Without evoformer_attn/evoformer_attn.so and model.aux_heads.pae.linear.weight".

[forever-up]

Dear team，

I'm following the deployment tutorial and encountered two errors when running inference:

First Error:
After l installation, when running:

run_openfold predict --query_json=examples/example_inference_inputs/query_ubiquitin.json

I got the extension error:

[ERROR] Failed for query_id(s) query1: /home/yl/.cache/torch_extensions/py310_cu126/evoformer_attn/evoformer_attn.so
Second Error:
Based on community suggestions, I tried using the runner.yml config:

run_openfold predict --query_json=examples/example_inference_inputs/query_ubiquitin.json --runner_yaml=runner.yml
But now I'm getting a different error about unexpected keys in the state_dict:

text
Unexpected key(s) in state_dict: "model.aux_heads.pae.layer_norm.weight", "model.aux_heads.pae.layer_norm.bias", "model.aux_heads.pae.linear.weight"
My runner.yml content:

model_update:
preset:
- predict
- pae_enabled # <-- Required for OpenFold3 current model release which has PAE output
- low_mem
custom:
settings:
memory:
eval:
use_cueq_triangle_kernels: true
use_deepspeed_evo_attention: false
Environment:
Linux with A100 500G
Python 3.10
CUDA 12.6
Any insights, workarounds, or pointers to the correct configuration would be immensely helpful. Thank you so much for your time and support!

[jnwei]

This issue was also observed in #17. The issue is that the field should be presets instead of preset, e.g. the yaml file should be:

model_update:
presets:  # not preset!
    - predict
    - pae_enabled # <-- Required for OpenFold3 current model release which has PAE output
    - low_mem
custom:
    settings:
        memory:
             eval:
                 use_cueq_triangle_kernels: true
                 use_deepspeed_evo_attention: false

The documentation is incorrect here, leading to this issue. I apologize for the confusion.

I will fix the documentation. In the next release, we will change the pydantic models to error on missing fields / incorrect fioelds.

[forever-up]

@jnwei Thank you very much for your guidance. It has been running smoothly.

[jnwei]

Hi, we currently don't have support for MSA free predictions, i.e. running with both use_msa_servers=False and not providing separate multiple sequence alignments is not supported. You can find more information in our documentation here. We hope to add better support for the MSA-free prediction soon.

One request: Technically this is an issue separate from the model.aux_heads.pae.linear.weight issue you originally reported. If you have more questions / comments, please consider opening a new issue.