TileFuse: A Fused Mixed-Precision Kernel Library for Efficient Quantized LLM Inference on AMD NPUs

This batch of arXiv papers (June 11, 2026) covers systems and ML-infrastructure advances: TileFuse describes a close-to-metal mixed-precision kernel library to run AWQ-style low-bit quantized LLM inference efficiently on AMD XDNA2 NPUs, improving GEMM/GEMV throughput and energy use; a concurrency-aware cost methodology and the vllm-cost-meter quantify large utilization-driven variance in effective $/M-token for LLM hosting and expose severe underutilization penalties; ForeMoE proposes foresight-driven micro-step load balancing for MoE during RL post-training to reduce expert-imbalance overhead