# ai-dynamo/nixl

NVIDIA Inference Xfer Library (NIXL)

Repository: https://github.com/ai-dynamo/nixl
Canonical: https://ross.abutalabs.com/products/nixl
Language: C++
License: NOASSERTION
License Family: other
Last push: 2026-09-02T16:24:35+00:00

## Health v2 (maintenance only)
Score: 88/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 100, release rhythm 100, longevity 39
- inputs: {"age_days": 546, "days_push": 0, "days_rel": 1, "gap_med": 18.5, "n_releases_24m": 27}
- flags: no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1227, forks 426 (observed 2026-09-03T02:15:08.870444+00:00)

## What it is
NVIDIA Inference Xfer Library (NIXL) is a C++/Python library that accelerates point-to-point communication in AI inference frameworks like NVIDIA Dynamo. It abstracts over various memory types (CPU, GPU) and storage backends (file, block, object store) through a modular plug-in architecture.

## Use cases
- accelerate KV cache transfer between inference engine instances
- abstract point-to-point data transfer across CPU and GPU memory
- plug different transport backends like UCX into an inference framework
- benchmark memory transfer performance for LLM serving
- transfer tensors between distributed inference workers
- read and write to object or block storage from GPU code

## When to choose
- building or extending AI inference frameworks on Linux with NVIDIA GPUs
- you need high-performance point-to-point transfers between heterogeneous memory and storage
- you want a pluggable abstraction over transports like UCX
- you're integrating with NVIDIA Dynamo or similar disaggregated serving stacks

## When to avoid
- you need macOS or Windows support - NIXL is Linux-only
- you need general-purpose networking rather than inference-focused data transfer
- your project doesn't involve GPU or accelerator memory movement
- you need a stable API on a non-NVIDIA stack

## Facets
- artifact type: library
- maturity: active
- function: networking, plugin-system, benchmarking, caching, sdk
- domain: machine-learning, large-language-models, microservices, gpu-computing, developer-tools
- platform: cpp, python
- tags: data-transfer, inference-infrastructure, nvidia, ucx, kv-cache-transfer, point-to-point-communication, memory-abstraction, cuda, linux, gpu

## Member repositories
- ai-dynamo/nixl (main) score 88

## Provenance
- Observed fields: from GitHub, fetched 2026-09-03T02:15:08.870444+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:17:14.225613+00:00, confidence not recorded.
  - readme: https://github.com/ai-dynamo/nixl (fetched 2026-09-03T02:15:08.870444+00:00, sha ea7359fbdc54)
- Data as of 2026-08-30T08:39:29.467469+00:00.