uccl-project/uccl

Health v2 · maintenance only

Adoption not part of the score

1496 stars · 169 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

UCCL is a high-performance GPU communication library written in C++ that provides collectives (as a drop-in NCCL/RCCL replacement), P2P transfers (KV cache, RL weight transfer), and expert-parallelism communication. It emphasizes flexibility for evolving ML workloads and portability across heterogeneous GPUs and network vendors (NVIDIA, AMD, Broadcom) using RDMA, InfiniBand, RoCE, and EFA.

Use cases

• replace NCCL for faster AllReduce in multi-node LLM training
• transfer KV cache between GPUs for LLM inference disaggregation
• sync RL training weights across GPU nodes
• run expert-parallel MoE communication across NVIDIA and AMD GPUs
• communicate over RDMA, InfiniBand, RoCE, or AWS EFA networks
• speed up distributed training on cloud instances with commodity NICs

When to choose

• you need faster collective communication than NCCL/RCCL provides
• your ML workloads span heterogeneous GPU vendors or network hardware
• you build LLM serving systems needing fast KV cache or weight transfer
• you run MoE models requiring efficient expert-parallel communication

When to avoid

• you only need simple single-node GPU communication
• your stack depends strictly on NCCL-specific extensions
• you need a battle-tested, widely-deployed communication library rather than a newer research-driven one

Facets

library · maturity active

networking gpu-computing machine-learning llm-training llm-inference machine-learning gpu-computing microservices large-language-models networking performance cpp cross-platform gpu-communication nccl-replacement rdma allreduce collective-communication kv-cache-transfer expert-parallelism moe rdma-transport p2p-transfer linux gpu cuda

Member repositories

For agents

markdown · JSON · MCP: product_card(name="uccl-project/uccl")