Hi dear authors,
Thanks a lot for you work!
Your D64 version tokenizer seems quite powerful to me (good reconstruction PSNR),
but none of your example scripts (interactive_infer.ipynb, or its 8B/8B 512x512 version) actually uses it.
Naively loading the D64 tokenizer for the 2B/8B/8B 512x512 model will result in a parameter shape mismatch error.
As a result I would like to ask if any of your models (2B, 8B or the future 20B?) can actually use the D64 tokenizer, and how.
Thanks a lot for your help!
Hi dear authors,
Thanks a lot for you work!
Your D64 version tokenizer seems quite powerful to me (good reconstruction PSNR),
but none of your example scripts (interactive_infer.ipynb, or its 8B/8B 512x512 version) actually uses it.
Naively loading the D64 tokenizer for the 2B/8B/8B 512x512 model will result in a parameter shape mismatch error.
As a result I would like to ask if any of your models (2B, 8B or the future 20B?) can actually use the D64 tokenizer, and how.
Thanks a lot for your help!