Tag
A suitcase robot uses a real gas sensor to live-adjust LLM sampler parameters (temperature, top_p, top_k) based on smoke detection, causing the robot's speech to become more unpredictable and loop-free.
Implemented a top-k kernel from the kernel design section of the MSA paper, using exp-free comparison and warp-level tree merging with CUDA shuffles. The code is available on GitHub.