Skip to content

Version 0.12.3

Latest
Latest

Choose a tag to compare

View all tags
@kohya-ss kohya-ss released this 26 Oct 13:11
v0.12.3
cb70b96
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed

  • Avoid graph breaks in attention() under torch.compile by @kohya-ss in #2470
  • Add truncated_logit_normal timestep sampling (MOSS-Image pre-training schedule) and offset support for sigma sampling by @kohya-ss in #2471
  • Round timesteps to the nearest histogram bin in --show_timesteps by @kohya-ss in #2472
  • Register ConvRot INT8 Linear as torch.library custom ops so --compile covers it (MOSS-Image) by @kohya-ss in #2473
  • ConvRot INT8 backward: fused dequantization + rotation of the weight (MOSS-Image) by @kohya-ss in #2474
  • Extend the autotune configs of the ConvRot INT8 GEMM for Ampere / Blackwell GPUs by @kohya-ss in #2475
  • docs: update README for the 0.12.3 release by @kohya-ss in #2476
  • Version 0.12.3 by @kohya-ss in #2477

Full Changelog: v0.12.2...v0.12.3

Contributors

kohya-ss
Assets 2
Loading