[QUESTION] evoformer_attention

[kkbiu]

Describe the bug
Hello, thank you very much for your contribution to the community. Just like #17 , I also encountered the issue of not being able to evoformer_attention. I saw a solution to disable this mechanism by specifying runner.yml, but I still want to try building it to improve the inference speed of OpenFold3-Pre.

Therefore, in light of my situation, I would like to ask if you have any solutions, and I would also like to ask if you could provide a more detailed explanation of the configuration requirements and setup instructions for the evoformer_attention section in the subsequent documents.

Once again, I would like to express my deep gratitude for the contributions made by your team.

To Reproduce
Regular operation example：
run_openfold predict --query_json=~/software/openfold3/examples/example_inference_inputs/query_ubiquitin.json

Stack trace
Timestamp: 2025-11-29 08:51:09
Query ID(s): ubiquitin
Error Type: ImportError
Error Message: /.../.cache/torch_extensions/evoformer_attn/evoformer_attn.so: cannot open shared object file: No such file or directory

File "/ocean/projects/bio240017p/lnankai/tools/conda/anaconda3/envs/openfold3-pre/lib/python3.10/site-packages/torch/utils/cpp_extension.py", line 2634, in _import_module_from_library
module = importlib.util.module_from_spec(spec)
File "", line 571, in module_from_spec
File "", line 1176, in create_module
File "", line 241, in _call_with_frames_removed
ImportError: /ocean/projects/bio240017p/lnankai/.cache/torch_extensions/evoformer_attn/evoformer_attn.so: cannot open shared object file: No such file or directory

Configuration (please complete the following information):

• H100 80G
• Python version (if relevant): Both 3.10 and 3.13 have been tried
• Installation path : conda

Additional context
CUDA: 12.4.0 / 12.6.1

OpenFold Consortium Members⭐️: Please make use of the OpenFold Consortium Member label on your issue report for expedited support!

[jnwei]

Hi @kkbiu, with an H100 80G it should be possible to build DeepSpeed evoformer attention

A few troubleshooting questions:

• Is the /ocean/projects/bio240017p/lnankai/.cache/torch_extensions/evoformer_attn/ directory actually non-existent? If it is missing, it suggests there was an installation error with evoformer attention. If it is present, then there is instead a linking issue.
• Is there an error message that occurs earlier in the output that relates to the failed installation of Deepspeed evoformer attention? Perhaps something that looks like this message?

L/.../jnwei/miniforge3/envs/openfold_env/lib/python3.9/site-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda_cu -ltorch_cuda_cpp -ltorch -ltorch_python -L/opt/nvidia/hpc_sdk/Linux_x86_64/22.7/cuda/11.7/lib64 -lcudart -o evoformer_attn.so
FAILED: evoformer_attn.so
c++ attention.o attention_back.cuda.o attention_cu.cuda.o -shared -lcurand -L/.../jnwei/miniforge3/envs/openfold_env/lib/python3.9/site-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda_cu -ltorch_cuda_cpp -ltorch -ltorch_python -L/opt/nvidia/hpc_sdk/Linux_x86_64/22.7/cuda/11.7/lib64 -lcudart -o evoformer_attn.so
/usr/bin/ld: cannot find -lcurand: No such file or directory
collect2: error: ld returned 1 exit status
ninja: build stopped: subcommand failed.

In the example message above, the error cannot find -lcurand: No such file or directory suggests that the cuda libraries were not properly linked.

A few suggestions that may help, without knowing more specifics of the issue:

• Try setting PATH variables with export LIBRARY_PATH=$CONDA_PREFIX/lib:$LIBRARY_PATH && export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
• Check that the CUDA_HOME environment variable is set to your cuda home, on some clusters, I also needed to set export PATH="$CUDA_HOME/bin:$PATH"

Also, perhaps the fastest way to test just deepspeed installation with openfold3 is to use one of the unit tests, e.g.:

$ pytest -x openfold3/tests/test_kernels.py -k dsk_bf16

[kkbiu]

Thank you for your answer. I still haven't identified the problem so far. For the time being, I have temporarily disabled this mechanism by specifying runner.yml and successfully run the example.

Also, I would like to ask if whether deepspeed is enabled or not has a significant impact on the speed of model inference？or does this module mainly affect model training?

[jnwei]

We don't have benchmark numbers with OpenFold3 preview for deepspeed and cuequivariance yet. In general, deepspeed can have an impact for reducing memory requirements. This will especially have an impact for large queries with many tokens.

With an H100 80GB, and using cuequivariance, I think you might be fine for most queries.

[kkbiu]

Ok. Thank you for your patient answers. I have no other questions for the moment. Once again, I would like to thank your team for their contributions to the community！

[qurat-ul-ain95]

@jnwei Hi, I am also facing this error. seems like evoformer_attn skips downloaded the .so file.

[jnwei]

@qurat-ul-ain95 thanks for the report. A few questions to help identify the issue:

• Can you tell me if this error is occuring on a system where you previously had a working OF3 installation? Or is this error occuring upon installation of OF3 on a new device?
• Assuming it is the same device, did the versions of the drivers of your device (especially CUDA) change since your last working installation.

[qurat-ul-ain95]

Hi @jnwei, me and other fellows at my lab are unable to follow installation instructions on a system where previously OF3 worked. This is especially the case when we try to pip install openfold3[cuequivariance] (without cuequivariance works just fine).
Torch does not download the extension evoformer_attn.so. In my previous installations folder which worked I see this file, but the new installation does not have this file.

Thanks! :)

[jnwei]

Your observation about the missing evoformer_attn.so suggests to me that there's an issue with building and installing Deepspeed, rather than the torch download.

When you try to run inference or unit tests, did you see an error message that looked like this?

L/.../jnwei/miniforge3/envs/openfold_env/lib/python3.9/site-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda_cu -ltorch_cuda_cpp -ltorch -ltorch_python -L/opt/nvidia/hpc_sdk/Linux_x86_64/22.7/cuda/11.7/lib64 -lcudart -o evoformer_attn.so
FAILED: evoformer_attn.so

That error message would indicate the build failure.

To debug, could you please provide the versions of deepspeed and pytorch for the environments, with and without cuequivariance? Thanks!