NVIDIA NIXL
An AI inference data transfer library for Linux that connects CPU and GPU memory with storage through pluggable backends.
NVIDIA NIXL is a data transfer library for developers building AI inference systems such as NVIDIA Dynamo. It provides a common interface for moving data between inference components.
- Memory and storage: Transfer data across CPU and GPU memory, file storage, block storage and object stores.
- Pluggable backends: Work with transports and storage systems such as UCX, POSIX, Mooncake and GDS.
- Language bindings: Integrate the library into C++, Python or Rust applications.
- Distributed coordination: Exchange metadata between nodes through optional ETCD support and assess transfers with NIXLBench or KVBench.
NIXL supports Linux, including container and virtual-machine deployments. NVIDIA GPU transfers use CUDA; AMD GPU transfers require ROCm-enabled UCX. Backend availability varies by hardware, and CUDA-dependent backends are unavailable on ROCm-only hosts.
The core uses Apache 2.0 and the DeepEP-derived example uses MIT. The Python wheels also bundle NVIDIA networking modules under a proprietary license.
Logotyper
