tt-bio bundles third-party code under tt_bio/_vendor/ and tt_bio/boltzgen/.
Each component keeps its upstream license alongside it. The core Boltz-2 / Boltz-1
code that tt-bio is built on (parsing, tokenization, featurization, the model) is
MIT-licensed and credited in the top-level LICENSE.

1. tt_bio/_vendor/esm/ — ESMFold2 host-side reference: sequence/atom
   featurization, MSA handling, and mmCIF assembly (no neural compute; that runs
   on Tenstorrent via tt_bio.esmc / tt_bio.esmfold2).
   Source:        github.com/Biohub/esm @ b6b0e88
   License:       MIT — Copyright 2026 Chan Zuckerberg Biohub, Inc.
                  (tt_bio/_vendor/esm/LICENSE)
   Modifications: only the files on the ESMFold2 fold path are included; absolute
                  `esm.` imports rewritten to `tt_bio._vendor.esm.`.

2. tt_bio/_vendor/esmfold2_hf/ — the ESMFold2 model definition for HuggingFace
   `transformers` (runs on top of the stock `transformers` wheel).
   Source:        Biohub `transformers` fork (based on transformers 4.57.6),
                  src/transformers/models/esmfold2
   License:       Apache-2.0 — Copyright The HuggingFace team and Chan Zuckerberg
                  Biohub, Inc. (tt_bio/_vendor/esmfold2_hf/LICENSE)
   Modifications: the three inference files only; transformers-core relative
                  imports rewritten to absolute `transformers.`.

3. tt_bio/boltzgen/ — BoltzGen binder-design source, stripped to Tenstorrent-only
   inference (training code, PyTorch Lightning, and Hydra removed); the heavy
   compute runs on Tenstorrent via tt_bio.boltzgen.adapter.
   Source:        github.com/HannesStark/boltzgen
   License:       MIT — Copyright (c) 2025 Hannes Stärk (tt_bio/boltzgen/LICENSE).
                  Some modules carry their own upstream headers (lucidrains/
                  alphafold3-pytorch, MIT; PyTorch3D, BSD) — kept in-file.
   Modifications: inference-only subset; Lightning/Hydra dependencies eliminated.

4. tt_bio/protenix.py, protenix_data.py, protenix_weights.py — Protenix-v2
   (AlphaFold3-family) support. This is an INDEPENDENT ttnn reimplementation for
   Tenstorrent; no upstream Protenix source is vendored. The architecture and the
   model weights (downloaded on first use from the Hugging Face mirror, not
   redistributed in this repo) are from ByteDance's Protenix.
   Source:        github.com/bytedance/Protenix
   License:       Apache-2.0 — Copyright ByteDance (code and model parameters).

5. tt_bio/_vendor/openfold3/ — OpenFold3 host-side data pipeline: the query
   schema, CCD/ligand lookup, tokenization, and structure / reference-conformer /
   MSA / template featurization (no neural compute; that runs on Tenstorrent via
   tt_bio.openfold3*).
   Source:        github.com/aqlaboratory/openfold-3, PyPI `openfold3` 0.4.3
   License:       Apache-2.0 — Copyright 2026 AlQuraishi Laboratory
                  (tt_bio/_vendor/openfold3/LICENSE)
   Modifications: only the query-to-features path is included; the Lightning
                  Dataset/DataModule framework, the LMDB training dataset caches,
                  and the S3 template-cache build pipeline are dropped, and the
                  memory_profiler import was made lazy.

6. tt_bio/openfold3*.py — OpenFold3 (AlphaFold3-family) on-device model. Like
   Protenix-v2 this is an INDEPENDENT ttnn reimplementation; no upstream compute
   code is vendored (only the data pipeline above). The weights
   (of3-p2-155k.pt, the OpenFold3 preview2 checkpoint) are the OpenFold
   consortium's ungated public parameter release; tt-bio neither redistributes
   nor downloads them — set OF3_CKPT or place the file at ~/.boltz/.
   Source:        github.com/aqlaboratory/openfold-3, weights from
                  s3://openfold3-data/openfold3-parameters/of3-p2-155k.pt
   License:       Apache-2.0 — Copyright 2026 AlQuraishi Laboratory. Upstream
                  states the project is free for academic and commercial use and
                  publishes no separate parameter license.

7. tt_bio/_vendor/atomworks/ — AtomWorks: the AF3-style structure-processing and
   transform pipeline RF3 featurizes with (CCD lookup, bond perception,
   atomization, chirality perception, MSA pairing, templating). Host-side only.
   Source:        github.com/RosettaCommons/atomworks, PyPI `atomworks` 2.2.1
   License:       BSD-3-Clause — Copyright (c) 2025, Institute for Protein Design,
                  University of Washington (tt_bio/_vendor/atomworks/LICENSE)
   Modifications: absolute imports repointed at the vendor namespace, and two
                  Python 3.11-only imports replaced so the package runs on 3.10,
                  which is tt-bio's deployed runtime: `enum.StrEnum` via
                  tt_bio/_vendor/atomworks/_compat.py, and `typing.Never` via
                  typing_extensions. Verified bit-exact against the upstream 3.12
                  behaviour; see scripts/rf3_port/.

8. tt_bio/_vendor/rf3/ and tt_bio/_vendor/foundry/ — the RF3-specific host
   transforms (ground-truth templating, cyclic chains, the AF3 pipeline
   definition) and three helpers they call. No model or training code.
   Source:        github.com/RosettaCommons/foundry, models/rf3 and src/foundry
   License:       BSD-3-Clause — Copyright (c) 2025, Institute for Protein Design,
                  University of Washington (tt_bio/_vendor/rf3/LICENSE)
   Modifications: only the featurization subset is included. `foundry/__init__.py`,
                  a .env / beartype / jaxtyping / cuEquivariance bootstrap, and
                  `foundry.utils.ddp`, which is Lightning-based, are replaced by
                  minimal stubs, so no Lightning, hydra or environs dependency is
                  pulled in.
