Re-export 3b 8da4w with the right stop token, and correct the configs
#1
by msluszniak - opened
The published 3b 8da4w stops on <|endoftext|> (151643) where its bf16 sibling and both smaller sizes stop on <|im_end|> (151645). It traces to a copy error in the old recipe, so the file is re-exported to match.
All three configs also gain get_n_layers (it was null) and the real 2047 sequence dim (they declared 2048). The bf16 variants are correct as published and untouched.
The .pte was probed against its config and matches.