numba-cuda 0.19.1__py3-none-any.whl → 0.20.1__py3-none-any.whl

numba_cuda/numba/cuda/simulator/compiler.py

@@ -11,3 +11,7 @@ compile_for_current_device = None
 compile_ptx = None
 compile_ptx_for_current_device = None
 declare_device_function = None
+
+
+def run_frontend(func, inline_closures=False, emit_dels=False):
+    pass

numba_cuda/numba/cuda/simulator/cudadrv/devicearray.py

@@ -7,7 +7,7 @@ stubs to allow tests to import correctly.
 """
 
 from contextlib import contextmanager
-from numba.np.numpy_support import numpy_version
+from numba.cuda.np.numpy_support import numpy_version
 
 import numpy as np
 

numba_cuda/numba/cuda/simulator/kernelapi.py

@@ -13,7 +13,7 @@ import traceback
 from numba.core import types
 import numpy as np
 
-from numba.np import numpy_support
+from numba.cuda.np import numpy_support
 
 from .vector_types import vector_types
 

numba_cuda/numba/cuda/simulator/memory_management/nrt.py

@@ -1,9 +1,21 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: BSD-2-Clause
 
-from numba import config
+from numba.cuda import config
 
-rtsys = None
+
+class RTSys:
+    def __init__(self, *args, **kwargs):
+        pass
+
+    def memsys_enable_stats(self):
+        pass
+
+    def get_allocation_stats(self):
+        pass
+
+
+rtsys = RTSys()
 
 config.CUDA_NRT_STATS = False
 config.CUDA_ENABLE_NRT = False

numba_cuda/numba/cuda/target.py

@@ -6,27 +6,31 @@ from functools import cached_property
 import llvmlite.binding as ll
 from llvmlite import ir
 import warnings
-from numba.cuda import cgutils, itanium_mangler
-from numba.core import (
-    compiler,
-    config,
-    targetconfig,
-    types,
-    typing,
-)
+
+from numba.core import types
+
 from numba.core.compiler_lock import global_compiler_lock
 from numba.core.dispatcher import Dispatcher
 from numba.core.errors import NumbaWarning
-from numba.core.base import BaseContext
-from numba.cuda.core.callconv import BaseCallConv, MinimalCallConv
+from numba.cuda.core.base import BaseContext
 from numba.core.typing import cmathdecl
 from numba.core import datamodel
 
 from .cudadrv import nvvm
-from numba.cuda import codegen, ufuncs
+from numba.cuda import (
+    cgutils,
+    itanium_mangler,
+    compiler,
+    codegen,
+    ufuncs,
+    typing,
+)
 from numba.cuda.debuginfo import CUDADIBuilder
 from numba.cuda.flags import CUDAFlags
 from numba.cuda.models import cuda_data_manager
+from numba.cuda.core.callconv import BaseCallConv, MinimalCallConv
+from numba.cuda.core import config, targetconfig
+
 
 # -----------------------------------------------------------------------------
 # Typing
@@ -34,8 +38,15 @@ from numba.cuda.models import cuda_data_manager
 
 class CUDATypingContext(typing.BaseContext):
     def load_additional_registries(self):
-        from . import cudadecl, cudamath, fp16, libdevicedecl, vector_types
-        from numba.core.typing import enumdecl, cffi_utils
+        from . import (
+            cudadecl,
+            cudamath,
+            fp16,
+            bf16,
+            libdevicedecl,
+            vector_types,
+        )
+        from numba.cuda.typing import enumdecl, cffi_utils
 
         self.install_registry(cudadecl.registry)
         self.install_registry(cffi_utils.registry)
@@ -45,6 +56,7 @@ class CUDATypingContext(typing.BaseContext):
         self.install_registry(enumdecl.registry)
         self.install_registry(vector_types.typing_registry)
         self.install_registry(fp16.typing_registry)
+        self.install_registry(bf16.typing_registry)
 
     def resolve_value_type(self, val):
         # treat other dispatcher object as another device function
@@ -141,12 +153,14 @@ class CUDATargetContext(BaseContext):
         self._target_data = None
 
     def load_additional_registries(self):
-        # side effect of import needed for numba.cpython.*, the builtins
+        # side effect of import needed for numba.cpython.*, numba.cuda.cpython.*, the builtins
         # registry is updated at import time.
-        from numba.cpython import numbers, tupleobj, slicing  # noqa: F401
+        from numba.cpython import tupleobj, slicing  # noqa: F401
+        from numba.cuda.cpython import numbers  # noqa: F401
         from numba.cpython import rangeobj, iterators, enumimpl  # noqa: F401
         from numba.cpython import unicode, charseq  # noqa: F401
-        from numba.cpython import cmathimpl
+        from numba.cuda.cpython import cmathimpl, mathimpl
+        from numba.core import optional  # noqa: F401
         from numba.misc import cffiimpl
         from numba.np import arrayobj  # noqa: F401
         from numba.np import npdatetime  # noqa: F401
@@ -155,8 +169,9 @@ class CUDATargetContext(BaseContext):
             fp16,
             printimpl,
             libdeviceimpl,
-            mathimpl,
+            mathimpl as cuda_mathimpl,
             vector_types,
+            bf16,
         )
 
         # fix for #8940
@@ -168,8 +183,11 @@ class CUDATargetContext(BaseContext):
         self.install_registry(libdeviceimpl.registry)
         self.install_registry(cmathimpl.registry)
         self.install_registry(mathimpl.registry)
+        self.install_registry(numbers.registry)
+        self.install_registry(cuda_mathimpl.registry)
         self.install_registry(vector_types.impl_registry)
         self.install_registry(fp16.target_registry)
+        self.install_registry(bf16.target_registry)
 
     def codegen(self):
         return self._internal_codegen

numba_cuda/numba/cuda/testing.py

@@ -6,17 +6,18 @@ import platform
 import shutil
 import pytest
 from datetime import datetime
-from numba.core.utils import PYVERSION
-from numba.cuda.cuda_paths import get_conda_ctk
+from numba.cuda.utils import PYVERSION
+from numba.cuda.cuda_paths import get_conda_ctk_libdir
 from numba.cuda.cudadrv import driver, devices, libs
 from numba.cuda.dispatcher import CUDADispatcher
-from numba.core import config
+from numba.cuda import config
 from numba.cuda.tests.support import TestCase
 from pathlib import Path
 
 from typing import Iterable, Union
 from io import StringIO
 import unittest
+import numpy as np
 
 if PYVERSION >= (3, 10):
     from filecheck.matcher import Matcher
@@ -44,6 +45,8 @@ class CUDATestCase(TestCase):
     matches FileCheck checks, and is not specific to CUDADispatcher.
     """
 
+    FLOAT16_RTOL = np.finfo(np.float16).eps
+
     def setUp(self):
         self._low_occupancy_warnings = config.CUDA_LOW_OCCUPANCY_WARNINGS
         self._warn_on_implicit_copy = config.CUDA_WARN_ON_IMPLICIT_COPY
@@ -211,7 +214,7 @@ def skip_unless_cudasim(reason):
 
 def skip_unless_conda_cudatoolkit(reason):
     """Skip test if the CUDA toolkit was not installed by Conda"""
-    return unittest.skipUnless(get_conda_ctk() is not None, reason)
+    return unittest.skipUnless(get_conda_ctk_libdir() is not None, reason)
 
 
 def skip_if_external_memmgr(reason):
@@ -272,21 +275,6 @@ def skip_if_mvc_enabled(reason):
     )
 
 
-def skip_if_mvc_libraries_unavailable(fn):
-    libs_available = False
-    try:
-        import cubinlinker  # noqa: F401 # type: ignore
-        import ptxcompiler  # noqa: F401 # type: ignore
-
-        libs_available = True
-    except ImportError:
-        pass
-
-    return unittest.skipUnless(
-        libs_available, "Requires cubinlinker and ptxcompiler"
-    )(fn)
-
-
 def cc_X_or_above(major, minor):
     if not config.ENABLE_CUDASIM:
         cc = devices.get_context().device.compute_capability

numba_cuda/numba/cuda/tests/__init__.py

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: BSD-2-Clause
 
 from fnmatch import fnmatch
-from numba.testing import unittest
+import unittest
 from numba import cuda
 from os.path import dirname, isfile, join, normpath, relpath, splitext
 

numba_cuda/numba/cuda/tests/cloudpickle_main_class.py

@@ -0,0 +1,9 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: BSD-2-Clause
+
+# Expected to run this module as __main__
+
+
+# Cloudpickle will think this is a dynamic class when this module is __main__
+class Klass:
+    classvar = None

numba_cuda/numba/cuda/tests/core/test_serialize.py

@@ -12,9 +12,9 @@ from multiprocessing import get_context
 
 import numba
 from numba.core.errors import TypingError
-from numba.tests.support import TestCase
+from numba.cuda.tests.support import TestCase
 from numba.core.target_extension import resolve_dispatcher_from_str
-from numba.cloudpickle import dumps, loads
+from numba.cuda.cloudpickle import dumps, loads
 
 
 class TestDispatcherPickling(TestCase):
@@ -230,7 +230,7 @@ class TestDispatcherPickling(TestCase):
 class TestSerializationMisc(TestCase):
     def test_numba_unpickle(self):
         # Test that _numba_unpickle is memorizing its output
-        from numba.core.serialize import _numba_unpickle
+        from numba.cuda.serialize import _numba_unpickle
 
         random_obj = object()
         bytebuf = pickle.dumps(random_obj)
@@ -318,7 +318,7 @@ def issue_7356():
 def check_main_class_reset_on_unpickle():
     # Load module and get its global dictionary
     glbs = runpy.run_module(
-        "numba.tests.cloudpickle_main_class",
+        "numba.cuda.tests.cloudpickle_main_class",
         run_name="__main__",
     )
     # Get the Klass and check it is from __main__

numba_cuda/numba/cuda/tests/cudadrv/test_cuda_devicerecord.py

@@ -10,7 +10,7 @@ from numba.cuda.cudadrv.devicearray import (
 )
 from numba.cuda.testing import unittest, CUDATestCase
 from numba.cuda.testing import skip_on_cudasim
-from numba.np import numpy_support
+from numba.cuda.np import numpy_support
 from numba import cuda
 
 N_CHARS = 5

numba_cuda/numba/cuda/tests/cudadrv/test_cuda_libraries.py

@@ -3,7 +3,7 @@
 
 from numba.cuda.testing import unittest
 from numba.cuda.testing import skip_on_cudasim, skip_unless_conda_cudatoolkit
-from numba.misc.findlib import find_lib
+from numba.cuda.misc.findlib import find_lib
 
 
 @skip_on_cudasim("Library detection unsupported in the simulator")

numba_cuda/numba/cuda/tests/cudadrv/test_deallocations.py

@@ -13,7 +13,7 @@ from numba.cuda.testing import (
     CUDATestCase,
 )
 from numba.cuda.tests.support import captured_stderr
-from numba.core import config
+from numba.cuda.core import config
 
 
 @skip_on_cudasim("not supported on CUDASIM")

numba_cuda/numba/cuda/tests/cudadrv/test_detect.py

@@ -60,8 +60,8 @@ class TestCUDAFindLibs(CUDATestCase):
     @unittest.skipIf(not sys.platform.startswith("linux"), "linux only")
     def test_cuda_find_lib_errors(self):
         """
-        This tests that the find_libs works as expected in the case of an
-        environment variable being used to set the path.
+        This tests that driver discovery attempts to load from typical system
+        locations and fails gracefully if pointed at an invalid directory.
         """
         # one of these is likely to exist on linux, it's also unlikely that
         # someone has extracted the contents of libdevice into here!
@@ -76,8 +76,11 @@ class TestCUDAFindLibs(CUDATestCase):
         # This is the testing part, the test will only run if there's a valid
         # path in which to look
         if looking_for is not None:
+            # We no longer support NUMBA_CUDA_DRIVER. Still run the subprocess
+            # to ensure importing and a trivial kernel launch path work, but
+            # do not set any Numba-specific driver env vars.
             out, err = self.run_test_in_separate_process(
-                "NUMBA_CUDA_DRIVER", looking_for
+                "DUMMY_UNUSED", looking_for
             )
             self.assertTrue(out is not None)
             self.assertTrue(err is not None)

numba_cuda/numba/cuda/tests/cudadrv/test_emm_plugins.py

@@ -6,7 +6,7 @@ import numpy as np
 import weakref
 
 from numba import cuda
-from numba.core import config
+from numba.cuda.core import config
 from numba.cuda.testing import unittest, CUDATestCase, skip_on_cudasim
 from numba.cuda.tests.support import linux_only
 

numba_cuda/numba/cuda/tests/cudadrv/test_linker.py

@@ -3,9 +3,13 @@
 
 import numpy as np
 import warnings
-from numba import config
+from numba.cuda import config
 from numba.cuda.testing import unittest
-from numba.cuda.testing import skip_on_cudasim, skip_if_cuda_includes_missing
+from numba.cuda.testing import (
+    skip_on_cudasim,
+    skip_if_cuda_includes_missing,
+    skip_if_nvjitlink_missing,
+)
 from numba.cuda.testing import CUDATestCase, test_data_dir
 from numba.cuda.cudadrv.driver import CudaAPIError, _Linker, LinkerError
 from numba.cuda import require_context
@@ -329,6 +333,18 @@ class TestLinker(CUDATestCase):
         calc_size = np.dtype(np.float64).itemsize * LMEM_SIZE
         self.assertGreaterEqual(local_mem_size, calc_size)
 
+    @skip_if_nvjitlink_missing("nvJitLink not installed or new enough (>12.3)")
+    def test_link_for_different_cc(self):
+        linker = _Linker.new(cc=(7, 5), lto=True)
+        code = """
+__device__ int foo(int x) {
+    return x + 1;
+}
+"""
+        linker.add_cu(code, "foo")
+        ptx = linker.get_linked_ptx().decode()
+        assert "target sm_75" in ptx
+
 
 if __name__ == "__main__":
     unittest.main()

numba_cuda/numba/cuda/tests/cudadrv/test_module_callbacks.py

@@ -6,7 +6,8 @@ import threading
 
 import numpy as np
 
-from numba import cuda, config
+from numba import cuda
+from numba.cuda import config
 from numba.cuda.cudadrv.linkable_code import CUSource
 from numba.cuda.testing import (
     CUDATestCase,

numba_cuda/numba/cuda/tests/cudadrv/test_nvjitlink.py

@@ -8,7 +8,7 @@ from numba.cuda import get_current_device
 from numba.cuda.cudadrv.driver import _Linker, _have_nvjitlink
 
 from numba import cuda
-from numba import config
+from numba.cuda import config
 
 import os
 import io

numba_cuda/numba/cuda/tests/cudadrv/test_ptds.py

@@ -15,7 +15,7 @@ from numba.cuda.tests.support import linux_only
 
 def child_test():
     from numba import cuda, int32, void
-    from numba.core import config
+    from numba.cuda.core import config
     import io
     import numpy as np
     import threading

numba_cuda/numba/cuda/tests/cudapy/extensions_usecases.py

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: BSD-2-Clause
 
 from numba import types
-from numba.core import config
+from numba.cuda.core import config
 
 
 class TestStruct:

numba_cuda/numba/cuda/tests/cudapy/test_array.py

@@ -5,7 +5,8 @@ import numpy as np
 
 from numba.cuda.testing import unittest, CUDATestCase
 from numba.cuda.testing import skip_on_cudasim, skip_unless_cudasim
-from numba import config, cuda
+from numba import cuda
+from numba.cuda import config
 
 
 if config.ENABLE_CUDASIM:

numba_cuda/numba/cuda/tests/cudapy/test_atomics.py

@@ -6,7 +6,7 @@ from textwrap import dedent
 
 from numba import cuda, uint32, uint64, float32, float64
 from numba.cuda.testing import unittest, CUDATestCase, cc_X_or_above
-from numba.core import config
+from numba.cuda.core import config
 
 
 @cuda.jit(device=True)