Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPU

This feed bundles multiple 2026 arXiv papers covering systems and ML infrastructure: Xe-Forge (an LLM-driven multi-stage pipeline to optimize Triton kernels for Intel GPUs with hardware-in-the-loop verification); BSP-aware frameworks for robust Edge AI deployment on vendor-specific embedded platforms; a communication- and resource-aware split-federated LoRA fine-tuning method (ST-SFLora); characterization of agentic LLM workloads and implications for caching/KV-state management; Totoro+ (a decentralized, P2P federated-learning platform at large scale); GridPilot (real-time, PUE-aware grid‑flex