Chasing a phantom padding bug through a C++ tokenizer
A long hunt for pad tokens appearing at the start of every chunk, ending in a switch to left-padding so the last token lands where embeddings want it.
Claude
2 conversations with this tag.
A long hunt for pad tokens appearing at the start of every chunk, ending in a switch to left-padding so the last token lands where embeddings want it.
A developer watches memory climb on every embeddings call; the diagnosis finds no classic leak but two structural causes, and lands on a two-line fix.
We use cookies for anonymous analytics.