I made a custom llama.cpp build optimized for 7900xtx (one or two). for qwen 3.8 next and 27B. includes optimizations for PciE x4 and tensor parallel. read inside! (no AI slop)

A developer created a custom llama.cpp build optimized for AMD 7900xtx GPUs, specifically for Qwen 3.8 next and 27B models. This build includes optimizations for PCIe x4 and tensor parallel processing. Key improvements include a +58.88% Flash increase from lazy PLE/load path, +19.95% from GPU MoE expert cache, and +14.32% Flash from MoE MMQ sizing RDNA3, demonstrating significant performance gains for these specific hardware and model configurations.