tt-bio

# Changelog

All notable changes to TT-Bio are recorded here. Versioning is [SemVer](https://semver.org);
releases are cut from a commit that has passed the on-hardware test suite (see `RELEASING.md`).

## [Unreleased]

## [0.26.1] - 2026-10-29

### Fixed
- **ESMFold2 no longer runs out of device memory on a worker that has folded many different inputs.** A worker keeps
  its chip open across jobs, and the compiled programs for every input shape it has folded stay in device memory.
  After about ten hours of mixed lengths a Wormhole chip had too little room left for a 722-residue fold, and JapanFold
  failed up to 65 % of such tasks with "Ran out of device memory". A fold that is refused memory while programs are
  cached now drops them and runs again; folds that fit are unchanged. Seen on JapanFold with 0.25.0 and 0.26.0.

## [0.26.0] - 2026-10-29

### Changed
- **ESMFold2 and ESMFold2-Fast normal mode on Wormhole run the token attention as one fused program.** A 512-residue
  fold on a Wormhole Galaxy chip drops about 0.85 s (ESMFold2 25.9 -> 25.0 s, ESMFold2-Fast 16.0 -> 15.1 s). Structures
  change at the precision level: graded over 44 paired folds per model in normal mode, accuracy and docking success
  unchanged. Blackhole and `--fast` already ran it.
- **BoltzGen normal mode on Wormhole conditions its diffusion steps ahead of time.** A 514-token design on a
  Galaxy chip at 1000 MHz drops from 43.3 to 40.9 s. Designs change at the precision level: graded over 64 paired
  designs, design pTM, ipTM and complex pLDDT unchanged within the seed-to-seed spread.

## [0.25.0] - 2026-10-28

### Changed
- **BoltzGen's token diffusion transformer does less attention work.** In normal mode its pair bias was padded to
  768 tokens for a 544-token design; it now pads to 576, and on Blackhole the attention runs as one fused program
  in both modes. A 514-token design drops about 1.2 s on a Wormhole Galaxy chip (46 -> 45 s) and on a Blackhole
  chip about 2 s in normal mode (22.9 -> 21 s) and 1.5 s with `--fast`. Designs change at the precision level:
  graded over 64 paired designs per arch in normal mode and 32 with `--fast`, confidence metrics and filter pass
  counts unchanged. Other models keep their padding.
- **BoltzGen normal mode on Blackhole replays the diffusion step from a device trace by default.** Output is
  identical, the 1831-residue size ceiling holds, and with the triangle-multiplication change above a 514-token
  design on a p300 chip drops from 25.2 to 23.0 s. `--no_diffusion_trace` restores the eager step. `--fast` replays it too, and
  computes each step's conditioning inside the trace instead of several steps ahead: a 514-token fast design on a
  p300 chip drops from 32.4 to 23.3 s on a host-bound machine.

## [0.24.0] - 2026-10-27

### Changed
- **RoseTTAFold3 on Wormhole runs the fast triangle and transition kernels in normal mode.** A 512-residue fold
  on a Galaxy chip drops from 55.2 to 38.1 s. Structures change at the precis