_njit(
         compile_function_src(
             src,
             "to_tuple",

pytensor/link/numba/dispatch/elemwise.py

@@ -359,13 +359,13 @@ def numba_funcify_Sum(op, node, **kwargs):
 
     if ndim_input == len(axes):
         # Slightly faster than `numba_funcify_CAReduce` for this case
-        @numba_njit
+        @numba_basic.numba_njit
         def impl_sum(array):
             return np.asarray(array.sum(), dtype=np_acc_dtype).astype(out_dtype)
 
     elif len(axes) == 0:
         # These cases should be removed by rewrites!
-        @numba_njit
+        @numba_basic.numba_njit
         def impl_sum(array):
             return np.asarray(array, dtype=out_dtype)
 
@@ -615,25 +615,25 @@ def numba_funcify_Dot(op, node, **kwargs):
 
     if x_dtype == dot_dtype and y_dtype == dot_dtype:
 
-        @numba_njit
+        @numba_basic.numba_njit
         def dot(x, y):
             return np.asarray(np.dot(x, y))
 
     elif x_dtype == dot_dtype and y_dtype != dot_dtype:
 
-        @numba_njit
+        @numba_basic.numba_njit
         def dot(x, y):
             return np.asarray(np.dot(x, y.astype(dot_dtype)))
 
     elif x_dtype != dot_dtype and y_dtype == dot_dtype:
 
-        @numba_njit
+        @numba_basic.numba_njit
         def dot(x, y):
             return np.asarray(np.dot(x.astype(dot_dtype), y))
 
     else:
 
-        @numba_njit()
+        @numba_basic.numba_njit
         def dot(x, y):
             return np.asarray(np.dot(x.astype(dot_dtype), y.astype(dot_dtype)))
 
@@ -642,7 +642,7 @@ def dot(x, y):
 
     else:
 
-        @numba_njit
+        @numba_basic.numba_njit
         def dot_with_cast(x, y):
             return dot(x, y).astype(out_dtype)
 
@@ -653,7 +653,7 @@ def dot_with_cast(x, y):
 def numba_funcify_BatchedDot(op, node, **kwargs):
     dtype = node.outputs[0].type.numpy_dtype
 
-    @numba_njit
+    @numba_basic.numba_njit
     def batched_dot(x, y):
         # Numba does not support 3D matmul
         # https://github.com/numba/numba/issues/3804

pytensor/link/numba/dispatch/linalg/decomposition/lu.py

@@ -2,24 +2,24 @@
 from typing import Literal
 
 import numpy as np
-from numba import njit as numba_njit
 from numba.core.extending import overload
 from numba.np.linalg import ensure_lapack
 from scipy import linalg
 
+from pytensor.link.numba.dispatch import basic as numba_basic
 from pytensor.link.numba.dispatch.linalg.decomposition.lu_factor import _getrf
 from pytensor.link.numba.dispatch.linalg.utils import _check_scipy_linalg_matrix
 
 
-@numba_njit
+@numba_basic.numba_njit
 def _pivot_to_permutation(p, dtype):
     p_inv = np.arange(len(p)).astype(dtype)
     for i in range(len(p)):
         p_inv[i], p_inv[p[i]] = p_inv[p[i]], p_inv[i]
     return p_inv
 
 
-@numba_njit
+@numba_basic.numba_njit
 def _lu_factor_to_lu(a, dtype, overwrite_a):
     A_copy, IPIV, _INFO = _getrf(a, overwrite_a=overwrite_a)
 

pytensor/link/numba/dispatch/linalg/solve/tridiagonal.py

@@ -6,8 +6,8 @@
 from numpy import ndarray
 from scipy import linalg
 
+from pytensor.link.numba.dispatch import basic as numba_basic
 from pytensor.link.numba.dispatch import numba_funcify
-from pytensor.link.numba.dispatch.basic import numba_njit
 from pytensor.link.numba.dispatch.linalg._LAPACK import (
     _LAPACK,
     _get_underlying_float,
@@ -27,7 +27,7 @@
 )
 
 
-@numba_njit
+@numba_basic.numba_njit
 def tridiagonal_norm(du, d, dl):
     # Adapted from scipy _matrix_norm_tridiagonal:
     # https://github.com/scipy/scipy/blob/0f1fd4a7268b813fa2b844ca6038e4dfdf90084a/scipy/linalg/_basic.py#L356-L367
@@ -346,7 +346,7 @@ def numba_funcify_LUFactorTridiagonal(op: LUFactorTridiagonal, node, **kwargs):
     overwrite_d = op.overwrite_d
     overwrite_du = op.overwrite_du
 
-    @numba_njit(cache=False)
+    @numba_basic.numba_njit(cache=False)
     def lu_factor_tridiagonal(dl, d, du):
         dl, d, du, du2, ipiv, _ = _gttrf(
             dl,
@@ -368,7 +368,7 @@ def numba_funcify_SolveLUFactorTridiagonal(
     overwrite_b = op.overwrite_b
     transposed = op.transposed
 
-    @numba_njit(cache=False)
+    @numba_basic.numba_njit(cache=False)
     def solve_lu_factor_tridiagonal(dl, d, du, du2, ipiv, b):
         x, _ = _gttrs(
             dl,

pytensor/link/numba/dispatch/nlinalg.py

@@ -30,14 +30,14 @@ def numba_funcify_SVD(op, node, **kwargs):
 
     if not compute_uv:
 
-        @numba_basic.numba_njit()
+        @numba_basic.numba_njit
         def svd(x):
             _, ret, _ = np.linalg.svd(inputs_cast(x), full_matrices)
             return ret
 
     else:
 
-        @numba_basic.numba_njit()
+        @numba_basic.numba_njit
         def svd(x):
             return np.linalg.svd(inputs_cast(x), full_matrices)
 

pytensor/link/numba/dispatch/random.py

@@ -91,7 +91,7 @@ def {name}(rng, {input_signature}):
 def numba_core_BernoulliRV(op, node):
     out_dtype = node.outputs[1].type.numpy_dtype
 
-    @numba_basic.numba_njit()
+    @numba_basic.numba_njit
     def random(rng, p):
         return (
             direct_cast(0, out_dtype)

pytensor/link/numba/dispatch/shape.py

@@ -3,6 +3,7 @@
 import numpy as np
 from numba.np.unsafe import ndarray as numba_ndarray
 
+from pytensor.link.numba.dispatch import basic as numba_basic
 from pytensor.link.numba.dispatch import numba_funcify
 from pytensor.link.numba.dispatch.basic import create_arg_string, numba_njit
 from pytensor.link.utils import compile_function_src
@@ -12,7 +13,7 @@
 
 @numba_funcify.register(Shape)
 def numba_funcify_Shape(op, **kwargs):
-    @numba_njit
+    @numba_basic.numba_njit
     def shape(x):
         return np.asarray(np.shape(x))
 
@@ -23,7 +24,7 @@ def shape(x):
 def numba_funcify_Shape_i(op, **kwargs):
     i = op.i
 
-    @numba_njit
+    @numba_basic.numba_njit
     def shape_i(x):
         return np.asarray(np.shape(x)[i])
 
@@ -61,13 +62,13 @@ def numba_funcify_Reshape(op, **kwargs):
 
     if ndim == 0:
 
-        @numba_njit
+        @numba_basic.numba_njit
         def reshape(x, shape):
             return np.asarray(x.item())
 
     else:
 
-        @numba_njit
+        @numba_basic.numba_njit
         def reshape(x, shape):
             # TODO: Use this until https://github.com/numba/numba/issues/7353 is closed.
             return np.reshape(

pytensor/link/numba/dispatch/signal/conv.py

@@ -1,8 +1,8 @@
 import numpy as np
 from numba.np.arraymath import _get_inner_prod
 
+from pytensor.link.numba.dispatch import basic as numba_basic
 from pytensor.link.numba.dispatch import numba_funcify
-from pytensor.link.numba.dispatch.basic import numba_njit
 from pytensor.tensor.signal.conv import Convolve1d
 
 
@@ -13,7 +13,7 @@ def numba_funcify_Convolve1d(op, node, **kwargs):
     out_dtype = node.outputs[0].type.dtype
     innerprod = _get_inner_prod(a_dtype, b_dtype)
 
-    @numba_njit
+    @numba_basic.numba_njit
     def valid_convolve1d(x, y):
         nx = len(x)
         ny = len(y)
@@ -30,7 +30,7 @@ def valid_convolve1d(x, y):
 
         return ret
 
-    @numba_njit
+    @numba_basic.numba_njit
     def full_convolve1d(x, y):
         nx = len(x)
         ny = len(y)
@@ -59,7 +59,7 @@ def full_convolve1d(x, y):
 
         return ret
 
-    @numba_njit
+    @numba_basic.numba_njit
     def convolve_1d(x, y, mode):
         if mode:
             return full_convolve1d(x, y)

pytensor/link/numba/dispatch/slinalg.py

@@ -3,7 +3,8 @@
 import numpy as np
 
 from pytensor import config
-from pytensor.link.numba.dispatch.basic import numba_funcify, numba_njit
+from pytensor.link.numba.dispatch import basic as numba_basic
+from pytensor.link.numba.dispatch.basic import numba_funcify
 from pytensor.link.numba.dispatch.linalg.decomposition.cholesky import _cholesky
 from pytensor.link.numba.dispatch.linalg.decomposition.lu import (
     _lu_1,
@@ -63,7 +64,7 @@ def numba_funcify_Cholesky(op, node, **kwargs):
     if dtype in complex_dtypes:
         raise NotImplementedError(_COMPLEX_DTYPE_NOT_SUPPORTED_MSG.format(op=op))
 
-    @numba_njit
+    @numba_basic.numba_njit
     def cholesky(a):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(a), np.isnan(a))):
@@ -95,7 +96,7 @@ def pivot_to_permutation(op, node, **kwargs):
     inverse = op.inverse
     dtype = node.outputs[0].dtype
 
-    @numba_njit
+    @numba_basic.numba_njit
     def numba_pivot_to_permutation(piv):
         p_inv = _pivot_to_permutation(piv, dtype)
 
@@ -118,7 +119,7 @@ def numba_funcify_LU(op, node, **kwargs):
     if dtype in complex_dtypes:
         NotImplementedError(_COMPLEX_DTYPE_NOT_SUPPORTED_MSG.format(op=op))
 
-    @numba_njit(inline="always")
+    @numba_basic.numba_njit(inline="always")
     def lu(a):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(a), np.isnan(a))):
@@ -165,7 +166,7 @@ def numba_funcify_LUFactor(op, node, **kwargs):
     if dtype in complex_dtypes:
         NotImplementedError(_COMPLEX_DTYPE_NOT_SUPPORTED_MSG.format(op=op))
 
-    @numba_njit
+    @numba_basic.numba_njit
     def lu_factor(a):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(a), np.isnan(a))):
@@ -185,7 +186,7 @@ def numba_funcify_BlockDiagonal(op, node, **kwargs):
     dtype = node.outputs[0].dtype
 
     # TODO: Why do we always inline all functions? It doesn't work with starred args, so can't use it in this case.
-    @numba_njit
+    @numba_basic.numba_njit
     def block_diag(*arrs):
         shapes = np.array([a.shape for a in arrs], dtype="int")
         out_shape = [int(s) for s in np.sum(shapes, axis=0)]
@@ -235,7 +236,7 @@ def numba_funcify_Solve(op, node, **kwargs):
         )
         solve_fn = _solve_gen
 
-    @numba_njit
+    @numba_basic.numba_njit
     def solve(a, b):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(a), np.isnan(a))):
@@ -267,7 +268,7 @@ def numba_funcify_SolveTriangular(op, node, **kwargs):
             _COMPLEX_DTYPE_NOT_SUPPORTED_MSG.format(op="Solve Triangular")
         )
 
-    @numba_njit
+    @numba_basic.numba_njit
     def solve_triangular(a, b):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(a), np.isnan(a))):
@@ -304,7 +305,7 @@ def numba_funcify_CholeskySolve(op, node, **kwargs):
     if dtype in complex_dtypes:
         raise NotImplementedError(_COMPLEX_DTYPE_NOT_SUPPORTED_MSG.format(op=op))
 
-    @numba_njit
+    @numba_basic.numba_njit
     def cho_solve(c, b):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(c), np.isnan(c))):
@@ -337,7 +338,7 @@ def numba_funcify_QR(op, node, **kwargs):
     integer_input = dtype in integer_dtypes
     in_dtype = config.floatX if integer_input else dtype
 
-    @numba_njit(cache=False)
+    @numba_basic.numba_njit(cache=False)
     def qr(a):
         if check_finite:
             if np.any(np.bitwise_or(np.isinf(a), np.isnan(a))):

pytensor/link/numba/dispatch/sort.py

@@ -2,8 +2,8 @@
 
 import numpy as np
 
+from pytensor.link.numba.dispatch import basic as numba_basic
 from pytensor.link.numba.dispatch import numba_funcify
-from pytensor.link.numba.dispatch.basic import numba_njit
 from pytensor.tensor.sort import ArgSortOp, SortOp
 
 
@@ -18,7 +18,7 @@ def numba_funcify_SortOp(op, node, **kwargs):
             UserWarning,
         )
 
-    @numba_njit
+    @numba_basic.numba_njit
     def sort_f(a, axis):
         axis = axis.item()
 
@@ -45,7 +45,7 @@ def numba_funcify_ArgSortOp(op, node, **kwargs):
             UserWarning,
         )
 
-    @numba_njit
+    @numba_basic.numba_njit
     def argort_f(X, axis):
         axis = axis.item()