Struct ESM2Runner 

pub struct ESM2Runner { /* private fields */ }

Implementations

impl ESM2Runner

pub fn from_pretrained( modeltype: ESM2Models, device: Device, ) -> Result<ESM2Runner>

Load model from HuggingFace hub at F32, downloading config.json, tokenizer files, and weights.

Use from_pretrained_with to load at F16 or BF16 — T36_3B and T48_15B are not loadable at F32 on most machines (~12 GB and ~60 GB respectively).

pub fn load_model(modeltype: ESM2Models, device: Device) -> Result<ESM2Runner>

👎Deprecated since 0.4.0:

renamed to from_pretrained

Renamed to from_pretrained (ferritin-100.8).

pub fn load_model_with( modeltype: ESM2Models, opts: &LoadOptions, ) -> Result<ESM2Runner>

👎Deprecated since 0.4.0:

renamed to from_pretrained_with

Renamed to from_pretrained_with (ferritin-100.8).

pub fn from_pretrained_with( modeltype: ESM2Models, opts: &LoadOptions, ) -> Result<ESM2Runner>

Load model with an explicit device and dtype.

Reduced precision halves the memory footprint but changes the numerics; see tests/test_plm_dtype_parity.rs for the tolerance each dtype actually achieves against the Python reference (ferritin-100.9).

pub fn run_forward(&self, prot_sequence: &str) -> Result<ESM2Output>

pub fn predict_contacts(&self, prot_sequence: &str) -> Result<Tensor>

Predict residue-residue contact probabilities for a single protein sequence.

Returns a (seq_len, seq_len) contact probability matrix (BOS/EOS stripped, so dimensions equal the number of amino acids in prot_sequence).

pub fn decode_logits(&self, output: ESM2Output) -> Result<String>

pub fn get_pseudo_probabilities( &self, prot_sequence: &str, ) -> Result<Vec<PseudoProbability>>

Run ESM2 and return per-residue pseudo-probabilities for the 20 standard amino acids.

BOS and EOS tokens are stripped before softmax is applied. Only amino acid / position pairs with probability > 0.01 are included in the result.

Trait Implementations

impl PlmRunner for ESM2Runner

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

Run the ESM2 transformer and return per-residue embeddings (pre-LM-head).

Shape: (1, L, hidden_size) where L includes BOS and EOS tokens.

fn special_tokens(&self) -> SpecialTokenLayout

ESM-2 wraps sequences in <cls> and <eos>; encode_with_special adds them because the bundled tokenizer.json has no post-processor.

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

SaProt reads two characters per residue; stock ESM-2 reads one.

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

Masked-LM 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).

Padding is config.pad_token_id, which matters twice over: ESM-1b and ESM-1v number their absolute positions by counting non-pad tokens, and the (batch, seq_len) mask handed to ESM2::embed both keeps padded keys out of self-attention and supplies the per-row sequence length that token-dropout compensation divides by.

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 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.