Skip to content

Official float32 inference for seedvr2_ema_7b.pth #65

Description

@photoakos

Request: official float32 inference path for seedvr2_ema_7b.pth.

The public 7B apply script (projects/inference_seedvr2_7b.py) loads seedvr2_ema_7b.pth then runs torch.autocast("cuda", torch.bfloat16). configs_7b/main.yaml sets VAE dtype to bfloat16.

Will you publish an official lossless float32 inference config and command for the original seedvr2_ema_7b.pth (no dtype cast on weights or activations), including VAE ema_vae.pth in float32?

in simpler terms: I use the original seedvr2_ema_7b.pth.

Your public 7B script loads that file but then runs the calculations in bfloat16, not full 32-bit. The VAE is also set to bfloat16.

Please publish an official way to run seedvr2_ema_7b.pth and ema_vae.pth in full 32-bit, with no conversion of the weights or the math.

Thank you.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions