Coerce dtype __props__ to string due to invalid hash of `np.dtype()` objects

pytensor/sparse/sandbox/sp2.py

@@ -96,7 +96,7 @@ class Binomial(Op):
 
     def __init__(self, format, dtype):
         self.format = format
-        self.dtype = dtype
+        self.dtype = np.dtype(dtype).name
 
     def make_node(self, n, p, shape):
         n = pt.as_tensor_variable(n)

pytensor/tensor/basic.py

@@ -1090,6 +1090,8 @@ class Tri(Op):
     def __init__(self, dtype=None):
         if dtype is None:
             dtype = config.floatX
+        else:
+            dtype = np.dtype(dtype).name
         self.dtype = dtype
 
     def make_node(self, N, M, k):
@@ -1368,6 +1370,8 @@ class Eye(Op):
     def __init__(self, dtype=None):
         if dtype is None:
             dtype = config.floatX
+        else:
+            dtype = np.dtype(dtype).name
         self.dtype = dtype
 
     def make_node(self, n, m, k):
@@ -3225,7 +3229,7 @@ class ARange(COp):
     __props__ = ("dtype",)
 
     def __init__(self, dtype):
-        self.dtype = dtype
+        self.dtype = np.dtype(dtype).name
 
     def make_node(self, start, stop, step):
         from math import ceil
@@ -3407,7 +3411,8 @@ def arange(start, stop=None, step=1, dtype=None):
                     # We use the same dtype as numpy instead of the result of
                     # the upcast.
                     dtype = str(numpy_dtype)
-
+    else:
+        dtype = np.dtype(dtype).name
     if dtype not in _arange:
         _arange[dtype] = ARange(dtype)
     return _arange[dtype](start, stop, step)

pytensor/tensor/elemwise.py

@@ -1234,8 +1234,8 @@ class CAReduce(COp):
             else:
                 self.axis = tuple(axis)
 
-        self.dtype = dtype
-        self.acc_dtype = acc_dtype
+        self.dtype = dtype if dtype is None else np.dtype(dtype).name
+        self.acc_dtype = acc_dtype if acc_dtype is None else np.dtype(acc_dtype).name
         self.upcast_discrete_output = upcast_discrete_output
 
     @property

pytensor/tensor/io.py

@@ -25,7 +25,7 @@ class LoadFromDisk(Op):
     __props__ = ("dtype", "shape", "mmap_mode")
 
     def __init__(self, dtype, shape, mmap_mode=None):
-        self.dtype = np.dtype(dtype)  # turn "float64" into np.float64
+        self.dtype = np.dtype(dtype).name
         self.shape = shape
         if mmap_mode not in (None, "c"):
             raise ValueError(

pytensor/tensor/random/op.py

@@ -112,6 +112,8 @@ class RandomVariable(Op):
             else:
                 self.signature = safe_signature(self.ndims_params, [self.ndim_supp])
 
+        if isinstance(dtype, np.dtype):
+            dtype = dtype.name
         self.dtype = dtype or getattr(self, "dtype", None)
 
         self.inplace = (

tests/tensor/test_basic.py

@@ -2869,6 +2869,18 @@ class TestARange:
         assert np.arange(1.3, 17.48, 2.67).shape == arange(1.3, 17.48, 2.67).type.shape
         assert np.arange(-64, 64).shape == arange(-64, 64).type.shape
 
+    def test_c_cache_bug(self):
+        # Regression test for bug caused by issues in hash of `np.dtype()` objects
+        # https://github.com/numpy/numpy/issues/17864
+        end = iscalar("end")
+        arange1 = ARange(np.dtype("float64"))(0, end, 1)
+        arange2 = ARange("float64")(0, end + 1, 1)
+        assert arange1.owner.op == arange2.owner.op
+        assert hash(arange1.owner.op) == hash(arange2.owner.op)
+        fn = function([end], [arange1, arange2])
+        res1, res2 = fn(10)
+        np.testing.assert_array_equal(res1, res2[:-1], strict=True)
+
 
 class TestNdGrid:
     def setup_method(self):