Synthyra/ESMFold2

This checkpoint packages the FastPLMs ESMFold2 implementation.

Accepted inputs are raw amino-acid sequences or typed molecular-complex specifications; low-level forward accepts prepared feature tensors. Supported Transformers entry points are AutoConfig, AutoModel.

Capabilities

Feature	Status
Sequence classification	Unavailable: no advertised AutoClass
Token classification	Unavailable: no advertised AutoClass
PEFT fine-tuning	Supported pattern: attach LoRA to the pretrained model
Embeddings	Special: ESMC state mixture to 256-wide residue embeddings
Test-time training	Special: opt-in folding TTT on the ESMC backbone
Attention variants	Supported: eager, sdpa, flex_attention
Compliance	Declared: exact release evidence is required

A supported interface is not a pretrained downstream predictor. Classification heads start untrained, and declared compliance metadata is not a claim that an arbitrary local build passed its release gate.

Install and platform requirements

Install the direct dependencies published with this model:

python -m pip install -r \
  "https://huggingface.co/Synthyra/ESMFold2/resolve/main/requirements.txt"

The FastPLMs implementation itself is embedded in the model repository and loaded by Transformers through trust_remote_code=True.

Python 3.11-3.14, PyTorch 2.13, and Transformers 5.13 are required. The artifact requirements include the direct structure dependencies. The published execution contract requires a CUDA device. The current validated release target is the exact NVIDIA GH200 on Linux aarch64; Linux x86-64, CPU-only, Windows, and macOS structure runs are not current release evidence. The Hub quick start below requires network access on first download. For an air-gapped run, first build the manifest-pinned local artifact and use the offline form shown in the example.

Quick start

from transformers import AutoModel

model_id = "Synthyra/ESMFold2"
model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    attn_implementation="sdpa",
).eval()

For offline validation, replace model_id with the manifest-built dist/hub/ESMFold2 path and pass local_files_only=True.

Attention and compliance

The quick start selects sdpa explicitly. Declared variants are eager, sdpa, flex_attention. An unavailable requested backend raises instead of silently switching implementations. output_attentions=True may use the documented, one-call eager fallback solely to materialize attention tensors; the configured backend remains unchanged.

This family declares the compliance tier. Release evidence binds the exact checkpoint, backend, dtype, hardware, inputs, and reference revision.

PEFT fine-tuning

Install the direct training dependencies, then attach LoRA to the loaded checkpoint:

python -m pip install "datasets>=4.8,<5" "peft>=0.19,<0.20"

from peft import LoraConfig, get_peft_model

peft_model = get_peft_model(
    model,
    LoraConfig(
        r=8,
        lora_alpha=16,
        target_modules="all-linear",
    ),
)

This checkpoint has no advertised classifier. Supply the task-specific objective and preserve any new head through modules_to_save. All FastPLMs checkpoints follow the Transformers PreTrainedModel contract and can be adapted with PEFT. The ESM2-specific shipped CLI is an example, not a support boundary. Record the target modules, base revision, data identity, and trainable parameter scope.

Alignment-conditioning contract

This is a full 48-block ESMFold2 checkpoint. It supports both single-sequence inference and optional MSA-conditioned inference. Typed multichain and multimolecule inputs may attach an MSA to each applicable protein chain.

Protein folding

The single-protein helper returns typed structure and confidence outputs:

result = model.fold_protein(
    "MSTNPKPQRKTKRNT",
    num_loops=1,
    num_sampling_steps=200,
    num_diffusion_samples=1,
    seed=7,
)
pdb_text = model.result_to_pdb(result)
cif_text = model.result_to_cif(result)
print(result.ptm, result.plddt.mean().item())

No target structure is required. For complexes, construct the input from the types exposed by the loaded artifact:

types = model.input_types
complex_input = types.StructurePredictionInput(
    sequences=[
        types.ProteinInput(id="A", sequence="MSTNPKPQRKTKRNT"),
        types.ProteinInput(id="B", sequence="MKTIIALSYIFCLVFA"),
        types.DNAInput(id="C", sequence="ATGC"),
        types.LigandInput(id="L", smiles="O"),
    ]
)
complex_result = model.fold(
    complex_input,
    num_loops=1,
    num_sampling_steps=200,
    seed=7,
)
print(complex_result.ptm, complex_result.plddt.mean().item())

The typed interface also supports RNA, protein MSAs, modifications, and covalent bonds. The public schema recognizes PocketConditioning and DistogramConditioning, but the pinned official forward consumes neither. Its feature builder hard-codes a zero pocket feature and constructs distogram tensors that the released model ignores. FastPLMs therefore rejects non-null pocket and distogram conditioning instead of silently ignoring scientific inputs. Prepared ref_pos values are component reference geometries created during featurization, not target coordinates. Predicted coordinates and confidence scores are outputs and do not establish biochemical activity.

Learned representation and ESMC precision

ESMFold2 applies its learned state mixture and projection as H: (b, l, 81, 2560) -> Z: (b, l, 256). Retrieve Z through the public embedding API:

representations = model.embed_dataset(
    ["MSTNPKPQRKTKRNT", "MKTIIALSYIFCLVFA"],
    batch_size=2,
    full_embeddings=True,
)
print(representations[0].tensor.shape)  # (sequence_length, 256)

model.embed_dataset(..., full_embeddings=True) returns one (l, 256) residue tensor per single-chain input. It rejects complexes, ligands, MSAs, chain-separated inputs, cls, and parti in the embedding path.

Set esmc_precision to auto, bf16, fp32, or fp8 when loading. auto always resolves to BF16. Explicit FP8 is experimental, inference-only, and strict:

model.reload_esmc(precision="fp8", device="cuda:0")
print(model.esmc_precision_status)

FP8 raises when the validated CUDA and Transformer Engine path is unavailable. Canonical BF16 weights are retained, and transient quantization state is never serialized.

The ESMC backbone uses SDPA as the recommended highest-fidelity path. Flex Attention is supported and non-experimental but can be numerically divergent; ESMFold2 does not advertise FlashAttention for the folding interface.

Backend	Support	Measurement status
sdpa	Recommended fidelity path	Pending release measurement
eager	Supported	Pending release measurement
flex_attention	Supported, numerically divergent	Pending release measurement

Detailed backend measurements, release guardrails, and the GH200 package compatibility exception are maintained in the attention backend guide and release evidence manifest.

Hash-pinned CCD runtime asset

Structure preparation requires ccd.pkl from biohub/ESMFold2. The manifest pins its 417,306,584-byte size and SHA-256 9ff44b1927c6b9198e38ffe0928706827a09a350c15530beeeabebfa88038fc5 under MIT terms. This is a trusted-deserialization boundary: FastPLMs only allows the exact manifest repository/revision snapshot link to resolve within that repository's contained blob directory; user-supplied asset and cache_dir symlinks are rejected. The loader creates a private temporary snapshot, verifies its size and SHA-256, and unpickles only that loader-owned snapshot, closing path-replacement and in-place source-write races. Offline execution requires the exact cache object and never downloads a replacement.

Optional folding TTT

The standard and Fast checkpoints expose opt-in folding TTT on their ESMC backbone:

adapted = model.fold_protein_ttt(
    "MSTNPKPQRKTKRNT",
    num_loops=1,
    num_sampling_steps=50,
    seed=7,
    ttt_config={"steps": 3, "batch_size": 1, "seed": 7},
)
print(adapted.ttt_metrics)

Entering a gradient-enabled path reloads canonical BF16 ESMC weights. TTT adds latency and memory, can worsen a prediction, and does not calibrate confidence or establish biological validity. Folding TTT is result-scoped: its transient ESMC adapter modules are excluded from checkpoint state, so it is not a generic save_pretrained adapter-persistence path.

Runtime contract

• Public input: Raw amino-acid sequences or typed molecular-complex specifications; low-level forward accepts prepared feature tensors
• Advertised AutoClasses: AutoConfig, AutoModel
• AutoClass weight status: AutoConfig = FastPLMs extension, AutoModel = pretrained
• Attention implementations: eager, sdpa, flex_attention
• Precision policies: auto, fp32, bf16, fp8 (experimental)
• BF16 execution: fp32_parameters_autocast
• Generation contract: not_applicable
• Artifact dependency set: core + structure
• Weight publication allowed: true
• Weight license status: resolved
• Redistributable: true
• Complete weight publication required: false

Release record

• FastPLMs weights: Synthyra/ESMFold2
• Runtime revision: recorded separately in the built artifact and published commit
• Source-tree and runtime-bundle SHA-256: recorded in provenance.json
• Official checkpoint: biohub/ESMFold2
• Artifact source: fast
• State transform: identity
• Pinned upstreams: biohub-esm, biohub-transformers, protein-ttt
• Release tiers: check, compliance, structure, feature, artifact, benchmark
• Unresolved required file identities: 0

provenance.json records exact file identities, conversion, source revisions, legal texts, schema, and attestations. A nonzero unresolved count blocks release.

Validation boundary

Declared tiers compare applicable configuration, tokenizer behavior, state, and representative inference with the pinned reference. Metadata alone does not claim a build passed, a backend is faster, or an output is biologically valid.

License

Checkpoint terms: MIT. The Hub model-card identifier is mit. Applicable source licenses, notices, attribution, and conversion records are distributed with the local artifact. Review them before use.