NCCL EP: Towards a Unified Expert Parallel Communication API for NCCL

This collection contains research that touches multiple infrastructure and application-layer technologies with security implications: a new NCCL-based Expert Parallel (NCCL EP) library enabling GPU-initiated RDMA/NVLink MoE communication (expands device-level DMA and GPU-driver attack surface); critiques of SmartNIC offload architectures (Wave) that identify a Forward-In-Time-Only (FITO) communication model producing stale-decision/atomicity windows and advocating bilateral atomic primitives (implications for race/TOCTOU and latency-exploitable vulnerabilities); trusted-parallel BFT work that,