Tag
A user shares achieving 80 tok/s on a Qwen3.6 35B model with Q4_K_M quantization and 100k context on a single Nvidia P40 using TheTom's TurboQuant fork of llama.cpp, highlighting various optimizations.