Enum T5Models 

pub enum T5Models {
    ProtT5XlHalfUniref50Enc,
    AnkhBase,
    AnkhLarge,
}

Available T5-family models.

Variants

ProtT5XlHalfUniref50Enc

Rostlab/prot_t5_xl_half_uniref50-enc — the encoder-only half-precision build, which is the one almost everyone uses. The full encoder-decoder prot_t5_xl_uniref50 doubles the download for a decoder that embedding work never runs.

AnkhBase

ElnaggarLab/ankh-base — 48 encoder layers at d_model 768.

AnkhLarge

ElnaggarLab/ankh-large — 48 encoder layers at d_model 1536.

Implementations

impl T5Models

pub const fn registry_id(&self) -> &'static str

This variant’s registry id.

pub fn card(&self) -> &'static ModelCard

This variant’s ModelCard.

pub const fn published_dtype(&self) -> DType

The precision this checkpoint is published at.

ProtT5’s encoder-only build ships float16, so loading it at F32 would double 2.4 GB in memory for values that are F16 either way. Ankh ships float32, and rounding it down would lose real precision rather than save anything the checkpoint has. Either can be overridden through T5Runner::from_pretrained_with (ferritin-100.9).

pub fn model_info(&self) -> (WeightSource, &'static str, T5Config)

Where this variant’s weights live, plus its built-in fallback config.