Module pretrained 

ESMC pretrained model loading.

Downloads weights from biohub/ESMC-{300M,600M,6B} on HuggingFace and wraps the ESMC model for sequence embedding.

Weight key layout (biohub/ESMC-* safetensors)

The weights are saved from ESMCForMaskedLM, a HuggingFace wrapper that nests the backbone under an esmc attribute. All backbone keys are therefore prefixed with esmc.:

Python (HF safetensors key)	Rust VarBuilder path
esmc.embed.weight	embed.weight
esmc.transformer.blocks.{i}.attn.layernorm_qkv.0.*	transformer.blocks.{i}.attn.layernorm_qkv.0.*
esmc.transformer.blocks.{i}.attn.layernorm_qkv.1.*	transformer.blocks.{i}.attn.layernorm_qkv.1.*
esmc.transformer.blocks.{i}.attn.out_proj.weight	transformer.blocks.{i}.attn.out_proj.weight
esmc.transformer.blocks.{i}.attn.q_ln.weight	transformer.blocks.{i}.attn.q_ln.weight
esmc.transformer.blocks.{i}.attn.k_ln.weight	transformer.blocks.{i}.attn.k_ln.weight
esmc.transformer.blocks.{i}.attn.rotary.*	transformer.blocks.{i}.attn.rotary.*
esmc.transformer.blocks.{i}.ffn.0.*	transformer.blocks.{i}.ffn.0.*
esmc.transformer.blocks.{i}.ffn.1.weight	transformer.blocks.{i}.ffn.1.weight
esmc.transformer.blocks.{i}.ffn.3.weight	transformer.blocks.{i}.ffn.3.weight
esmc.transformer.norm.weight	transformer.norm.weight
esmc.sequence_head.0.weight/bias	sequence_head.0.weight/bias
esmc.sequence_head.2.weight	sequence_head.2.weight
esmc.sequence_head.3.weight/bias	sequence_head.3.weight/bias

The from_pretrained loader auto-detects whether the esmc. prefix is present and sets the VarBuilder root accordingly, so the same Rust model code works against both the wrapped and unwrapped formats.

Structs

ESMCRunner

Wraps a loaded ESMC model for sequence embedding inference.

Enums

ESMCModels

Available ESMC model variants hosted on HuggingFace.