Incorrect Weights in Model Repositories
Hi there! 😊
I've noticed that the models seem to have the wrong weights, specifically the weights for Emu3-Chat. It appears that the weights in the following repositories are all the same:
BAAI/Emu3-Gen-hf
BAAI/Emu3-Chat-hf
BAAI/Emu3-Chat
I suspect there was an error during the migration from the old repository to the new *-hf versions.
Thanks in advance for looking into this!
@francescortu
hey! Thanks for reporting this. The hf weights should be mirrors of original weight and thus it is expected that Chat and Chat-hf share same weight with slightly different state dict keys
For the Gen-hf model, do you mean that the values of weights are same as for Chat-hf and is it causing bad generation quality?
Yes, exactly. The BAAI/Emu3-Gen-hf model has the same weights as the BAAI/Emu3-Chat-hf model—not just identical state dict keys (which is expected), but the actual weight tensors are the same. I suspect there was an error during the conversion, where the input used was from Emu3-Chat instead of Emu3-Gen
Very weird, oke, I can verify that quite later. Might have been an error indeed since the Gen-hf was updated recently