Coordinate cross-node memory transfers and metadata exchange with NIXL.
npx skills add https://github.com/air-gapped/skills --skill nvidia-nixl
Please help me install this Agent Skill. Skill: nvidia-nixl Source: https://github.com/air-gapped/skills/tree/main/.claude/skills/nvidia-nixl Command: npx skills add https://github.com/air-gapped/skills --skill nvidia-nixl
NVIDIA Inference Xfer Library (NIXL) provides a transport layer for high-performance cross-node memory transfers and metadata coordination in distributed inference stacks.
Launch two processes acting as target and initiator and perform a minimal two-peer tensor transfer using the Python API.
High-intent search queries and answers about installing and using this skill.
Cross-node KV cache transfers for disaggregated inference are coordinated through NIXL, which manages memory registration, descriptor lists, and transfer posting across multiple network and storage backends. It supports Dynamo, vLLM, and SGLang deployments to enable scalable distributed inference.
Cross-node memory transfers support pluggable backends including UCX, libfabric, GDS, POSIX, and HF3FS via a plugin manager. This allows the transport layer to cover various network and storage fabrics for distributed inference stacks.
Metadata exchange for elastic clusters works via either side-channel TCP or ETCD-based coordination. This enables Dynamo, vLLM, and SGLang deployments to dynamically scale by sharing transfer descriptors and state across nodes.
Yes, NIXL can be used with vLLM and SGLang, as well as Dynamo deployments. It provides the transport layer for high-performance cross-node memory transfers and metadata coordination required to scale these distributed inference frameworks.
The transfer lifecycle involves registering memory, creating descriptor lists, initializing the transfer, posting it, and checking the transfer state. Per-transfer telemetry is also provided for observability throughout the cross-node memory transfer process.
A minimal two-peer tensor transfer is performed by launching two processes acting as target and initiator using the Python API. These processes coordinate memory registration and descriptor exchange to execute the cross-node tensor transfer.
sayalinvidiaAssess NIXL/UCX/NCCL transport readiness for Dynamo disaggregated serving.
uv-xiaoOrchestrates distributed ML training across multi-node clusters using Ray Train.
air-gappedCoordinate vLLM KV caching across native offload, LMCache, Nixl, and Mooncake backends.
theadriannBuild a Node-native Droid integration layer with SQLite-backed session management.
