Is it desirable to utilize tensor cores by modifying matmul precision in bin files?

[TristynAlxander]

When running rfd3 there is a line that says:

You are using a CUDA device ('NVIDIA GeForce RTX 4060 Laptop GPU') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision

If I modify the rfd3 in the bin to this:

#!/opt/.miniconda/envs/protein_foundry/bin/python3.12
try:
    import torch
    # TO-DO: make if-statement for tensor cores? 
    torch.set_float32_matmul_precision("high")
except Exception:
    print("Failed to Set matmul precision.")

import sys
from rfd3.cli import app
if __name__ == '__main__':
    if sys.argv[0].endswith('.exe'):
        sys.argv[0] = sys.argv[0][:-4]
    sys.exit(app())

The comment goes away. That said, I have no idea if this is a good idea or if the precision loss will be harmful in some way.

I tried using seed=42 in the command line interface; however, I got subtle differences even without this modification that make me think I might need an additional setting to make it deterministic. Weirdly, the "high" precision setting gave me something more different than the "medium" precision setting. My best guess is time-derived pseudo-randomness somewhere is messing with the outputs.

Since the medium gave me something more similar, I'm thinking I will leave this in my bin file, but since I really don't know what I'm doing I thought it wise to ask the experts. Thanks!