Sunday let's GEEK on!! Oh boy, let's get after "nondeterminism", "kernel batch-invariant", and "floating-point non-associativity". Researchers at Thinking Machines Lab are looking to tackle "Defeating Nondeterminism in LLM Inference" in this recent blog post.
The post is trying to help other LLM providers address nondeterminism in inference systems and to encourage work to develop a full understanding of their systems. https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/ #AI #LLMs #ThinkingMachinesLabs #Consistency #Inference #Nondeterminism

