repoGitHubTrust 82 · PrimaryPublished 11h agoLive · 11h ago

uccl-project/uccl

UCCL is an efficient communication library for GPUs, covering collectives, P2P (e.g., KV cache transfer, RL weight transfer), and EP (e.g., GPU-driven)

Lineage graph

Paper → model → repo connections mined from source citations (Tier-1 exact match).

Implements

paperWattGPU: Predicting Inference Power and Latency on Unseen GPUs and LLMs paperGPU Parallelization Strategies for Forward and Backward Propagation in Shallow Neural Networks: A CUDA-Based Comparative Study

Covers

newsAnyone using TensTorrent gpus for your local ai? What's been your experience?newsBuild real agentic apps using CUGA: two dozen working examples on a lightweight harness

Related across the graph

newsBuild real agentic apps using CUGA: two dozen working examples on a lightweight harness paperWattGPU: Predicting Inference Power and Latency on Unseen GPUs and LLMs paperGPU Parallelization Strategies for Forward and Backward Propagation in Shallow Neural Networks: A CUDA-Based Comparative Study newsAnyone using TensTorrent gpus for your local ai? What's been your experience?

Topics

ai allreduce amd broadcom collective cuda gpu hpc kvcache llm