Struct AmplifyRunner 

pub struct AmplifyRunner { /* private fields */ }

Implementations

impl AmplifyRunner

pub fn from_pretrained( modeltype: AmplifyModels, device: Device, ) -> Result<AmplifyRunner>

Load model from HuggingFace hub at F32.

Use from_pretrained_with for F16 or BF16.

pub fn load_model( modeltype: AmplifyModels, device: Device, ) -> Result<AmplifyRunner>

👎Deprecated since 0.4.0:

renamed to from_pretrained

Renamed to from_pretrained (ferritin-100.8).

pub fn load_model_with( modeltype: AmplifyModels, opts: &LoadOptions, ) -> Result<AmplifyRunner>

👎Deprecated since 0.4.0:

renamed to from_pretrained_with

Renamed to from_pretrained_with (ferritin-100.8).

pub fn from_pretrained_with( modeltype: AmplifyModels, opts: &LoadOptions, ) -> Result<AmplifyRunner>

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 achievable tolerance (ferritin-100.9).

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

pub fn get_best_prediction( &self, prot_sequence: &str, ) -> Result<String, Box<dyn Error + Send + Sync>>

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

Per-residue pseudo-probabilities over the full AMPLIFY vocabulary.

position is the residue index: BOS and EOS rows are stripped first, so there is exactly one position per input residue. Previously these rows were included, which both over-counted positions by two and shifted every label by one — position 0 was BOS, not the first residue (ferritin-100.18).

Unlike crate::ESM2Runner::get_pseudo_probabilities, every amino acid is returned rather than only those above a probability threshold.

pub fn get_contact_map(&self, prot_sequence: &str) -> Result<Vec<ContactMap>>

Trait Implementations

impl PlmRunner for AmplifyRunner

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

Run AMPLIFY and return the last-layer hidden states as per-residue embeddings.

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

fn special_tokens(&self) -> SpecialTokenLayout

AMPLIFY’s tokenizer.json has a TemplateProcessing post-processor that adds <bos> (3) and <eos> (4), so encode(.., true) wraps the sequence.

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).

AMPLIFY’s pad_mask is additive, not 1 = real — it is added to the attention scores — so the (batch, seq_len) mask goes through additive_padding_mask in the model’s dtype before it is passed in. Handing forward a 0/1 mask instead raises no error and masks nothing: real keys would get +1.0 and pads +0.0, so an unpadded row shifts uniformly and cancels in softmax, while a padded row leaves its pad keys with the highest bias of all.

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.