NVIDIA CUDA CUPTI Runtime Libraries

13.2.23 · active · verified Thu Apr 09

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.

Warnings

gotcha 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.
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.
gotcha 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.
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.
gotcha 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.
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.

Install

pip install nvidia-cuda-cupti Install via pip

Quickstart

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.

# This library provides low-level CUDA profiling tools runtime libraries.
# It does NOT expose direct Python APIs for import and use via `import nvidia_cuda_cupti`.
# Its primary purpose is to make the necessary C/C++ shared libraries (e.g., libcupti.so)
# available in the environment for other Python libraries (e.g., PyTorch, TensorFlow, Numba, CuPy)
# that interface with CUDA and CUPTI for profiling and performance analysis.

print("nvidia-cuda-cupti is installed as a runtime dependency.")
print("It provides the CUPTI shared libraries for CUDA-enabled applications.")
print("To verify its effect, you would typically use a profiling tool or framework that relies on CUPTI,")
print("e.g., enabling profilers in PyTorch or TensorFlow, which would then leverage these libraries.")
print("No direct Python interaction with this package is expected after installation.")

view raw JSON →