fabric-lib: RDMA Point-to-Point Communication for LLM Systems

fabric-lib addresses a critical bottleneck in next-generation LLM architectures: the need for efficient, hardware-independent Point-to-Point (P2P) communication. Standard collective libraries like NCCL are optimized for all-reduce/all-gather patterns used in training, but they struggle with the sparse, asynchronous networking required for Mixture-of-Experts (MoE) or disaggregated KV cache management. With 0 stars and 4 forks, the project is currently in a nascent research phase, likely tied to its accompanying arXiv paper. Its defensibility is moderate; while RDMA programming is notoriously difficult and requires deep domain expertise (the moat), the project lacks the ecosystem gravity of established alternatives like UCX (Unified Communication X) or Microsoft's MSCCL. Platform domination risk is high because major cloud providers (AWS with Nitro, Google with Falcon/ICI) and hardware vendors (NVIDIA, Broadcom) are incentivized to provide their own optimized drivers for P2P communication to lock users into their fabric. If a project like vLLM or DeepSpeed adopts a specific P2P standard, fabric-lib would need to be the underlying provider to survive. Its best opportunity is to become the 'OpenSSL of RDMA' for AI inference engines, but it currently lacks the adoption velocity (0.0/hr) to suggest this is happening yet.