numba-cuda 0.23.0__cp313-cp313-win_amd64.whl → 0.24.0__cp313-cp313-win_amd64.whl

numba_cuda/VERSION

@@ -1 +1 @@
-0.23.0
+0.24.0

numba_cuda/numba/cuda/__init__.py

@@ -1,6 +1,8 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: BSD-2-Clause
 
+# delvewheel: patch
+
 import importlib
 from numba.cuda.core import config
 from .utils import _readenv
@@ -23,7 +25,8 @@ if not (
 ):
     raise ImportError(
         "NVIDIA CUDA Python bindings not found. Install the 'cuda' package "
-        "(e.g. pip install nvidia-cuda-python or numba-cuda[cuXY])."
+        '(e.g. pip install "cuda-bindings==XY.*" or "numba-cuda[cuXY]", '
+        "with XY=12 or XY=13)."
     )
 
 if config.ENABLE_CUDASIM:

numba_cuda/numba/cuda/_compat.py

@@ -0,0 +1,47 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: BSD-2-Clause
+from packaging import version
+from cuda import core
+
+
+CUDA_CORE_VERSION = version.parse(core.__version__)
+if CUDA_CORE_VERSION < version.parse("0.5.0"):
+    from cuda.core.experimental import (
+        Program,
+        ProgramOptions,
+        Linker,
+        LinkerOptions,
+        Stream,
+        Device,
+        launch,
+        ObjectCode,
+        LaunchConfig,
+    )
+    from cuda.core.experimental._utils.cuda_utils import CUDAError, NVRTCError
+else:
+    from cuda.core import (
+        Program,
+        ProgramOptions,
+        Linker,
+        LinkerOptions,
+        Stream,
+        Device,
+        launch,
+        ObjectCode,
+        LaunchConfig,
+    )
+    from cuda.core._utils.cuda_utils import CUDAError, NVRTCError
+
+__all__ = [
+    "Program",
+    "ProgramOptions",
+    "Linker",
+    "LinkerOptions",
+    "Stream",
+    "Device",
+    "launch",
+    "CUDAError",
+    "NVRTCError",
+    "ObjectCode",
+    "LaunchConfig",
+]

numba_cuda/numba/cuda/cext/_dispatcher.cpp

@@ -29,7 +29,7 @@
  *
  */
 
-#if (PY_MAJOR_VERSION >= 3) && ((PY_MINOR_VERSION == 12) || (PY_MINOR_VERSION == 13))
+#if (PY_MAJOR_VERSION >= 3) && ((PY_MINOR_VERSION == 12) || (PY_MINOR_VERSION == 13) || (PY_MINOR_VERSION == 14))
 
 #ifndef Py_BUILD_CORE
     #define Py_BUILD_CORE 1
@@ -1004,12 +1004,18 @@ static PyTypeObject DispatcherType = {
     0,                                           /* tp_version_tag */
     0,                                           /* tp_finalize */
     0,                                           /* tp_vectorcall */
-#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION == 12)
+#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION >= 12)
 /* This was introduced first in 3.12
  * https://github.com/python/cpython/issues/91051
  */
     0,                                           /* tp_watched */
 #endif
+#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION >= 13)
+/* This was introduced in 3.13
+ * https://github.com/python/cpython/pull/114900
+ */
+    0,                                           /* tp_versions_used */
+#endif
 
 /* WARNING: Do not remove this, only modify it! It is a version guard to
  * act as a reminder to update this struct on Python version update! */

numba_cuda/numba/cuda/cext/_hashtable.cpp

@@ -110,7 +110,12 @@ _Numba_hashtable_hash_int(const void *key)
 extern "C" Py_uhash_t
 _Numba_hashtable_hash_ptr(const void *key)
 {
+/* Use public API on Python 3.13+; _Py_HashPointer is deprecated on 3.14+ */
+#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION >= 13)
+    return (Py_uhash_t)Py_HashPointer((void *)key);
+#else
     return (Py_uhash_t)_Py_HashPointer((void *)key);
+#endif
 }
 
 extern "C" int

numba_cuda/numba/cuda/cext/_pymodule.h

@@ -33,6 +33,6 @@
         Py_DECREF(tmp); } while (0)
 
 
-#define NB_SUPPORTED_PYTHON_MINOR ((PY_MINOR_VERSION == 10) || (PY_MINOR_VERSION == 11) || (PY_MINOR_VERSION == 12) || (PY_MINOR_VERSION == 13))
+#define NB_SUPPORTED_PYTHON_MINOR ((PY_MINOR_VERSION == 10) || (PY_MINOR_VERSION == 11) || (PY_MINOR_VERSION == 12) || (PY_MINOR_VERSION == 13) || (PY_MINOR_VERSION == 14))
 
 #endif /* NUMBA_PY_MODULE_H_ */

numba_cuda/numba/cuda/cext/_typeof.cpp

@@ -17,11 +17,24 @@
     #include <numpy/npy_2_compat.h>
 #endif
 
-#if (PY_MAJOR_VERSION >= 3) && (PY_MINOR_VERSION == 13)
-    #ifndef Py_BUILD_CORE
-        #define Py_BUILD_CORE 1
-    #endif
-    #include "internal/pycore_setobject.h"  // _PySet_NextEntry()
+#ifndef Py_BUILD_CORE
+#define Py_BUILD_CORE 1
+#endif
+
+#if (PY_MAJOR_VERSION >= 3) && (PY_MINOR_VERSION >= 13)
+// required include from Python 3.13+
+#include "internal/pycore_setobject.h"
+#ifndef PySet_NextEntry
+#define PySet_NextEntry _PySet_NextEntryRef
+#endif
+#else
+#ifndef PySet_NextEntry
+#define PySet_NextEntry _PySet_NextEntry
+#endif
+#endif
+
+#ifdef Py_BUILD_CORE
+#undef Py_BUILD_CORE
 #endif
 
 
@@ -410,17 +423,52 @@ compute_fingerprint(string_writer_t *w, PyObject *val)
         Py_hash_t h;
         PyObject *item;
         Py_ssize_t pos = 0;
+        int rc;
+
+#if PY_MAJOR_VERSION >= 3 && PY_MINOR_VERSION >= 13
+        // needed when using _PySet_NextEntryRef
+        Py_BEGIN_CRITICAL_SECTION(val);
+#endif
         /* Only one item is considered, as in typeof.py */
-        if (!_PySet_NextEntry(val, &pos, &item, &h)) {
+        rc = PySet_NextEntry(val, &pos, &item, &h);
+
+#if PY_MAJOR_VERSION >= 3 && PY_MINOR_VERSION >= 13
+        // needed when using _PySet_NextEntryRef
+        Py_END_CRITICAL_SECTION();
+#endif
+
+        if (!rc) {
             /* Empty set */
             PyErr_SetString(PyExc_ValueError,
                             "cannot compute fingerprint of empty set");
             return -1;
         }
-        TRY(string_writer_put_char, w, OP_SET);
-        TRY(compute_fingerprint, w, item);
+
+        if (string_writer_put_char(w, OP_SET)) {
+            goto fingerprint_error;
+        }
+
+        if (compute_fingerprint(w, item)) {
+            goto fingerprint_error;
+        }
+
+        goto fingerprint_success;
+
+fingerprint_error:
+#if PY_MAJOR_VERSION >= 3 && PY_MINOR_VERSION >= 13
+        // extra ref if using python >= 3.13
+        Py_XDECREF(item);
+#endif
+        return -1;
+
+fingerprint_success:
+#if PY_MAJOR_VERSION >= 3 && PY_MINOR_VERSION >= 13
+        // extra ref if using python >= 3.13
+        Py_XDECREF(item);
+#endif
         return 0;
     }
+
     if (PyObject_CheckBuffer(val)) {
         Py_buffer buf;
         int flags = PyBUF_ND | PyBUF_STRIDES | PyBUF_FORMAT;

numba_cuda/numba/cuda/cext/mviewbuf.c

@@ -337,12 +337,18 @@ static PyTypeObject MemAllocType = {
     0,                                          /* tp_version_tag */
     0,                                          /* tp_finalize */
     0,                                          /* tp_vectorcall */
-#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION == 12)
+#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION >= 12)
 /* This was introduced first in 3.12
  * https://github.com/python/cpython/issues/91051
  */
     0,                                           /* tp_watched */
 #endif
+#if (PY_MAJOR_VERSION == 3) && (PY_MINOR_VERSION >= 13)
+/* This was introduced in 3.13
+ * https://github.com/python/cpython/pull/114900
+ */
+    0,                                           /* tp_versions_used */
+#endif
 
 /* WARNING: Do not remove this, only modify it! It is a version guard to
  * act as a reminder to update this struct on Python version update! */

numba_cuda/numba/cuda/cloudpickle/cloudpickle.py

@@ -422,10 +422,9 @@ HAVE_ARGUMENT = dis.HAVE_ARGUMENT
 EXTENDED_ARG = dis.EXTENDED_ARG
 
 
-_BUILTIN_TYPE_NAMES = {}
-for k, v in types.__dict__.items():
-    if type(v) is type:
-        _BUILTIN_TYPE_NAMES[v] = k
+_BUILTIN_TYPE_NAMES = {
+    v: k for k, v in types.__dict__.items() if type(v) is type
+}
 
 
 def _builtin_type(name):
@@ -463,7 +462,7 @@ def _extract_class_dict(cls):
             base_value = inherited_dict[name]
             if value is base_value:
                 to_remove.append(name)
-        except KeyError:
+        except KeyError:  # noqa: PERF203
             pass
     for name in to_remove:
         clsdict.pop(name)

numba_cuda/numba/cuda/codegen.py

@@ -209,7 +209,9 @@ class CUDACodeLibrary(serialize.ReduceMixin, CodeLibrary):
             return cc
 
         device = devices.get_context().device
-        return device.compute_capability
+        cc = device.compute_capability
+        cc = (cc[0], cc[1], "a" if cc >= (9, 0) else "")
+        return cc
 
     def get_asm_str(self, cc=None):
         return "\n".join(self.get_asm_strs(cc=cc))
@@ -347,7 +349,7 @@ class CUDACodeLibrary(serialize.ReduceMixin, CodeLibrary):
         cufunc = self._cufunc_cache.get(device.id, None)
         if cufunc:
             return cufunc
-        cubin = self.get_cubin(cc=device.compute_capability)
+        cubin = self.get_cubin()
         module = ctx.create_module_image(
             cubin, self._setup_functions, self._teardown_functions
         )

numba_cuda/numba/cuda/compiler.py

@@ -852,15 +852,15 @@ def kernel_fixup(kernel, debug):
     return_value = kernel.args[0]
 
     for block in kernel.blocks:
-        remove_list = []
-
         # Find all stores first
-        for inst in block.instructions:
+        remove_list = [
+            inst
+            for inst in block.instructions
             if (
                 isinstance(inst, ir.StoreInstr)
                 and inst.operands[1] == return_value
-            ):
-                remove_list.append(inst)
+            )
+        ]
 
         # Remove all stores
         for to_remove in remove_list:

numba_cuda/numba/cuda/core/annotations/pretty_annotate.py

@@ -273,7 +273,7 @@ class Annotate:
         ann = function.get_annotation_info(signature=signature)
         self.ann = ann
 
-        for k, v in ann.items():
+        for v in ann.values():
             res = hllines(reform_code(v), style)
             rest = htlines(reform_code(v), style)
             v["pygments_lines"] = [

numba_cuda/numba/cuda/core/base.py

@@ -63,11 +63,11 @@ class OverloadSelector(object):
         """
         Select all compatible signatures and their implementation.
         """
-        out = {}
-        for ver_sig, impl in self.versions:
-            if self._match_arglist(ver_sig, sig):
-                out[ver_sig] = impl
-        return out
+        return {
+            ver_sig: impl
+            for ver_sig, impl in self.versions
+            if self._match_arglist(ver_sig, sig)
+        }
 
     def _best_signature(self, candidates):
         """
@@ -1322,11 +1322,7 @@ class _wrap_missing_loc(object):
             # ignore attributes if not available (i.e fix py2.7)
             attrs = "__name__", "libs"
             for attr in attrs:
-                try:
-                    val = getattr(fn, attr)
-                except AttributeError:
-                    pass
-                else:
+                if (val := getattr(fn, attr, None)) is not None:
                     setattr(wrapper, attr, val)
 
             return wrapper

numba_cuda/numba/cuda/core/bytecode.py

@@ -14,7 +14,7 @@ from numba.cuda import utils
 from numba.cuda.utils import PYVERSION
 
 
-if PYVERSION in ((3, 12), (3, 13)):
+if PYVERSION in ((3, 12), (3, 13), (3, 14)):
     from opcode import _inline_cache_entries
 
     # Instruction/opcode length in bytes
@@ -112,7 +112,7 @@ class ByteCodeInst(object):
         # https://bugs.python.org/issue27129
         # https://github.com/python/cpython/pull/25069
         assert self.is_jump
-        if PYVERSION in ((3, 13),):
+        if PYVERSION in ((3, 13), (3, 14)):
             if self.opcode in (
                 dis.opmap[k]
                 for k in ["JUMP_BACKWARD", "JUMP_BACKWARD_NO_INTERRUPT"]
@@ -141,7 +141,7 @@ class ByteCodeInst(object):
         else:
             raise NotImplementedError(PYVERSION)
 
-        if PYVERSION in ((3, 10), (3, 11), (3, 12), (3, 13)):
+        if PYVERSION in ((3, 10), (3, 11), (3, 12), (3, 13), (3, 14)):
             if self.opcode in JREL_OPS:
                 return self.next + self.arg * 2
             else:
@@ -179,7 +179,7 @@ NO_ARG_LEN = 1
 OPCODE_NOP = dis.opname.index("NOP")
 
 
-if PYVERSION in ((3, 13),):
+if PYVERSION in ((3, 13), (3, 14)):
 
     def _unpack_opargs(code):
         buf = []
@@ -418,7 +418,7 @@ class _ByteCode(object):
 
 
 def _fix_LOAD_GLOBAL_arg(arg):
-    if PYVERSION in ((3, 11), (3, 12), (3, 13)):
+    if PYVERSION in ((3, 11), (3, 12), (3, 13), (3, 14)):
         return arg >> 1
     elif PYVERSION in (
         (3, 9),
@@ -452,12 +452,13 @@ class ByteCodePy311(_ByteCode):
         """
         Returns the exception entry for the given instruction offset
         """
-        candidates = []
-        for ent in self.exception_entries:
-            if ent.start <= offset < ent.end:
-                candidates.append((ent.depth, ent))
+        candidates = [
+            (ent.depth, ent)
+            for ent in self.exception_entries
+            if ent.start <= offset < ent.end
+        ]
         if candidates:
-            ent = max(candidates)[1]
+            _, ent = max(candidates)
             return ent
 
 
@@ -592,7 +593,7 @@ class ByteCodePy312(ByteCodePy311):
                 if not next_inst.opname == "FOR_ITER":
                     continue
 
-                if PYVERSION in ((3, 13),):
+                if PYVERSION in ((3, 13), (3, 14)):
                     # Check end of pattern, two instructions.
                     # Check for the corresponding END_FOR, exception table end
                     # is non-inclusive, so subtract one.
@@ -601,8 +602,14 @@ class ByteCodePy312(ByteCodePy311):
                     if not curr_inst.opname == "END_FOR":
                         continue
                     next_inst = self.table[self.ordered_offsets[index - 1]]
-                    if not next_inst.opname == "POP_TOP":
-                        continue
+                    if PYVERSION in ((3, 13),):
+                        if not next_inst.opname == "POP_TOP":
+                            continue
+                    elif PYVERSION in ((3, 14),):
+                        if not next_inst.opname == "POP_ITER":
+                            continue
+                    else:
+                        raise NotImplementedError(PYVERSION)
                     # END_FOR must be followed by SWAP(2)
                     next_inst = self.table[self.ordered_offsets[index]]
                     if not next_inst.opname == "SWAP" and next_inst.arg == 2:
@@ -646,6 +653,7 @@ if PYVERSION == (3, 11):
 elif PYVERSION in (
     (3, 12),
     (3, 13),
+    (3, 14),
 ):
     ByteCode = ByteCodePy312
 elif PYVERSION < (3, 11):