Checkpoint `~/.boltz/boltz2_conf.ckpt` missing `pairformer_module.layers.*.attention.norm_s.*` keys vs latest `main`

[Peter-obi]

Hi Boltz team,

I’m seeing a model/checkpoint mismatch when loading Boltz2 with the current main code.

Environment

• OS: Colab (Linux)
• Python: 3.12
• PyTorch: 2.9.0+cu128
• PyTorch Lightning: 2.5.0
• Boltz code: latest main (cloned from jwohlwend/boltz)
• Checkpoint: ~/.boltz/boltz2_conf.ckpt (downloaded by default flow)
• Size: ~2180.63 MB
• state_dict keys: 5102

Main warning:

• Found keys that are in the model state dict but not in the checkpoint: ['pairformer_module.layers.0.attention.norm_s.weight', ... 'pairformer_module.layers.63.attention.norm_s.bias']
• Confirmed:
  • norm_s keys in ckpt: 0

Question

Which code revision + checkpoint pair should be used for fully matched Boltz2 inference?

• Should users pin to a specific commit/tag?
• Is there an updated checkpoint that includes attention.norm_s weights?

[volgin]

This is fixed in boltz-community, a community-maintained fork with 20+ bug fixes, broader hardware compatibility (including Apple Silicon MPS), expanded test coverage (170+ tests including regression and determinism tests), and CI.

Fix: Added **kwargs to AtomDiffusion.__init__() so it gracefully absorbs extra checkpoint hparams like mse_rotational_alignment that are present in the released Boltz-2 weights but not in the current constructor signature.

Install: pip install "boltz-community @ git+https://github.com/Novel-Therapeutics/boltz-community.git"