The error is pretty straight forward. Your GPU doesn't have enough memory to predict the structure you're trying to. What are you trying to predict?

I had the same issues when I tried on an RTX4090 with 24Gb of VRAM. I could only fold monomers with a few hundred peptides in length. I just tested a bunch of different lengths until I found the limit of my GPU.

The memory use has improved significantly since they launched the project, but it's still a while off AF2s performance in terms of memory utilisation.

Depending on the length of your protein or complex you may be better off using the Neurosnap implementation of Boltz-1 which can be accessed using this link: https://neurosnap.ai/service/Boltz-1%20(AlphaFold3)

Our implementation uses H200s which have substantially more VRAM allowing you to predict substantially larger structures.

Full disclosure, I am the founder of Neurosnap.

Best of luck with your research!

The existing answers are pointing in the right direction: this is very likely a GPU memory/runtime limit rather than a failed pip install.

I would separate the checks into two parts.

First, confirm the environment:

python --version
python -c "import torch; print(torch.__version__, torch.cuda.is_available(), torch.version.cuda)"
nvidia-smi

Then run the smallest possible Boltz test input. If a tiny monomer works but your real target fails, the installation is probably fine and the failure is input size, complex size, or VRAM.

While the job is running, watch GPU memory:

watch -n 1 nvidia-smi

If memory climbs and then the process is killed or throws CUDA OOM, reducing the input size, number of samples/recycles if configurable, or moving to a larger GPU is the practical fix. Long proteins, multimers, and protein-ligand/nucleic-acid complexes can exceed a consumer GPU even when the software is installed correctly. As noted above, even a 24 GB RTX 4090 can be limiting for larger Boltz-1 jobs.

Also check that your PyTorch/CUDA build matches the installed NVIDIA driver. Some errors look like model failures but are really a driver/runtime mismatch.

Finally, keep the interpretation separate from successful execution. A completed Boltz/AlphaFold-like run is still a prediction: inspect confidence metrics, chain/interface confidence where available, and whether results are stable across seeds/settings before making biological claims.