Memory issues on GPU

Hi all,

Getting issues running on A100 GPU with batch size of 1 on the CASP 14 PDBs. Running the following command
`python scripts/tokenize_pdb.py --token_save_path ./test_tokens --pdb_dir casp14_pdbs --backend gpu --batch_size_per_device 1`

gets the following error:
```
... [Omitted] ...
---Devices---
local device count: 1
2025-07-22 08:48:33 | 179132 | INFO | tokenize_pdb:main:60 | params loaded
0%| | 0/31 [00:00

Start with scripts/tokenize_pdb.py and reproduce the reported command on an A100 using the listed environment, focusing on the cuDNN initialization failure. Done means the command can tokenize a single structure with batch size 1, or the incompatibility and required environment changes are documented with a reproducible result.

Written by the indexing model from the issue text.