Cross-Platform Fused MoE Dispatch in Triton: Portable Expert Routing Without CUDA

This batch contains systems and ML infra papers with multiple security- and privacy-relevant implications. Key items: TritonMoE (a fully-Triton fused MoE dispatch kernel) increases cross-vendor portability and performance but broadens the runtime/kernel attack surface by moving a large fused kernel into portable, open-source Triton code. LoRA fusion and the prune-train-recover + LoRA-CR work highlight that naively fusing adapters from private edge devices fails and can produce degraded or conflicting parameter updates—raising confidentiality, integrity (poisoning/conflict), and provenance/atf-