Skip to content

0.7.0

Choose a tag to compare

@nsakkine nsakkine released this 25 Sep 10:42
· 141 commits to main since this release
a712e90

What's Changed

  • Fix DiffusionOutput to not wrap videos=None into a list by @Arech8 in #771
  • fix(sd3,hunyuandit): drop callback pops for undefined negative embeds by @Anai-Guo in #776
  • fix(flux): drop callback pops for nonexistent negative embeds by @Anai-Guo in #775
  • Fix FA3 ring attention argument forwarding by @zjn20030811 in #774
  • [bugfix] take start time of all ranks into account for performance measurements by @Arech8 in #772
  • new MHA v4 features, bugs, fixes by @jcaraban in #779
  • Add determinism check by @Arech8 in #778
  • Enable MiniMax-H3 tile-parallel VAE decode behind --use_parallel_vae by @hakankiymaz-amd in #777
  • FP8 Ulysses All-to-All Communication by @tjkemp in #753
  • FastH3 VSA-H3 T2VA support by @hakankiymaz-amd in #780
  • feat: add MXFP6 GEMM and unified GEMM configuration by @ksikiric in #782
  • Extended FP8 comms support by @pds-amd in #783
  • restore torch<2.10 compute-communication overlap passes by @lauri9 in #784
  • [Fix] Recompiles on Minimax-H3 by @nsakkine in #781
  • fix: skip torchao-quantized linear weights in quantization paths by @pds-amd in #786
  • Extended Wan CFG parallel support by @pds-amd in #787
  • Add FastH3 V2 runner and V1 weight variant aliases by @jjuvonen-amd in #789
  • [Perf] MiniMax-H3: quantize decoded video on device instead of on the host by @jjuvonen-amd in #790
  • Add Lumina-Image-2.0 support with CFG and GQA-aware Ulysses by @lavinal712 in #785
  • feat: support CFG parallelism for Qwen Image Edit by @pds-amd in #788
  • [Perf] LTX2.5: balance decode ranks using tile volume based cost proxy by @lauri9 in #791
  • Enable FP8 communications with AITER MHA v4 by @ksikiric in #792
  • Update CODEOWNERS list by @avjves in #793
  • [Perf] FastH3: faster and compilable VSA-H3 attention by @jjuvonen-amd in #794
  • FastH3 V1 Dense model variant runner support by @jjuvonen-amd in #796
  • [AMD] Let dense AITER MHA v4 backends serve trailing-padded packed sequences by @jjuvonen-amd in #797

New Contributors

Full Changelog: 0.6.0...0.7.0