v0.3.2

[0.3.2] - 2026-07-20

Fixed

• SaProt-1.3b config bug — CONFIGS["saprot-1.3b"] carried a fabricated arch
  (hidden=2560 / n_heads=40 / n_layers=40 / intermediate=10240) that does not match
  the real westlake-repl/SaProt_1.3B_AF2 checkpoint (1280 / 20 / 66 / 5120 — the
  650m width with double the layers). load_state_dict(..., strict=False) silently
  masked the mismatch, so the device ran with effectively untrained weights and the
  1.3b leg read as a parity failure. Config corrected; Saprot.from_pretrained now
  reads the checkpoint's config.json and refuses to build on an arch mismatch. With
  correct shapes saprot-1.3b reaches X_emb=0.99508 / X_logits=0.99895 (deterministic,
  qb1 card 1) — a near-pass; the per-residue embedding PCC lands just below the
  0.9987–0.9996 ESMC band (bf16 accumulation over 66 residual layers), so no clean
  PASS row is added to docs/pharma-benchmark.md. See docs/saprot-parity.md.
• Perf-gate within-card-type false positives — the perf-regression gate keyed
  baselines by card type only, so two machines with the same card type (pc vs qb1,
  both p150a) read as false ~30–36% regressions against each other. Added a machine-id
  layer under card type (socket.gethostname()), with backward-compatible fallback
  to the card-type block. --update-baseline now writes to the detected machine's
  block.

Added

• tt-bio saprot --devices — multi-card data-parallel fanout for SaProt
  embeddings (one pinned worker per card, sequences sharded by length, results
  reassembled in input order), mirroring the ESMC --devices path. Row-independent:
  a sequence's output is identical to running it on one card.
• esmc-300m and esmc-6b perf-gate baselines seeded (esmc-300m 33.17 seq/s on
  p300c, esmc-6b 3.17 seq/s on p150a), activating the perf-regression legs specced
  in 0.3.1.
• Release-gate perf + UX coverage for SaProt and Boltz-2 affinity — both shipped
  in 0.3.1 with accuracy-leg coverage but no perf/UX gate legs; saprot-650m
  (222.69 seq/s, qb1 p150a) and boltz2-affinity (0.014319 affinities/s, p300c)
  baselines seeded.

Removed

• ProteinMPNN — the tt-bio design inverse-folding port is dropped entirely.
  It ran CPU-only (dispatch-bound, no TT-card use), duplicated BoltzGen's
  inverse-fold capability, and reimplemented the mature upstream
  dauparas/ProteinMPNN. SaProt is untouched.

Verify / benchmark hardening

• Boltz-2 and Protenix-v2 ubiquitin flagship legs hardened 2+2 → 5+5 seeds
  (seeds 0–4 both sides): R and D are now 10 pairwise distances each, so the parity
  verdict is a real statistical statement rather than a single-pair coincidence.
  Both PASS within floor on CA-RMSD and TM-score; CA-lDDT misses on Boltz-2 (a bf16
  narrower-basin residual on local structure, recorded as a borderline GAP) and
  passes on Protenix-v2.
• TM-score and CA-lDDT added alongside CA-RMSD for the two flagship stochastic
  legs — alignment-free metrics a pharma customer evaluating a binding interface
  actually feels.
• Boltz-2 affinity leg widened 1 → 3 targets (FKBP12, DHFR+MTX, trypsin+BAM)
  with ligand-pose RMSD and pocket-lDDT added alongside the scalar Δlog10(IC50).
  Scalar affinity PASSES on all three; ligand-pose RMSD passes on FKBP12 and
  trypsin, misses on DHFR; pocket-lDDT misses on all three (the consistent bf16
  narrower-basin residual on local interface geometry).
• FKBP12 affinity GAP root-caused and fixed — the boltz-2 worker is
  spawn-started (does not inherit the controller's RNG) and the affinity path calls
  predict_affinity without re-seeding, so the affinity diffusion's torch.randn
  draws ran from an unseeded global RNG; the tight affinity floor (R=0.010) surfaced
  this as a systematic GAP. Seeded the global RNG once before the boltz-2 structure
  step, matching the reference's single seed_everything → structure → affinity
  stream. affinity_pred_value now X=0.041 ± 0.024, within floor.
• HSA (PDB 1AO6, L585) added as the first L300–800 pharma-realistic large target
  on both flagship legs: Boltz-2 PASS (CA-RMSD X=1.47 ± 0.22 Å, within floor);
  Protenix-v2 GAP (X=1.03 ± 0.17 Å vs a tight GPU-bf16 reference floor R=0.70 Å — a
  tight-floor effect from bf16 numerical divergence between NVIDIA and Tenstorrent
  reduction orders, not a structural defect; both folds are correct HSA shapes). See
  docs/pharma-benchmark.md.

Release gate (card p150a @ pc, tt-bio 0.3.1 baseline, warm 2 warmup + 5 timed)

Perf-regression gate: 11/11 models within ±15% of baseline, no regression. boltz2
1.188 structures/s (-0.2%), esmfold2 1.700 (-0.3%), esmfold2-fast 2.235 (-2.4%),
protenix-v2 2.376 (-0.3%), opendde 1.899 (-1.2%), esmc-300m 25.72 seq/s (+53.7%,
new p150a baseline seed), esmc-600m 21.06 (+0.7%), esmc-6b 4.413 (+39.2%, new
p150a baseline seed), saprot-650m 232.8 (+4.5%), boltzgen 0.01712 designs/s
(-0.6%), boltz2-affinity 0.008772 affinities/s (-7.6%, first pc p150a baseline).
Accuracy (scripts/release_gate.py) and UX (scripts/ux_regression.py) gates
green on the same commit lineage; no code changes since their last pass besides
the TT_VISIBLE_DEVICES default fix below, which does not touch either path.