Will you be doing1. text encoders too?
Hey City
Are you planning to add
1.3b and text encoders
Text encoder (at least T5) most likely, assuming I can figure out how to map the lcpp tokenizer back to spiece.
Not too sure about 1.3B, though I think the arch should be the same so I'm sure someone will convert it if there's demand.
Thanks for this. Would love I2V-14B.
Working on it here though there's a few things I still need to figure out for it to actually be usable.
Main issue was the tokenizer (which needs to be in a special format embedded in the state dict for comfy to load, this part works now) + the fact that the current code randomly OOMs due to the token embedding tensor being massive compared to regular T5 (this part is harder to fix, probably need to add batching logic for dequantizing the token_embed tensor, or just offload it to CPU manually somehow).