Struct ESMCRunner 

pub struct ESMCRunner { /* private fields */ }

Wraps a loaded ESMC model for sequence embedding inference.

Implementations

impl ESMCRunner

pub fn from_pretrained(model: ESMCModels, device: Device) -> Result<Self>

Download weights from HuggingFace and load the model.

Handles the ESMCForMaskedLM wrapper prefix (esmc.) transparently: probes for esmc.embed.weight in the safetensors and sets the VarBuilder root to vb.pp("esmc") when found, otherwise uses the flat (unwrapped) layout.

pub fn from_pretrained_with( model: ESMCModels, opts: &LoadOptions, ) -> Result<Self>

Load with an explicit device and dtype.

ESMC6B is ~24 GB at F32 but ~12 GB at BF16 — its on-disk precision — so reduced precision is the difference between loadable and not on a 32 GB machine. See tests/test_plm_dtype_parity.rs for the numerical cost (ferritin-100.9).

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

Tokenize sequence and run a forward pass.

Returns per-residue embeddings with shape (1, L, d_model) where L includes the BOS and EOS tokens added by the tokenizer.

Trait Implementations

impl PlmRunner for ESMCRunner

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

Delegate to embed_sequence, returning per-residue embeddings (1, L, d_model).

fn special_tokens(&self) -> SpecialTokenLayout

ESMC::encode calls tokenize_sequence(.., true), which wraps the sequence in BOS and EOS.

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

Sequence logits (1, L + 2, vocab_size).

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

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

ESM-C’s sequence_id is a where_cond predicate — 1 keeps a key, 0 replaces its score with -inf — so the U8 mask goes in as-is. It is passed explicitly rather than relying on ESMC::forward’s fallback (tokens.ne(pad_id)), which would happen to agree here but ties correctness to the padding id rather than to the batch itself.

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.