ValueError: implementation='triton' is unsupported on this GPU generation.

[yipy0005]

I was running the following:

#!/bin/bash

AF="af3"

singularity exec \
     --nv \
     --bind $AF/af_input:/root/af_input \
     --bind $AF/af_output:/root/af_output \
     --bind $AF/models:/root/models \
     --bind $AF/public_databases:/root/public_databases \
     alphafold3.sif \
     python alphafold3/run_alphafold.py \
     --json_path=/root/af_input/fold_input.json \
     --model_dir=/root/models \
     --db_dir=/root/public_databases \
     --output_dir=/root/af_output

and I got the error ValueError: implementation='triton' is unsupported on this GPU generation. when it was running model inference.

Appreciate if I could get some help on this. Thank you! 😁

[championsnet]

This happened for me when using an RTX 4090 with a compute capability <8.0. If you go with an A100 and compute capability >8.0, as recommended, you should be avoiding this problem.

[jacobjinkelly]

Hi @yipy0005, you can use the flag --flash_attention_implementation=xla (as outlined in our performance docs) to disable flash attention. As @championsnet says, this will work on GPUs with compute capability <8.0. Closing the issue for now, but please feel free to re-open if you still run into any issues! :)

[jacobjinkelly]

Just wanted to flag that we improved the error message for this in 3599612

[smg3d]

This happened for me when using an RTX 4090 with a compute capability <8.0. If you go with an A100 and compute capability >8.0, as recommended, you should be avoiding this problem.

Just to point out that the GeForce RTX 4090 has a compute capability > 8.0. Not that it matters, but in fact it has a higher compute capability (8.9) than the A100 (8.0). https://developer.nvidia.com/cuda-gpus