v0.2.3

[0.2.3] - 2026-07-09

Multi-card fanout parity for predict, a designability (scRMSD) verify script for tt-bio gen,
and tt-bio embed input/UX polish. No structure-model code changed vs 0.2.2 (tt_bio/boltz2.py,
protenix.py, esmfold2.py, tenstorrent.py are byte-identical) — only esmc.py and the CLI
(main.py) changed, so the release gate below is a confirmation run, not a re-verification.

Release gate (scripts/release_gate.py, examples/prot.yaml, 200 steps / 5 samples, seed 0):

model	CA-RMSD	TM	floor	result
Boltz-2	1.60 Å	0.931	≤3.0 Å / ≥0.75	PASS
ESMFold2	2.28 Å	0.832	≤4.0 Å / ≥0.65	PASS
ESMFold2-fast	1.74 Å	0.907	≤4.5 Å / ≥0.60	PASS
Protenix-v2	3.87 Å	0.706	≤6.0 Å / ≥0.50	PASS

Full test suite: 71 passed, 46 skipped (missing optional reference checkpoints/packages, same
gap as prior releases), 0 failed. No OOM: examples/615.yaml and examples/1303.yaml
(Boltz-2 --fast) completed cleanly; the full supported range up to examples/3233.yaml
(4-chain multimer + ligand) was already verified OOM-free on this same unchanged model code
(docs/boltz2-tt-vs-nvidia.md). No perf regression: Boltz-2 --fast warm e2e at L=615 is
43.4 s, matching the 0.2.2-era baseline exactly (same code path since before 0.2.2).

Added

• tt-bio predict --devices — alias for --device_ids (comma-separated card ids), matching tt-bio embed's flag name; --device_ids still works for back-compat.
• BoltzGen designability (scRMSD) verify script — scripts/boltzgen_designability.py harvests the self-consistency RMSD tt-bio gen already computes and summarizes/gates on it; see docs/boltzgen-designability.md.
• tt-bio embed --devices wall-clock scaling measured (docs/esmc-multicard-scaling.md) — real ~2x @ 4 cards for esmc-600m on large batches, but flat/worse for small batches and for esmc-6b beyond 2 cards (concurrent weight-load contention); README softened to match. Performance-only finding, no change to the (already bit-exact) sharding correctness.

Changed

• tt-bio embed input handling — DATA now also accepts a YAML {id: sequence} mapping or a bare sequence string (previously FASTA file/directory only), writes a manifest.json (model/pool/shapes/dtype + which output file holds each sequence) alongside the embeddings, and reports bad input as a one-line error instead of a raw traceback.