expert-parallelism
Profile-guided layout optimization for llama.cpp GGUF MoE models — BOLT/PGO for LLM weights. Trace expert routing, rewrite the file so co-activated experts sit together, prefetch as merged reads.
