{"id":1586,"library":"nvidia-cuda-cupti","title":"NVIDIA CUDA CUPTI Runtime Libraries","description":"The `nvidia-cuda-cupti` package provides the CUDA Profiling Tools Interface (CUPTI) runtime libraries, essential for profiling CUDA applications. It's a low-level dependency that makes the necessary shared libraries available in the environment. The current version is 13.2.23, and it's released in alignment with NVIDIA's CUDA Toolkit versions.","status":"active","version":"13.2.23","language":"python","source_language":"en","source_url":"https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html","tags":["nvidia","cuda","cupti","profiling","gpu","runtime","low-level","dependency"],"install":[{"cmd":"pip install nvidia-cuda-cupti","lang":"bash","label":"Install via pip"}],"dependencies":[],"imports":[],"quickstart":{"code":"# This library provides low-level CUDA profiling tools runtime libraries.\n# It does NOT expose direct Python APIs for import and use via `import nvidia_cuda_cupti`.\n# Its primary purpose is to make the necessary C/C++ shared libraries (e.g., libcupti.so)\n# available in the environment for other Python libraries (e.g., PyTorch, TensorFlow, Numba, CuPy)\n# that interface with CUDA and CUPTI for profiling and performance analysis.\n\nprint(\"nvidia-cuda-cupti is installed as a runtime dependency.\")\nprint(\"It provides the CUPTI shared libraries for CUDA-enabled applications.\")\nprint(\"To verify its effect, you would typically use a profiling tool or framework that relies on CUPTI,\")\nprint(\"e.g., enabling profilers in PyTorch or TensorFlow, which would then leverage these libraries.\")\nprint(\"No direct Python interaction with this package is expected after installation.\")","lang":"python","description":"This quickstart demonstrates that `nvidia-cuda-cupti` is a runtime dependency without direct Python imports. Its presence is primarily for other CUDA-aware libraries to utilize its underlying C/C++ shared objects for profiling."},"warnings":[{"fix":"Simply install the package. Its effects are indirect, making CUPTI libraries available for other Python libraries like PyTorch or TensorFlow that rely on CUDA profiling.","message":"This package is a runtime distribution of C/C++ libraries, not a Python package with direct Python APIs. Do not attempt to import `nvidia_cuda_cupti` in your Python code, as it does not expose any Python symbols for direct use.","severity":"gotcha","affected_versions":"All versions"},{"fix":"Refer to the NVIDIA CUDA Toolkit release notes and compatibility matrix for the recommended versions. Mismatches can lead to runtime errors or incorrect profiling data.","message":"Compatibility with your CUDA Toolkit and GPU drivers is crucial. Ensure the version of `nvidia-cuda-cupti` is compatible with the CUDA Toolkit version used by your main deep learning frameworks (e.g., PyTorch, TensorFlow) and your NVIDIA GPU driver version.","severity":"gotcha","affected_versions":"All versions"},{"fix":"Prefer using `pip install nvidia-cuda-cupti` if your other libraries are designed to dynamically link to system-provided CUPTI. If a library explicitly bundles CUPTI, ensure its version is compatible or consider using a consistent environment (e.g., containers) to manage library versions.","message":"Potential for conflicts if other Python packages (e.g., specific PyTorch or TensorFlow builds) bundle their own versions of CUPTI libraries. This can lead to symbol conflicts or unexpected behavior.","severity":"gotcha","affected_versions":"All versions"}],"env_vars":null,"search_vec":"'13.2.23':47 'align':53 'applic':25 'avail':39 'cuda':2,9,14,24,57,61 'cupti':3,10,18,62 'current':44 'depend':32,69 'environ':42 'essenti':21 'gpu':64 'interfac':17 'level':31,68 'librari':5,20,38 'low':30,67 'low-level':29,66 'make':34 'necessari':36 'nvidia':1,8,55,60 'nvidia-cuda-cupti':7 'packag':11 'profil':15,23,63 'provid':12 'releas':51 'runtim':4,19,65 'share':37 'tool':16 'toolkit':58 'version':45,59","created_at":"2026-04-09T03:54:29.798133+00:00","updated_at":"2026-04-16T17:26:39.671669+00:00","problems":[{"fix":"Add the directory containing `libcupti.so` (e.g., `/usr/local/cuda/extras/CUPTI/lib64` or `/usr/local/cuda-X.Y/extras/CUPTI/lib64`) to your `LD_LIBRARY_PATH` (Linux) or `PATH` (Windows) environment variable. Ensure the CUDA toolkit is fully and correctly installed and that its version is compatible with your profiling tool. For specific missing versions, creating a symbolic link to an existing minor version might work (e.g., `sudo ln -s /path/to/libcupti.so.X.Y /path/to/libcupti.so.X.0`).","cause":"The system's dynamic linker cannot find the `libcupti.so` library because its directory is not included in the `LD_LIBRARY_PATH` (Linux) or `PATH` (Windows) environment variables, or the library is missing/corrupted, potentially due to version mismatches.","error":"CUPTI error: CUPTI could not be loaded or symbol could not be found."},{"fix":"Verify that your NVIDIA driver, CUDA Toolkit, and any software utilizing CUPTI (e.g., PyTorch, TensorFlow, profiling tools) are mutually compatible by checking their respective documentation and compatibility matrices. Ensure CUPTI initialization happens after the CUDA context is fully established. In some cases, a `conda update --all` or selectively updating/downgrading packages might resolve conflicts.","cause":"This error typically indicates an incompatibility between the installed CUDA driver and the CUPTI version, or that CUPTI functions are being invoked too early before the CUDA environment (context) has been fully set up by the application, common in frameworks like PyTorch.","error":"CUPTI_ERROR_NOT_INITIALIZED (15)"},{"fix":"Update your NVIDIA GPU display driver to a version that is officially supported by your CUDA Toolkit and the specific CUPTI features you are trying to use. Consult the CUDA Toolkit Release Notes for driver compatibility information.","cause":"This error occurs when attempting to use specific CUPTI features, such as certain Event or Metric APIs, which are not supported by the currently installed NVIDIA GPU driver version, even if the CUDA Toolkit version appears compatible.","error":"CUPTI_ERROR_NOT_SUPPORTED (27)"},{"fix":"To allow non-admin users to profile, add `options nvidia \"NVreg_RestrictProfilingToAdminUsers=0\"` to a NVIDIA kernel module configuration file (e.g., `/etc/modprobe.d/nvidia-kernel-common.conf` on Linux), then run `sudo update-initramfs -u` and reboot the system. Alternatively, run the profiling application with root privileges (e.g., `sudo your_profiling_command`).","cause":"The user attempting to run the profiling operation does not have the necessary system-level permissions to access and utilize the GPU profiling capabilities provided by CUPTI.","error":"CUPTI has no sufficient privilege / CUPTI_ERROR_INSUFFICIENT_PRIVILEGES"}],"ecosystem":"pypi","meta_description":null,"install_score":null,"quickstart_score":null,"quickstart_tag":null,"pypi_latest":null,"cli_name":"","cli_version":null,"type":"library","homepage":"https://developer.nvidia.com/cuda-zone","github":null,"docs":null,"changelog":null,"pypi":"https://pypi.org/project/nvidia-cuda-cupti/","npm":null,"openapi_spec":null,"status_page":null,"smithery":null,"categories":["ai-ml","devops"],"base_url":null,"auth_type":null,"provenance":{"verified_status":null,"verified_at":null,"last_verified":"2026-04-09","next_check":"2026-07-08","install_tag":null}}