python3Packages.cuda-bindings: fix lib lookups (#512828)

  buildPythonPackage,

  fetchFromGitHub,

  cudaPackages,

  replaceVars,

  addDriverRunpath,

  # build-system

  symlinkJoin,

  # dependencies

  cuda-pathfinder,

  numpy,

  # tests

  numpy,

  pytestCheckHook,

  # passthru

    hash = "sha256-uRv27h2b6wXC8oOf5k2KxZ0bUFNvNu6XO05FBbJcU1k=";

  };

  # Apply patch relative to cuda_bindings

  patchFlags = [ "-p2" ];

  patches = [

    (replaceVars ./patch-nvidia-libs-paths.patch {

      libcudart = lib.getLib cudaPackages.cuda_cudart;

      libcufile = lib.getLib cudaPackages.libcufile;

      libnvfatbin = lib.getLib cudaPackages.libnvfatbin;

      libnvjitlink = lib.getLib cudaPackages.libnvjitlink;

      libnvml = addDriverRunpath.driverLink;

      libnvrtc = lib.getLib cudaPackages.cuda_nvrtc;

      libnvvm = "${cudaPackages.cuda_nvcc}/nvvm";

    })

  ];

  sourceRoot = "${finalAttrs.src.name}/cuda_bindings";

  postPatch =

    cudaPackages.libcufile # cufile.h

  ];

  pythonRemoveDeps = [

    # We circumvent cuda_pathfinder to localize nvidia libs with patches

    "cuda-pathfinder"

  ];

  dependencies = [

    cuda-pathfinder

    # Not explicitly listed as a dependency, but is required at import time

    numpy

  ];

  pythonImportsCheck = [

    "cuda"

    "cuda.cuda"

    "cuda.cudart"

    "cuda.nvrtc"

    "cuda.bindings.cufile"

    "cuda.bindings.driver"

    "cuda.bindings.nvfatbin"

    "cuda.bindings.nvjitlink"

    "cuda.bindings.nvml"

    "cuda.bindings.nvrtc"

    "cuda.bindings.nvvm"

    "cuda.bindings.runtime"

  ];

  preCheck = ''

  '';

  nativeCheckInputs = [

    numpy

    pytestCheckHook

  ];

    "tests/test_nvjitlink.py"

  ];

  disabledTests = [

    # sysfs cpu topology is not available in the sandbox, causing:

    #   cuda.bindings.nvml.UnknownError: Unknown Error

    #   hwloc/linux: failed to find sysfs cpu topology directory, aborting linux discovery.

    "test_device_get_cpu_affinity_within_scope"

    "test_device_get_memory_affinity"

    # Requires the nvidia_fs kernel module (GPUDirect Storage), causing:

    #   cuda.bindings.cufile.cuFileError: NVFS_SETUP_ERROR (5033): NVFS driver initialization error

    "test_buf_register_already_registered"

    "test_buf_register_host_memory"

    "test_buf_register_invalid_flags"

    "test_buf_register_large_buffer"

    "test_buf_register_multiple_buffers"

    "test_buf_register_simple"

    "test_driver_open"

  ];

  # Tests need access to a GPU

  doCheck = false;

  passthru.gpuCheck = cuda-bindings.overridePythonAttrs {

diff --git a/cuda_bindings/cuda/bindings/_bindings/cynvrtc.pyx.in b/cuda_bindings/cuda/bindings/_bindings/cynvrtc.pyx.in

index f684be6870..9437de642b 100644

--- a/cuda_bindings/cuda/bindings/_bindings/cynvrtc.pyx.in

+++ b/cuda_bindings/cuda/bindings/_bindings/cynvrtc.pyx.in

@@ -8,7 +8,7 @@ cimport cuda.bindings._lib.windll as windll

 {{else}}

 cimport cuda.bindings._lib.dlfcn as dlfcn

 {{endif}}

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 from libc.stdint cimport intptr_t, uintptr_t

 import threading

@@ -47,7 +47,7 @@ cdef int _cuPythonInit() except -1 nogil:

     # Load library

     with gil, __symbol_lock:

         {{if 'Windows' == platform.system()}}

-        handle = load_nvidia_dynamic_lib("nvrtc")._handle_uint

+        handle = CDLL("@libnvrtc@/lib/libnvrtc.so")._handle

         # Load function

         {{if 'nvrtcGetErrorString' in found_functions}}

@@ -156,7 +156,7 @@ cdef int _cuPythonInit() except -1 nogil:

         {{endif}}

         {{else}}

-        handle = <void*><uintptr_t>(load_nvidia_dynamic_lib("nvrtc")._handle_uint)

+        handle = <void*><uintptr_t>(CDLL("@libnvrtc@/lib/libnvrtc.so")._handle)

         # Load function

         {{if 'nvrtcGetErrorString' in found_functions}}

diff --git a/cuda_bindings/cuda/bindings/_internal/cufile_linux.pyx b/cuda_bindings/cuda/bindings/_internal/cufile_linux.pyx

index 78ea77b283..0b875a6cd3 100644

--- a/cuda_bindings/cuda/bindings/_internal/cufile_linux.pyx

+++ b/cuda_bindings/cuda/bindings/_internal/cufile_linux.pyx

@@ -9,7 +9,7 @@ import threading

 from .utils import FunctionNotFoundError, NotSupportedError

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 import cython

@@ -95,7 +95,7 @@ cdef void* __cuFileSetParameterString = NULL

 cdef void* load_library() except* with gil:

-    cdef uintptr_t handle = load_nvidia_dynamic_lib("cufile")._handle_uint

+    cdef uintptr_t handle = CDLL("@libcufile@/lib/libcufile.so")._handle

     return <void*>handle

diff --git a/cuda_bindings/cuda/bindings/_internal/nvfatbin_linux.pyx b/cuda_bindings/cuda/bindings/_internal/nvfatbin_linux.pyx

index f5a9bbd218..8271f7aa20 100644

--- a/cuda_bindings/cuda/bindings/_internal/nvfatbin_linux.pyx

+++ b/cuda_bindings/cuda/bindings/_internal/nvfatbin_linux.pyx

@@ -9,7 +9,7 @@ from libc.stdint cimport intptr_t, uintptr_t

 import threading

 from .utils import FunctionNotFoundError, NotSupportedError

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 ###############################################################################

@@ -73,7 +73,7 @@ cdef void* __nvFatbinAddTileIR = NULL

 cdef void* load_library() except* with gil:

-    cdef uintptr_t handle = load_nvidia_dynamic_lib("nvfatbin")._handle_uint

+    cdef uintptr_t handle = CDLL("@libnvfatbin@/lib/libnvfatbin.so")._handle

     return <void*>handle

diff --git a/cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx b/cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

index d676aac372..ed3c000566 100644

--- a/cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

+++ b/cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

@@ -9,7 +9,7 @@ from libc.stdint cimport intptr_t, uintptr_t

 import threading

 from .utils import FunctionNotFoundError, NotSupportedError

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 ###############################################################################

@@ -76,7 +76,7 @@ cdef void* __nvJitLinkVersion = NULL

 cdef void* load_library() except* with gil:

-    cdef uintptr_t handle = load_nvidia_dynamic_lib("nvJitLink")._handle_uint

+    cdef uintptr_t handle = CDLL("@libnvjitlink@/lib/libnvJitLink.so")._handle

     return <void*>handle

diff --git a/cuda_bindings/cuda/bindings/_internal/nvml_linux.pyx b/cuda_bindings/cuda/bindings/_internal/nvml_linux.pyx

index 28f0919423..852f75e46d 100644

--- a/cuda_bindings/cuda/bindings/_internal/nvml_linux.pyx

+++ b/cuda_bindings/cuda/bindings/_internal/nvml_linux.pyx

@@ -10,7 +10,7 @@ import threading

 from .utils import FunctionNotFoundError, NotSupportedError

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 ###############################################################################

@@ -406,7 +406,7 @@ cdef void* __nvmlDeviceSetRusdSettings_v1 = NULL

 cdef void* load_library() except* with gil:

-    cdef uintptr_t handle = load_nvidia_dynamic_lib("nvml")._handle_uint

+    cdef uintptr_t handle = CDLL("@libnvml@/lib/libnvidia-ml.so")._handle

     return <void*>handle

diff --git a/cuda_bindings/cuda/bindings/_internal/nvvm_linux.pyx b/cuda_bindings/cuda/bindings/_internal/nvvm_linux.pyx

index 8a84834a9a..a3ced66807 100644

--- a/cuda_bindings/cuda/bindings/_internal/nvvm_linux.pyx

+++ b/cuda_bindings/cuda/bindings/_internal/nvvm_linux.pyx

@@ -9,7 +9,7 @@ from libc.stdint cimport intptr_t, uintptr_t

 import threading

 from .utils import FunctionNotFoundError, NotSupportedError

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 ###############################################################################

@@ -75,7 +75,7 @@ cdef void* __nvvmGetProgramLog = NULL

 cdef void* load_library() except* with gil:

-    cdef uintptr_t handle = load_nvidia_dynamic_lib("nvvm")._handle_uint

+    cdef uintptr_t handle = CDLL("@libnvvm@/lib/libnvvm.so")._handle

     return <void*>handle

diff --git a/cuda_bindings/cuda/bindings/cyruntime.pyx.in b/cuda_bindings/cuda/bindings/cyruntime.pyx.in

index 5cd65fbd96..fbbcffbbb7 100644

--- a/cuda_bindings/cuda/bindings/cyruntime.pyx.in

+++ b/cuda_bindings/cuda/bindings/cyruntime.pyx.in

@@ -1874,7 +1874,7 @@ cdef cudaError_t cudaGraphicsVDPAURegisterOutputSurface(cudaGraphicsResource** r

 {{if True}}

 from libc.stdint cimport uintptr_t

-from cuda.pathfinder import load_nvidia_dynamic_lib

+from ctypes import CDLL

 {{if 'Windows' == platform.system()}}

 cimport cuda.bindings._lib.windll as windll

 {{else}}

@@ -1884,11 +1884,11 @@ cimport cuda.bindings._lib.dlfcn as dlfcn

 cdef cudaError_t getLocalRuntimeVersion(int* runtimeVersion) except ?cudaErrorCallRequiresNewerDriver nogil:

     # Load

     with gil:

-        loaded_dl = load_nvidia_dynamic_lib("cudart")

+        loaded_dl = CDLL("@libcudart@/lib/libcudart.so")

         {{if 'Windows' == platform.system()}}

-        handle = <uintptr_t>loaded_dl._handle_uint

+        handle = <uintptr_t>loaded_dl._handle

         {{else}}

-        handle = <void *><uintptr_t>loaded_dl._handle_uint

+        handle = <void *><uintptr_t>loaded_dl._handle

         {{endif}}

     {{if 'Windows' == platform.system()}}

diff --git a/cuda_bindings/cuda/bindings/nvml.pyx b/cuda_bindings/cuda/bindings/nvml.pyx

index 42c9fdcc87..0661379a68 100644

--- a/cuda_bindings/cuda/bindings/nvml.pyx

+++ b/cuda_bindings/cuda/bindings/nvml.pyx

@@ -27471,4 +27471,3 @@ cpdef str vgpu_type_get_name(unsigned int vgpu_type_id):

         __status__ = nvmlVgpuTypeGetName(<nvmlVgpuTypeId_t>vgpu_type_id, vgpu_type_name, <unsigned int*>size)

     check_status(__status__)

     return cpython.PyUnicode_FromStringAndSize(vgpu_type_name, size[0])

-