HYPIC: Accelerating Hybrid-Attention LLM Serving with Position-Independent Caching

Collection of new arXiv papers (announced 2026-07-03) focused on performance, scalability, and system design for large-model serving, training, and distributed storage. Key contributions: Hypic — first position-independent caching for hybrid-attention LLMs reducing TTFT 2.45x and improving peak throughput up to 2.0x via segment-cumulative transition operators and seam recomputation; Lynx — progressive split-stream KV transfer for long-context inference that starts decoding on an Anchor stream, improving TTFT vs. 8-bit KV by up to 1.43x while matching high-precision accuracy (up to +5.1%); SLFS