Tag
User seeking to optimize token generation speed on dual 3090s using DFlash and MTP speculative decoding techniques with llama.cpp and beellama, sharing their configuration and commands.
A user shares their experience setting up a dual 3090 GPU system to run the Qwen 3.6 27b model locally, achieving over 100 tokens/second after switching to Ubuntu and using the club-3090 tool with custom patches. They express excitement about the future of local AI.