v1.2.5

Bug fixes

A review against upstream (Video-Rebirth/hyperflow) found 5 porting bugs — all fixed in this release:

• SwiGLU LoRA rows were reversed. Diffusers and ComfyUI store the gate/value halves in opposite order, so the feed-forward calculation was wrong throughout the model. Existing weight files are auto-corrected on load — no re-download needed — and the converter now writes the correct layout.
• First-step audio got the wrong endpoint. Video and audio both start at t=0 but need different r values; audio was receiving video's endpoint on step 1. Distinct (t, r) rows are now preserved through the blocks and output heads.
• Merge mode contaminated the endpoint embedder. In merge mode the endpoint branch read LoRA-merged weights, double-applying the base time LoRA. The two small time projections now stay in bypass even when merging.
• Automatic weight downloads failed when the destination folder didn't exist, and broke Hugging Face cache symlinks on Windows.
• Sparse-attention recipe mapping was off. start_percent=0.25 silently kept 3 steps dense instead of 2; corrected to 0.16 in docs and both example workflows (plus extra_tokens, sink conditioning, and min_tokens aligned with upstream).

Improvements

• Bypass QKV adapters no longer carry ~1.07 GiB of zero padding — bit-identical to upstream-style separate projections, and slightly faster at realistic sequence lengths.
• 57 tests pass, including 17 new regression tests covering all of the above; validated against the published weight files.

Update: pull latest and restart ComfyUI. Old weight files keep working — the loader applies the SwiGLU correction automatically.