NIXL (NVIDIA Inference Xfer Library) is an open-source Python API designed to accelerate point-to-point communications in AI inference frameworks. It provides a unified abstraction layer over various memory types (CPU, GPU) and storage (file, block, object store) through a modular plugin architecture. The `nixl-cu12` package specifically targets CUDA 12 environments. NIXL is actively maintained with frequent releases, with version 1.0.0 being the current stable release.
pip install nixl-cu12Verified import paths — ran on the pinned version, not inferred.
This quickstart demonstrates how to import the `nixl` library and initialize a NIXL agent. The agent is the core component for managing data transfers within the NIXL framework.
Review and update your code to utilize the new Device API V2. For UCX backend configurations, adapt to the updated request handling mechanisms. Refer to the official NIXL 1.0.0 release notes and documentation for detailed migration guides.
Ensure only the package corresponding to your target CUDA version is installed (`nixl-cu12` for CUDA 12, `nixl-cu13` for CUDA 13). If using the `nixl` meta-package, specify the correct extra, e.g., `pip install "nixl[cu12]"`.
Run NIXL in a Linux host, container, or virtual machine environment.
Verify that all required backend libraries (e.g., UCX) are correctly installed and accessible. Check NIXL agent initialization parameters and ensure they match your system's hardware and software configuration. Review NIXL's backend guide for specific setup instructions.