Struct ESM3Runner 

pub struct ESM3Runner { /* private fields */ }

Wraps a loaded ESM3 model for sequence embedding and multi-track inference.

HuggingFace access

The ESM3 model is gated. Before calling from_pretrained the user must:

1. Accept the Cambrian Non-Commercial license at https://huggingface.co/EvolutionaryScale/esm3-sm-open-v1
2. Run huggingface-cli login (or set HF_TOKEN env var)

Implementations

impl ESM3Runner

pub fn from_pretrained(variant: ESM3Models, device: Device) -> Result<Self>

Download the ESM3 weights from HuggingFace and load the model.

The .pth checkpoint is loaded directly via candle_core::pickle::PthTensors.

pub fn from_pretrained_with( variant: ESM3Models, opts: &LoadOptions, ) -> Result<Self>

Load with an explicit device and dtype (ferritin-100.9).

pub fn embed_sequence(&self, sequence: &str) -> Result<Tensor>

Tokenize sequence and return per-residue embeddings (1, L, d_model).

Only the sequence track is provided; all other tracks are None. Embeddings are the pre-norm transformer hidden states (raw activations).

Trait Implementations

impl PlmRunner for ESM3Runner

fn special_tokens(&self) -> SpecialTokenLayout

embed_sequence calls tokenize_sequence(.., true), which prepends SEQUENCE_BOS_TOKEN and appends SEQUENCE_EOS_TOKEN.

fn logits(&self, sequence: &str) -> Result<Tensor>

Sequence-track logits (1, L + 2, d_sequence_vocab).

Only the sequence track is supplied; the other tracks are None.

fn embed_batch(&self, sequences: &[&str]) -> Result<Tensor>

One batched forward pass over right-padded sequences (ferritin-100.12).

ESM-3 reuses ESM-C’s MultiHeadAttention, so sequence_id is the same where_cond predicate (1 keeps a key). Unlike ESM-C, ESM3::forward has no fallback that derives one from the tokens: passing None — which is what every single-sequence path here does — means no masking at all, so a batch must supply it or the padding silently changes the real residues’ embeddings.

fn embed(&self, sequence: &str) -> Result<Tensor>

Run a forward pass on sequence and return raw per-residue embeddings.

fn model_name(&self) -> &str

Model name / identifier string (e.g. “esm2”, “amplify”, “esmc”).

fn metadata(&self) -> ModelMetadata

Architecture dimensions, for sizing downstream layers.

fn device(&self) -> &Device

Device the model’s weights live on.

fn residue_count(&self, sequence: &str) -> usize

How many residues sequence encodes.

fn embed_residues(&self, sequence: &str) -> Result<Tensor>

Per-residue embeddings with special-token rows removed.

fn embed_residues_batch(&self, sequences: &[&str]) -> Result<Tensor>

Per-residue embeddings for a batch, special-token rows removed.