Toward Compiler World Models: Learning Latent Dynamics for Efficient Tensor Program Search
Researchers have developed a novel approach to optimize tensor programs for machine learning systems by modeling schedule evaluation as latent dynamics. This method, inspired by world models, uses a lightweight transition model to predict program states in a continuous latent space, avoiding costly code mutations and encodings. When implemented in TVM AutoScheduler, it significantly improved subgraph latency on GPUs and CPUs and accelerated full-model inference compared to existing methods, all within a reduced measurement budget. AI
IMPACT This research could lead to more efficient AI model training and inference by optimizing the underlying tensor computations.