EPIC: Efficient and Parallel Inference under CFG Constraints for Diffusion Language Models
EPIC improves decoding of diffusion language models under Context-Free Grammar (CFG) constraints by reducing overhead from repeated lexing, DFA construction, and sequential validation. It achieves this by using lexing memoization, DFA-free parsing with Earley-style checks, and relaxed subset selection for parallel token commitment. This results in significant speedup while maintaining the correctness of grammar adherence.