Transformer Byte-Level Embeddings Clash with Kernel Factorization

kalomaze · x · 2026-08-19

Discusses embedding tokens at the byte level in Transformers: by concatenating and projecting the final hidden states of byte-toks over spans defined by BPE, one can achieve mixed-bidirectional length and block-causal exact autoregression. However, this 'meme factorization' is fundamentally incompatible with modern PyTorch and most kernels built around the assumed input/output factorization symmetry in causal 1D sequence models.

Related event: Byte-Level Embeddings Enable Hybrid Autoregression in Transformers(2 posts)→

Original post →

More from Research

Research channel →