evo-if

Experimental Evo2-NAIAD inverse-folding fusion prototype.

This repository contains a Q-Former-style adapter that connects intermediate Evo2-7B hidden states to the NAIAD inverse-folding decoder:

Evo2-7B blocks.28.mlp.l3 hidden states -> Q-Former bridge -> NAIAD hidden states -> nucleotide logits

Important status: adapter.pt is an initialized prototype checkpoint. It is structurally valid and was smoke-tested for inference, but it has not been trained on DAS/CATH or any downstream benchmark yet. Do not interpret outputs as a measured improvement over NAIAD.

Files

• adapter.pt: Q-Former bridge checkpoint.
• naiad_s1836.pt: NAIAD pretrained checkpoint used by the prototype.
• irm_enhanced_diffusion_training.json: NAIAD config.
• evo2_naiad_qformer_fusion.py: smoke/inference script.
• smoke.json: full Evo2 hidden-state smoke run.
• smoke_load_adapter.json: same run after loading adapter.pt.

Evo2-7B weights are not duplicated here. Load them from arcinstitute/evo2_7b or provide a local evo2_7b.pt.

Verified Smoke Test

The adapter checkpoint was loaded on an RTX 3090-class GPU with Evo2-7B and NAIAD frozen:

{
  "evo2_encoder": "full",
  "evo2_layer_name": "blocks.28.mlp.l3",
  "evo2_hidden_shape": [1, 97, 4096],
  "naiad_hidden_shape": [1, 97, 128],
  "qformer_hidden_shape": [1, 16, 128],
  "logits_shape": [1, 97, 33],
  "trainable_parameters": 1121921,
  "frozen_parameters": 6584016337
}

Example

Run from a NAIAD checkout with ArcInstitute evo2 installed:

PYTHONPATH=/path/to/evo2:/path/to/NAIAD/pdbx:/path/to/NAIAD \
python evo2_naiad_qformer_fusion.py \
  --evo2_encoder full \
  --evo2_checkpoint /path/to/evo2_7b.pt \
  --evo2_model_name evo2_7b \
  --evo2_layer_name blocks.28.mlp.l3 \
  --adapter_checkpoint adapter.pt \
  --naiad_checkpoint naiad_s1836.pt \
  --naiad_config irm_enhanced_diffusion_training.json \
  --pdb_path /path/to/input.pdb \
  --output_json output.json

On non-Hopper GPUs without FlashAttention, the patched local Evo2 runtime can fall back to bf16 projections and PyTorch scaled dot-product attention.

Provenance

• Evo2-7B base model: arcinstitute/evo2_7b
• NAIAD source/checkpoint: StarLiu714/NAIAD
• Adapter architecture: local experimental Q-Former bridge.