Backport #1243

.github/workflows/build-wheel.yml

@@ -25,7 +25,6 @@ jobs:
       fail-fast: false
       matrix:
         python-version:
-          - "3.9"
           - "3.10"
           - "3.11"
           - "3.12"

ci/test-matrix.json

@@ -4,8 +4,6 @@
   "_notes": "DRIVER: 'earliest' does not work with CUDA 12.9.1 and LOCAL_CTK: 0 does not work with CUDA 12.0.1",
   "linux": {
     "pull-request": [
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "latest" },
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "13.0.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.10", "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.10", "CUDA_VER": "13.0.1", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.11", "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "latest" },
@@ -17,8 +15,6 @@
       { "ARCH": "amd64", "PY_VER": "3.13t", "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.14", "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.14t", "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "latest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "latest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "13.0.1", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },
       { "ARCH": "arm64", "PY_VER": "3.10", "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },
       { "ARCH": "arm64", "PY_VER": "3.10", "CUDA_VER": "13.0.1", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "latest" },
       { "ARCH": "arm64", "PY_VER": "3.11", "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "latest" },
@@ -32,11 +28,6 @@
       { "ARCH": "arm64", "PY_VER": "3.14t", "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "latest" }
     ],
     "nightly": [
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "11.8.0", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "earliest" },
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "11.8.0", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "12.0.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "latest" },
-      { "ARCH": "amd64", "PY_VER": "3.9",  "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.10", "CUDA_VER": "11.8.0", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "earliest" },
       { "ARCH": "amd64", "PY_VER": "3.10", "CUDA_VER": "11.8.0", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.10", "CUDA_VER": "12.0.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
@@ -57,11 +48,6 @@
       { "ARCH": "amd64", "PY_VER": "3.13", "CUDA_VER": "12.0.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.13", "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "l4", "DRIVER": "latest" },
       { "ARCH": "amd64", "PY_VER": "3.13", "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "l4", "DRIVER": "latest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "11.8.0", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "earliest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "11.8.0", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "12.0.1", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "12.9.1", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "latest" },
-      { "ARCH": "arm64", "PY_VER": "3.9",  "CUDA_VER": "12.9.1", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },
       { "ARCH": "arm64", "PY_VER": "3.10", "CUDA_VER": "11.8.0", "LOCAL_CTK": "0", "GPU": "a100", "DRIVER": "earliest" },
       { "ARCH": "arm64", "PY_VER": "3.10", "CUDA_VER": "11.8.0", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },
       { "ARCH": "arm64", "PY_VER": "3.10", "CUDA_VER": "12.0.1", "LOCAL_CTK": "1", "GPU": "a100", "DRIVER": "latest" },

cuda_bindings/cuda/bindings/_internal/nvjitlink.pxd

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from ..cynvjitlink cimport *
 

cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from libc.stdint cimport intptr_t, uintptr_t
 
@@ -16,6 +16,8 @@ from cuda.pathfinder import load_nvidia_dynamic_lib
 # Extern
 ###############################################################################
 
+# You must 'from .utils import NotSupportedError' before using this template
+
 cdef extern from "<dlfcn.h>" nogil:
     void* dlopen(const char*, int)
     char* dlerror()
@@ -30,14 +32,32 @@ cdef extern from "<dlfcn.h>" nogil:
 
     const void* RTLD_DEFAULT 'RTLD_DEFAULT'
 
+cdef int get_cuda_version():
+    cdef void* handle = NULL
+    cdef int err, driver_ver = 0
+
+    # Load driver to check version
+    handle = dlopen('libcuda.so.1', RTLD_NOW | RTLD_GLOBAL)
+    if handle == NULL:
+        err_msg = dlerror()
+        raise NotSupportedError(f'CUDA driver is not found ({err_msg.decode()})')
+    cuDriverGetVersion = dlsym(handle, "cuDriverGetVersion")
+    if cuDriverGetVersion == NULL:
+        raise RuntimeError('Did not find cuDriverGetVersion symbol in libcuda.so.1')
+    err = (<int (*)(int*) noexcept nogil>cuDriverGetVersion)(&driver_ver)
+    if err != 0:
+        raise RuntimeError(f'cuDriverGetVersion returned error code {err}')
+
+    return driver_ver
+
+
 
 ###############################################################################
 # Wrapper init
 ###############################################################################
 
 cdef object __symbol_lock = threading.Lock()
 cdef bint __py_nvjitlink_init = False
-cdef void* __cuDriverGetVersion = NULL
 
 cdef void* __nvJitLinkCreate = NULL
 cdef void* __nvJitLinkDestroy = NULL
@@ -60,12 +80,16 @@ cdef void* load_library() except* with gil:
     return <void*>handle
 
 
-cdef int __check_or_init_nvjitlink() except -1 nogil:
+cdef int _init_nvjitlink() except -1 nogil:
     global __py_nvjitlink_init
 
     cdef void* handle = NULL
 
     with gil, __symbol_lock:
+        # Recheck the flag after obtaining the locks
+        if __py_nvjitlink_init:
+            return 0
+
         # Load function
         global __nvJitLinkCreate
         __nvJitLinkCreate = dlsym(RTLD_DEFAULT, 'nvJitLinkCreate')
@@ -173,7 +197,7 @@ cdef inline int _check_or_init_nvjitlink() except -1 nogil:
     if __py_nvjitlink_init:
         return 0
 
-    return __check_or_init_nvjitlink()
+    return _init_nvjitlink()
 
 cdef dict func_ptrs = None
 

cuda_bindings/cuda/bindings/_internal/nvjitlink_windows.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from libc.stdint cimport intptr_t
 
@@ -15,7 +15,7 @@ from libc.stddef cimport wchar_t
 from libc.stdint cimport uintptr_t
 from cpython cimport PyUnicode_AsWideCharString, PyMem_Free
 
-from .utils import NotSupportedError
+# You must 'from .utils import NotSupportedError' before using this template
 
 cdef extern from "windows.h" nogil:
     ctypedef void* HMODULE
@@ -61,10 +61,10 @@ cdef int get_cuda_version():
         raise NotSupportedError('CUDA driver is not found')
     cuDriverGetVersion = GetProcAddress(handle, 'cuDriverGetVersion')
     if cuDriverGetVersion == NULL:
-        raise RuntimeError('something went wrong')
+        raise RuntimeError('Did not find cuDriverGetVersion symbol in nvcuda.dll')
     err = (<int (*)(int*) noexcept nogil>cuDriverGetVersion)(&driver_ver)
     if err != 0:
-        raise RuntimeError('something went wrong')
+        raise RuntimeError(f'cuDriverGetVersion returned error code {err}')
 
     return driver_ver
 
@@ -93,10 +93,14 @@ cdef void* __nvJitLinkGetInfoLog = NULL
 cdef void* __nvJitLinkVersion = NULL
 
 
-cdef int __check_or_init_nvjitlink() except -1 nogil:
+cdef int _init_nvjitlink() except -1 nogil:
     global __py_nvjitlink_init
 
     with gil, __symbol_lock:
+        # Recheck the flag after obtaining the locks
+        if __py_nvjitlink_init:
+            return 0
+
         # Load library
         handle = load_nvidia_dynamic_lib("nvJitLink")._handle_uint
 
@@ -151,7 +155,7 @@ cdef inline int _check_or_init_nvjitlink() except -1 nogil:
     if __py_nvjitlink_init:
         return 0
 
-    return __check_or_init_nvjitlink()
+    return _init_nvjitlink()
 
 
 cdef dict func_ptrs = None

cuda_bindings/cuda/bindings/_internal/nvvm.pxd

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from ..cynvvm cimport *
 

cuda_bindings/cuda/bindings/_internal/nvvm_linux.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from libc.stdint cimport intptr_t, uintptr_t
 
@@ -16,6 +16,8 @@ from cuda.pathfinder import load_nvidia_dynamic_lib
 # Extern
 ###############################################################################
 
+# You must 'from .utils import NotSupportedError' before using this template
+
 cdef extern from "<dlfcn.h>" nogil:
     void* dlopen(const char*, int)
     char* dlerror()
@@ -49,13 +51,13 @@ cdef int get_cuda_version():
     return driver_ver
 
 
+
 ###############################################################################
 # Wrapper init
 ###############################################################################
 
 cdef object __symbol_lock = threading.Lock()
 cdef bint __py_nvvm_init = False
-cdef void* __cuDriverGetVersion = NULL
 
 cdef void* __nvvmGetErrorString = NULL
 cdef void* __nvvmVersion = NULL
@@ -77,12 +79,16 @@ cdef void* load_library() except* with gil:
     return <void*>handle
 
 
-cdef int __check_or_init_nvvm() except -1 nogil:
+cdef int _init_nvvm() except -1 nogil:
     global __py_nvvm_init
 
     cdef void* handle = NULL
 
     with gil, __symbol_lock:
+        # Recheck the flag after obtaining the locks
+        if __py_nvvm_init:
+            return 0
+
         # Load function
         global __nvvmGetErrorString
         __nvvmGetErrorString = dlsym(RTLD_DEFAULT, 'nvvmGetErrorString')
@@ -179,11 +185,11 @@ cdef int __check_or_init_nvvm() except -1 nogil:
         return 0
 
 
-cdef int _check_or_init_nvvm() except -1 nogil:
+cdef inline int _check_or_init_nvvm() except -1 nogil:
     if __py_nvvm_init:
         return 0
 
-    return __check_or_init_nvvm()
+    return _init_nvvm()
 
 
 cdef dict func_ptrs = None

cuda_bindings/cuda/bindings/_internal/nvvm_windows.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from libc.stdint cimport intptr_t
 
@@ -15,7 +15,7 @@ from libc.stddef cimport wchar_t
 from libc.stdint cimport uintptr_t
 from cpython cimport PyUnicode_AsWideCharString, PyMem_Free
 
-from .utils import NotSupportedError
+# You must 'from .utils import NotSupportedError' before using this template
 
 cdef extern from "windows.h" nogil:
     ctypedef void* HMODULE
@@ -61,10 +61,10 @@ cdef int get_cuda_version():
         raise NotSupportedError('CUDA driver is not found')
     cuDriverGetVersion = GetProcAddress(handle, 'cuDriverGetVersion')
     if cuDriverGetVersion == NULL:
-        raise RuntimeError('something went wrong')
+        raise RuntimeError('Did not find cuDriverGetVersion symbol in nvcuda.dll')
     err = (<int (*)(int*) noexcept nogil>cuDriverGetVersion)(&driver_ver)
     if err != 0:
-        raise RuntimeError('something went wrong')
+        raise RuntimeError(f'cuDriverGetVersion returned error code {err}')
 
     return driver_ver
 
@@ -92,10 +92,14 @@ cdef void* __nvvmGetProgramLogSize = NULL
 cdef void* __nvvmGetProgramLog = NULL
 
 
-cdef int __check_or_init_nvvm() except -1 nogil:
+cdef int _init_nvvm() except -1 nogil:
     global __py_nvvm_init
 
     with gil, __symbol_lock:
+        # Recheck the flag after obtaining the locks
+        if __py_nvvm_init:
+            return 0
+
         # Load library
         handle = load_nvidia_dynamic_lib("nvvm")._handle_uint
 
@@ -147,7 +151,7 @@ cdef inline int _check_or_init_nvvm() except -1 nogil:
     if __py_nvvm_init:
         return 0
 
-    return __check_or_init_nvvm()
+    return _init_nvvm()
 
 
 cdef dict func_ptrs = None

cuda_bindings/cuda/bindings/cufile.pyx

@@ -12,6 +12,7 @@ import numpy as _numpy
 from cpython cimport buffer as _buffer
 from cpython.memoryview cimport PyMemoryView_FromMemory
 from enum import IntEnum as _IntEnum
+cimport cpython
 
 import cython
 
@@ -1254,7 +1255,7 @@ cpdef str get_parameter_string(int param, int len):
     with nogil:
         status = cuFileGetParameterString(<_StringConfigParameter>param, desc_str, len)
     check_status(status)
-    return _desc_str_.decode()
+    return cpython.PyUnicode_FromString(desc_str)
 
 
 cpdef set_parameter_size_t(int param, size_t value):

cuda_bindings/cuda/bindings/cynvjitlink.pxd

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from libc.stdint cimport intptr_t, uint32_t
 

cuda_bindings/cuda/bindings/cynvjitlink.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from ._internal cimport nvjitlink as _nvjitlink
 

cuda_bindings/cuda/bindings/cynvvm.pxd

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 
 ###############################################################################

cuda_bindings/cuda/bindings/cynvvm.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from ._internal cimport nvvm as _nvvm
 

cuda_bindings/cuda/bindings/nvjitlink.pxd

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.0.1 to 12.9.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.0.1 to 13.0.2. Do not modify it directly.
 
 from libc.stdint cimport intptr_t, uint32_t