Add scalar argtypes to udf cache keys (#13194)

Closes #13193

Authors:
  - https://github.com/brandon-b-miller

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - GALI PREM SAGAR (https://github.com/galipremsagar)

URL: #13194

python/cudf/cudf/core/udf/groupby_utils.py

@@ -150,7 +150,7 @@ def jit_groupby_apply(offsets, grouped_values, function, *args):
     ngroups = len(offsets) - 1
 
     cache_key = _generate_cache_key(
-        grouped_values, function, suffix="__GROUPBY_APPLY_UDF"
+        grouped_values, function, args, suffix="__GROUPBY_APPLY_UDF"
     )
 
     if cache_key not in precompiled:

python/cudf/cudf/core/udf/utils.py

@@ -24,6 +24,7 @@
     column_from_udf_string_array,
     column_to_string_view_array,
 )
+from cudf.api.types import is_scalar
 from cudf.core.column.column import as_column
 from cudf.core.dtypes import dtype
 from cudf.core.udf.masked_typing import MaskedType
@@ -245,20 +246,22 @@ def _mask_get(mask, pos):
     return (mask[pos // MASK_BITSIZE] >> (pos % MASK_BITSIZE)) & 1
 
 
-def _generate_cache_key(frame, func: Callable, suffix="__APPLY_UDF"):
+def _generate_cache_key(frame, func: Callable, args, suffix="__APPLY_UDF"):
     """Create a cache key that uniquely identifies a compilation.
 
     A new compilation is needed any time any of the following things change:
     - The UDF itself as defined in python by the user
     - The types of the columns utilized by the UDF
     - The existence of the input columns masks
     """
+    scalar_argtypes = tuple(typeof(arg) for arg in args)
     return (
         *cudautils.make_cache_key(
             func, tuple(_all_dtypes_from_frame(frame).values())
         ),
         *(col.mask is None for col in frame._data.values()),
         *frame._data.keys(),
+        scalar_argtypes,
         suffix,
     )
 
@@ -285,9 +288,11 @@ def _compile_or_get(frame, func, args, kernel_getter=None):
     we then obtain the return type from that separate compilation and
     use it to allocate an output column of the right dtype.
     """
+    if not all(is_scalar(arg) for arg in args):
+        raise TypeError("only scalar valued args are supported by apply")
 
     # check to see if we already compiled this function
-    cache_key = _generate_cache_key(frame, func)
+    cache_key = _generate_cache_key(frame, func, args)
     if precompiled.get(cache_key) is not None:
         kernel, masked_or_scalar = precompiled[cache_key]
         return kernel, masked_or_scalar

python/cudf/cudf/tests/test_udf_masked_ops.py

@@ -775,6 +775,19 @@ def f(x):
 
     assert precompiled.currsize == 1
 
+    # validate that changing the type of a scalar arg
+    # results in a miss
+    precompiled.clear()
+
+    def f(x, c):
+        return x + c
+
+    data.apply(f, args=(1,))
+    assert precompiled.currsize == 1
+
+    data.apply(f, args=(1.5,))
+    assert precompiled.currsize == 2
+
 
 @pytest.mark.parametrize(
     "data", [[1.0, 0.0, 1.5], [1, 0, 2], [True, False, True]]