NVIDIA JIT LTO Library (CUDA 12)

Warnings

• gotcha This package `nvidia-nvjitlink-cu12` is a *metapackage* for a native C library (`nvJitLink`) and part of the CUDA Toolkit. It does not provide direct Python APIs for JIT linking functionality. Python users typically interact with `nvJitLink` through higher-level libraries like `pynvjitlink-cu12`, `Numba-CUDA`, or `CuPy` that are built to utilize it.
  Fix: Use `pynvjitlink-cu12` for direct Python bindings to `nvJitLink`, or rely on libraries like `numba-cuda` or `cupy` which integrate `nvJitLink` functionality.
• breaking The `nvJitLink` library has compatibility constraints regarding linking objects across different CUDA major versions, especially for LTO-IR (Link Time Optimization Intermediate Representation) inputs. Linking across major versions (e.g., CUDA 11.x with 12.x) may not work for LTO-IR inputs, although it generally works for ELF and PTX inputs.
  Fix: Ensure all inputs for JIT LTO (especially LTO-IR) are compiled with a CUDA Toolkit version compatible with the `nvJitLink` library being used. Compatibility is only guaranteed within a major CUDA release for LTO-IR. Update your CUDA Toolkit and associated `nvidia-nvjitlink-cu12` package to match if encountering linking errors.
• gotcha Installation of `nvidia-nvjitlink-cu12` from PyPI typically requires a pre-existing CUDA driver, even if it aims to install CUDA runtime components. For full development capabilities, a system-wide CUDA Toolkit installation is often still necessary as these pip wheels primarily provide runtime components and may not include developer tools (like `nvcc`).
  Fix: Ensure a compatible NVIDIA GPU driver is installed. For full CUDA development, consider installing the complete NVIDIA CUDA Toolkit alongside these Python packages, or use `cupy-cudaXX[ctk]` for a more complete pip-based CUDA runtime environment with CuPy.
• deprecated The `cuLink*` APIs in the CUDA Driver, which provided similar functionality to `nvJitLink`, have been deprecated for use with LTO-IR. `nvJitLink` is the recommended path for JIT LTO.
  Fix: Migrate from deprecated `cuLink*` APIs to the `nvJitLink` library for JIT LTO functionality.

Install

• pip install nvidia-nvjitlink-cu12 Install latest version
• pip install --extra-index-url https://pypi.nvidia.com nvidia-pyindex pip install nvidia-nvjitlink-cu12 Install with NVIDIA PyPI index (recommended for full ecosystem)

Imports

• None

  This package primarily installs native libraries; direct Python imports for JIT linking functionality are not typically made from `nvidia-nvjitlink-cu12`. Higher-level libraries like `pynvjitlink-cu12`, `numba.cuda`, or `cupy` provide the Python interfaces that utilize the installed `nvJitLink` library.

  nvidia-nvjitlink-cu12 is a metapackage for the native library, not a direct Python API library for JIT linking.

Quickstart

Since `nvidia-nvjitlink-cu12` is a metapackage for native CUDA components, its functionality is indirectly exposed through other Python libraries that depend on the `nvJitLink` library. This quickstart demonstrates how to check for the availability of a CUDA-capable GPU and the CUDA runtime environment using `CuPy` and `Numba-CUDA`, which would implicitly rely on `nvJitLink` if their features that use it are invoked. This verifies that the underlying CUDA toolkit, which `nvidia-nvjitlink-cu12` helps install, is correctly set up.

import os

try:
    import cupy as cp
    print(f"CuPy is installed. CUDA available: {cp.cuda.is_available()}")
    if cp.cuda.is_available():
        print(f"CuPy CUDA Device Count: {cp.cuda.runtime.getDeviceCount()}")
except ImportError:
    print("CuPy not installed. Install with `pip install cupy-cuda12x` to verify CUDA environment.")

try:
    import numba.cuda
    print(f"Numba CUDA is installed. CUDA available: {numba.cuda.is_available()}")
    if numba.cuda.is_available():
        print(f"Numba CUDA Device Count: {numba.cuda.count_devices()}")
except ImportError:
    print("Numba-CUDA not installed. Install with `pip install numba-cuda` to verify CUDA environment.")

print("\nThis output indicates whether higher-level Python libraries can detect and use the CUDA environment, which includes the nvJitLink library provided by nvidia-nvjitlink-cu12.")