When training on a single large input (e.g. 1GB) whose profile supports chunking, we should sample chunks rather than the entire file. Currently, fail training unless we set `--total-size-mb`, whereas it should select a subset of chunks.
When training on a single large input (e.g. 1GB) whose profile supports chunking, we should sample chunks rather than the entire file.
Currently, fail training unless we set
--total-size-mb, whereas it should select a subset of chunks.