v0.4.0

[0.4.0] - 2026-07-26

First release shipping RFdiffusion3 (tt-bio design) — an all-atom generative model that
designs new protein structures and the sequences that support them from a specification, instead
of folding a sequence you already have. Protein-binder design, motif scaffolding, and
nucleic-acid-binder design run end to end from a real input structure. The checkpoint downloads
itself from the Institute for Protein Design on first use, so no rc-foundry install is needed.
Multiple designs per specification share device forwards, and --devices fans a design set
across cards. See docs/rfd3-design.md.

Also in this release: OpenDDE no longer runs its diffusion in fp32 (a >60x slowdown it inherited
from a Protenix-v2 default), opening a card that another process already holds now fails with a
clear error instead of colliding, and transformers moves to >= 5.5.0, clearing three dependabot
advisories.

Release gate (Blackhole P150a on pc, tt-bio 0.4.0): host suite 111 passed / 49 skipped;
packaging guard 15/15 data files and 31/31 declared dependencies present in the wheel and sdist;
UX gate PASS on every shipped surface (live-progress advancement, strict mmCIF/npz parse, and
results/manifest shape for boltz2, esmfold2, esmfold2-fast, protenix-v2, opendde, opendde-abag,
esmc-600m, saprot-650m, boltzgen, boltz2-affinity).

Perf gate (scripts/perf_regression.py, trpcage 20 aa single-sequence, 1 recycle / 10 steps /
1 sample, warm 2+5, ±15% threshold):

model	metric	baseline	current	delta	result
boltz2	structures/s	1.19	1.15	-3.4%	PASS
esmfold2	structures/s	1.705	1.601	-6.1%	PASS
esmfold2-fast	structures/s	2.29	2.141	-6.5%	PASS
protenix-v2	structures/s	2.383	2.166	-9.1%	PASS
opendde	structures/s	1.922	1.785	-7.1%	PASS
esmc-300m	seq/s	16.74	25.26	+50.9%	PASS
esmc-600m	seq/s	20.92	20.78	-0.7%	PASS
esmc-6b	seq/s	3.171	4.363	+37.6%	PASS
saprot-650m	seq/s	222.7	237.9	+6.8%	PASS
boltzgen	designs/s	0.01723	0.01695	-1.6%	PASS
boltz2-affinity	affinities/s	0.009498	0.008148	-14.2%	PASS

No model regressed beyond the threshold. No OOM through the gate targets.

Parity gate (scripts/full_parity_gate.py, 21 legs): ESMC-300m/600m, SaProt-35m/650m,
ESMFold2 (4 targets, L20 to L129), OpenDDE-abag (global DockQ 0.853, fnat 0.932) and BoltzGen
(scRMSD pass-rate 100%, median 1.12 Å) all reproduce their recorded verdicts, as do five of the
six Boltz-2 affinity legs; the sixth, FKBP12+SB3 with MSA, improves on its recorded gap. The
FKBP12+SB3 no-MSA affinity scalar is now recorded GAP-evidenced under the integration-envelope
test on the same cross-backend bf16 floor already accepted for its MSA counterpart (see
docs/implementation-parity.md).

The eight envelope structure legs report BLOCKED-REF-REGEN-NEEDED rather than a verdict: the
CPU reference structures those legs compare against are not distributed with the repository, so
they cannot be reproduced from a clean checkout. Their recorded verdicts in
docs/implementation-parity.md are unchanged and remain the evidence of record; restoring the
reference set is tracked as follow-up work.

Added

• RFdiffusion3 (RFD3) — tt-bio design specs.json --from_pdb --out_dir ./designs. Designs
  are specified with a contig mini-language (fixed regions taken verbatim from the input
  structure, designed regions of fixed or randomized length, chain breaks, indexed and unindexed
  motifs, per-atom fixing). Protein-binder design, motif scaffolding, and nucleic-acid-binder
  design accept a real PDB input; small-molecule-binder, enzyme, and symmetric-oligomer design
  run on device and are value-parity-verified against a captured reference, but the host
  featurizer does not build their input from a PDB yet and raises NotImplementedError.
  An independent ttnn reimplementation — no upstream RosettaCommons code is vendored, only the
  BSD-3-Clause checkpoint is fetched.
• Batched multi-design generation for RFD3 — --num_designs N produces N designs per
  specification (noise seed --seed + i), sharing device forwards in batches of up to
  --batch_size (default 8, reduced automatically for larger atom counts). Batching is
  accuracy-free: the device forward is bit-identical across batch size, so a batched design
  reproduces its standalone run exactly (min trajectory PCC 1.000000, maxabs 0, at 200 timesteps
  and batch 8). Throughput depends on design size — on one Blackhole p150a at 200 timesteps,
  batch 8 is 1.59x batch 1 at 40 residues and 1.21x at 80 residues, and within a few percent of
  batch 1 above roughly 150 residues, where --devices is the parallelism that matters. The
  per-size numbers are in docs/rfd3-design.md.
• tt-bio design --devices 0,1,2,3 — fans the (specification × --num_designs) jobs across
  the listed cards, one pinned subprocess per card, the same data-parallel pattern tt-bio embed
  and tt-bio predict use.
• Physical-card lease at device open — every device acquisition takes an exclusive flock on
  a per-card lock file for as long as the card is open. A second process opening the same card
  waits up to TT_BIO_LEASE_TIMEOUT (120 s) and then fails with DeviceInUseError naming the
  holder, instead of colliding at the fd level. The lock is released by the kernel on any process
  death, so a killed or orphaned job never leaves a phantom claim.

Fixed

• OpenDDE ran its diffusion in fp32, >60x slower than it needs to be. OpenDDE reuses the
  Protenix-v2 diffusion stack, and Protenix-v2 defaults that stack to fp32 on device
  (PROTENIX_DIFFUSION_FP32_DEVICE=1, correct for Protenix-v2, where fp32 is what its own
  reference uses). OpenDDE silently inherited it, and fp32 on OpenDDE's atom-level tensors is
  catastrophically slow. OpenDDE now pins its own already-validated bf16 diffusion config
  explicitly instead of reading the env default. A real device fold of 1AHW is back to 539 s at
  DockQ 0.862, matching the 0.863 pre-regression baseline — accuracy-neutral, speed-only.
• --write_pae was silently a no-op for --model opendde / opendde-abag. The flag was
  parsed but never reached the OpenDDE prediction path, so no PAE was written. Now wired.
• --msa_db_path was silently ignored by opendde-abag. An offline paired-MSA run fell
  through to the network path and failed if the public ColabFold service was unreachable. The
  local paired-MSA database is now honored, so opendde-abag folds fully offline.

Changed

• transformers >= 5.5.0, huggingface_hub >= 1.5.0 (were ==4.57.6 and <1.0). Clears
  three dependabot advisories (Trainer, config.json, and LightGlue remote-code execution; the
  last is not reachable from tt-bio). The vendored ESMFold2 model runs on the stock transformers
  core, and the bump is CPU-level bit-exact on the same torch, weights, and seed.