推文

@sanpsonfox · 2026-10-17 20:57

BERT’s tokenizer was spending most of its encoding time before it even picked the tokens. In Hugging Face’s profiled English-text run, normalization took 88%; WordPiece took 8%. One fix was an ASCII fast path that avoids expensive Unicode lookups where they aren’t needed. The write-up reports up to 50× throughput versus tokenizers 0.23, with the latest changes still at PR stage when published. That’s tokenizer throughput, not a 50× faster model. https://huggingface.co/blog/huggingface/tokenmaxxing-bert

曝光 3 · 评论 0 · 点赞 0 · 书签 0