Struct StructureTokenEncoder 

pub struct StructureTokenEncoder { /* private fields */ }

Encodes backbone coordinates into discrete structure tokens via a 2-layer geometric mini-transformer followed by nearest-neighbour VQ codebook lookup.

Weight layout (in encoder checkpoint):

• encoder.blocks.* — transformer
• pre_vq_proj.weight — (d_codebook, enc_d_model)
• codebook.embeddings— (n_codes, d_codebook)

Implementations

impl StructureTokenEncoder

pub fn load(vb: VarBuilder<'_>, config: VqVaeConfig) -> Result<Self>

Load from esm3_structure_encoder_v0.pth.

Every prefix here was read off the real checkpoint rather than inferred from the reference source: it roots at transformer (not encoder), and pre_vq_proj carries a bias. Both were wrong before, and both would have failed loudly — unlike the missing relative position embedding, which simply left the encoder computing from zeros.

pub fn encode( &self, coords: &Tensor, sequence_id: Option<&Tensor>, ) -> Result<Tensor>

Encode backbone coordinates into structure tokens.

• coords — (B, L, 3, 3) backbone (N, CA, C) positions.
• sequence_id — optional (B, L) bin-packing ids.

Returns (B, L) u32 structure token indices.

Why this is not a sequence transformer

This encoder is local. It does not run the transformer over the chain; it runs it over each residue’s 16 nearest neighbours in space, as (B*L, 16, d_model). The earlier port ran a full-sequence pass from a zero hidden state, which is a different computation entirely — the weights would load (once the prefixes were right) and the output would be quietly meaningless.