Make cuda_bindings build cleanly with clang (#1658)

* Make cuda_bindings build cleanly with clang

* Remove `-Werror`

* Fix Windows build

Author: mdboom

cuda_bindings/cuda/bindings/cufile.pyx

@@ -2901,25 +2901,25 @@ cpdef void batch_io_destroy(intptr_t batch_idp) except*:
 
 cpdef read_async(intptr_t fh, intptr_t buf_ptr_base, intptr_t size_p, intptr_t file_offset_p, intptr_t buf_ptr_offset_p, intptr_t bytes_read_p, intptr_t stream):
     with nogil:
-        __status__ = cuFileReadAsync(<Handle>fh, <void*>buf_ptr_base, <size_t*>size_p, <off_t*>file_offset_p, <off_t*>buf_ptr_offset_p, <ssize_t*>bytes_read_p, <void*>stream)
+        __status__ = cuFileReadAsync(<Handle>fh, <void*>buf_ptr_base, <size_t*>size_p, <off_t*>file_offset_p, <off_t*>buf_ptr_offset_p, <ssize_t*>bytes_read_p, <CUstream>stream)
     check_status(__status__)
 
 
 cpdef write_async(intptr_t fh, intptr_t buf_ptr_base, intptr_t size_p, intptr_t file_offset_p, intptr_t buf_ptr_offset_p, intptr_t bytes_written_p, intptr_t stream):
     with nogil:
-        __status__ = cuFileWriteAsync(<Handle>fh, <void*>buf_ptr_base, <size_t*>size_p, <off_t*>file_offset_p, <off_t*>buf_ptr_offset_p, <ssize_t*>bytes_written_p, <void*>stream)
+        __status__ = cuFileWriteAsync(<Handle>fh, <void*>buf_ptr_base, <size_t*>size_p, <off_t*>file_offset_p, <off_t*>buf_ptr_offset_p, <ssize_t*>bytes_written_p, <CUstream>stream)
     check_status(__status__)
 
 
 cpdef stream_register(intptr_t stream, unsigned flags):
     with nogil:
-        __status__ = cuFileStreamRegister(<void*>stream, flags)
+        __status__ = cuFileStreamRegister(<CUstream>stream, flags)
     check_status(__status__)
 
 
 cpdef stream_deregister(intptr_t stream):
     with nogil:
-        __status__ = cuFileStreamDeregister(<void*>stream)
+        __status__ = cuFileStreamDeregister(<CUstream>stream)
     check_status(__status__)
 
 

cuda_bindings/cuda/bindings/nvml.pxd

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.9.1 to 13.1.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.9.1 to 13.1.1, generator version 0.3.1.dev1283+gc7bc6fa75. Do not modify it directly.
 
 from libc.stdint cimport intptr_t
 
@@ -342,7 +342,6 @@ cpdef object device_get_grid_licensable_features_v4(intptr_t device)
 cpdef unsigned int get_vgpu_driver_capabilities(int capability) except? 0
 cpdef unsigned int device_get_vgpu_capabilities(intptr_t device, int capability) except? 0
 cpdef str vgpu_type_get_class(unsigned int vgpu_type_id)
-cpdef str vgpu_type_get_name(unsigned int vgpu_type_id)
 cpdef unsigned int vgpu_type_get_gpu_instance_profile_id(unsigned int vgpu_type_id) except? 0
 cpdef tuple vgpu_type_get_device_id(unsigned int vgpu_type_id)
 cpdef unsigned long long vgpu_type_get_framebuffer_size(unsigned int vgpu_type_id) except? 0

cuda_bindings/cuda/bindings/nvml.pyx

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.9.1 to 13.1.1. Do not modify it directly.
+# This code was automatically generated across versions from 12.9.1 to 13.1.1, generator version 0.3.1.dev1283+gc7bc6fa75. Do not modify it directly.
 
 cimport cython  # NOQA
 
@@ -23189,22 +23189,6 @@ cpdef str vgpu_type_get_class(unsigned int vgpu_type_id):
     return cpython.PyUnicode_FromString(vgpu_type_class)
 
 
-cpdef str vgpu_type_get_name(unsigned int vgpu_type_id):
-    """Retrieve the vGPU type name.
-
-    Args:
-        vgpu_type_id (unsigned int): Handle to vGPU type.
-
-    .. seealso:: `nvmlVgpuTypeGetName`
-    """
-    cdef unsigned int size = 64
-    cdef char[64] vgpu_type_name
-    with nogil:
-        __status__ = nvmlVgpuTypeGetName(<nvmlVgpuTypeId_t>vgpu_type_id, vgpu_type_name, <unsigned int*>size)
-    check_status(__status__)
-    return cpython.PyUnicode_FromString(vgpu_type_name)
-
-
 cpdef unsigned int vgpu_type_get_gpu_instance_profile_id(unsigned int vgpu_type_id) except? 0:
     """Retrieve the GPU Instance Profile ID for the given vGPU type ID. The API will return a valid GPU Instance Profile ID for the MIG capable vGPU types, else INVALID_GPU_INSTANCE_PROFILE_ID is returned.
 
@@ -26228,3 +26212,19 @@ cpdef str device_get_current_clock_freqs(intptr_t device):
         __status__ = nvmlDeviceGetCurrentClockFreqs(<Device>device, current_clock_freqs)
     check_status(__status__)
     return cpython.PyUnicode_FromString(current_clock_freqs[0].str)
+
+
+cpdef str vgpu_type_get_name(unsigned int vgpu_type_id):
+    """Retrieve the vGPU type name.
+
+    Args:
+        vgpu_type_id (unsigned int): Handle to vGPU type.
+
+    .. seealso:: `nvmlVgpuTypeGetName`
+    """
+    cdef unsigned int[1] size = [64]
+    cdef char[64] vgpu_type_name
+    with nogil:
+        __status__ = nvmlVgpuTypeGetName(<nvmlVgpuTypeId_t>vgpu_type_id, vgpu_type_name, <unsigned int*>size)
+    check_status(__status__)
+    return cpython.PyUnicode_FromString(vgpu_type_name)

cuda_bindings/setup.py

@@ -1,13 +1,15 @@
-# SPDX-FileCopyrightText: Copyright (c) 2021-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2021-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
 import atexit
 import contextlib
+import functools
 import glob
 import os
 import pathlib
 import platform
 import shutil
+import subprocess
 import sys
 import sysconfig
 import tempfile
@@ -270,6 +272,7 @@ def generate_output(infile, local):
         "-fpermissive",
         "-Wno-deprecated-declarations",
         "-fno-var-tracking-assignments",
+        "-Wno-unused-function",
     ]
     if "--debug" in sys.argv:
         extra_cythonize_kwargs["gdb_debug"] = True
@@ -391,6 +394,21 @@ def do_cythonize(extensions):
 building_wheel = False
 
 
+def is_clang(compiler):
+    @functools.lru_cache
+    def _is_clang(compiler_cxx):
+        try:
+            output = subprocess.check_output([*compiler_cxx, "--version"])  # noqa: S603
+        except subprocess.CalledProcessError:
+            return False
+        lines = output.decode().splitlines()
+        return len(lines) > 0 and "clang" in lines[0]
+
+    if not hasattr(compiler, "compiler_cxx"):
+        return False
+    return _is_clang(tuple(compiler.compiler_cxx))
+
+
 class WheelsBuildExtensions(bdist_wheel):
     def run(self):
         global building_wheel
@@ -405,6 +423,9 @@ def initialize_options(self):
             self.parallel = nthreads
 
     def build_extension(self, ext):
+        if is_clang(self.compiler):
+            ext.extra_compile_args = [x for x in ext.extra_compile_args if x != "-fno-var-tracking-assignments"]
+
         if building_wheel and sys.platform == "linux" and "--debug" not in sys.argv:
             # Strip binaries to remove debug symbols
             ext.extra_link_args.append("-Wl,--strip-all")