Hi, when will you release the THUDM/VisionReward-Image weight?

[shencuifeng]

when I run inference-image.py, it has this error:
url = MODEL_URLS[name]
KeyError: 'THUDM/VisionReward-Image'

[Uranusxer]

The model has now been uploaded to: https://huggingface.co/THUDM/VisionReward-Image.

[KT313]

@Uranusxer is the model fp32 20B parameters or why are there 75GB of files? I can't find info on that on the HF page. And would it be possible to upload a bf16 (safetensors) version?

[anvilarth]

@KT313 i have the same experience, one have to download https://huggingface.co/THUDM/VisionReward-Image fully as folder to github repo, then convert to bf16 weight, but it seems uncomfortable to convert from SAT format

[Uranusxer]

Yes, we initially uploaded the model with fp32 parameters. We have now uploaded a version with bf16 parameters, which should be more efficient for your use case. You can access it via this link: https://huggingface.co/THUDM/VisionReward-Image-bf16 . Thank you for using our model! If you have any further questions, feel free to ask.