DAK: Direct-Access-Enabled GPU Memory Offloading with Optimal Efficiency for LLM Inference

Collection of recent systems research arXiv announcements (Apr 30 2026) focused on large-model inference scalability, memory-tiering, and high-performance kernels. Key contributions: DAK — direct GPU access to remote memory using Tensor Memory Accelerator (TMA) and SMEM for efficient LLM weight/KV offload; DUAL-BLADE — dual-path NVMe-direct KV-cache offloading that maps KV tensors to contiguous LBAs to bypass filesystem/page cache; pipelined sharding — CPU/GPU hybrid scheduling for VRAM-constrained client xLM inference; SplitFT — adaptive federated split learning for LLM fine-tuning with cut‑层