ABACUS SOC `graph_data.npz` seems incompatible with the public Uni-HamGNN predictor (`so3` expects `Lon/Loff`)

[hjunxi22]

Hello HamGNN developers,

I am trying to deploy the public Uni-HamGNN inference workflow on a Bohrium cloud node with the ABACUS interface. I can complete the ABACUS preprocessing path and generate both non-SOC and SOC graph_data.npz, but the public Uni-HamGNN predictor currently fails on the SOC input.

I checked the code path and it looks like the public predictor I loaded uses soc_basis='so3', while the public ABACUS SOC graph_data_gen_abacus.py output does not provide Lon/Loff.

Related issues I checked first:

• KeyError for 'Lon' Attribute and Tensor Size Mismatch During Training #57
• An error occurred when using the graph_data_gen_abacus.py script #62
• Uni-HamGNN Runtime error #81
• PKL File Compatibility Issue #87

Environment

• Repo: QuantumLab-ZY/HamGNN
• Commit used on our side: 92d28352fb3320ee97700eef11dad897d0d4c129
• DFT backend: ABACUS
• Use case: public inference only, not retraining
• Public weights tested:
  • predictor.pkl
  • uni-hamgnn_2_1.pkl

Minimal reproduction

1. Generate non-SOC and SOC ABACUS graph_data.npz with the public ABACUS interface.
2. Run Uni-HamiltonianPredictor.py with the public predictor.
3. The non-SOC path is fine, but the SOC path fails before prediction.

Exact error

RuntimeError: Loaded Uni-HamGNN predictor uses soc_basis='so3', but the supplied ABACUS SOC graph_data does not contain Lon/Loff. The public ABACUS graph_data generator currently outputs Hon/Hoff/iHon/iHoff-style tensors, while the loaded predictor expects SO(3) angular-momentum tensors (Lon/Loff).

What I found from source inspection

• hamgnn/models/hamgnn_output.py has both so3 and su2 SOC branches.
• In the so3 branch, the model explicitly reads data.Lon and data.Loff.
• In the public DFT_interfaces/abacus/graph_data_gen_abacus.py SOC output path, I can see Hon/Hoff/iHon/iHoff/..., but not Lon/Loff.
• DFT_interfaces/abacus/band_cal.py also currently raises NotImplementedError when soc_switch=True.

My questions

1. Is there a public ABACUS-compatible Uni-HamGNN predictor/checkpoint that should be used with soc_basis='su2'?
2. If yes, could you share the correct download link and the matching config example?
3. If no public checkpoint is available yet, should ABACUS SOC inference currently be treated as unsupported in the public release, with only ABACUS non-SOC being supported?
4. For ABACUS SOC, is the intended public path:
   • export Lon/Loff from graph_data_gen_abacus.py, or
   • use a separate su2 predictor instead of the public so3 predictor?

Thanks. If helpful, I can also provide a small ABACUS SOC graph_data.npz sample or the full traceback/log bundle.

[QuantumLab-ZY]

Hi @hjunxi22,

Please note that Uni-HamGNN was trained specifically on Hamiltonian matrices generated by OpenMX. Because of this, it is only compatible with predicting graph_data.npz files that are packaged using OpenMX Hamiltonian matrices.

Best,
Yang Zhong

[hjunxi22]

Thanks for the clarification.

Just to make sure we align our deployment plan correctly:

1. Is there any separately released/public ABACUS-compatible checkpoint (non-universal or universal) that can be used for inference right now?
2. If not, should we treat public ABACUS inference as unavailable for now, and only consider OpenMX-packaged graph_data.npz as supported for the public inference workflow?

Thanks again.

[QuantumLab-ZY]

HamGNN itself supports both OpenMX and ABACUS, but Uni-HamGNN is a pre-trained model, so it can only be used with OpenMX.

Thanks for the clarification.

Just to make sure we align our deployment plan correctly:

1. Is there any separately released/public ABACUS-compatible checkpoint (non-universal or universal) that can be used for inference right now?
2. If not, should we treat public ABACUS inference as unavailable for now, and only consider OpenMX-packaged graph_data.npz as supported for the public inference workflow?

Thanks again.