Flash-SemiCRF: Exact segment-level inference on 100K+ sequences with 100x memory reduction

Flash-SemiCRF replaces the edge potential tensor with a prefix-sum array, reducing memory by a factor proportional to segment length × label count. A streaming forward-backward pass with checkpoint-boundary normalization keeps working memory sublinear in sequence length. Zero-centered cumulative scores control numerical drift and induce an adaptive duration prior. The fused Triton kernel enables exact semi-CRF inference on genomic-scale sequences exceeding 100,000 positions, previously intractable.

• #semi-crf
• #sequence-labeling
• #triton
• #genomics
• #memory-efficient