Using Tiberius v2.0.6 on AMD GPUs

Tiberius is a deep learning-based ab initio gene structure prediction tool that end-to-end integrates convolutional and long short-term memory layers with a differentiable HMM layer. It can be used to predict gene structures from genomic sequences only (ab initio), while matching the accuracy of tools that use extrinsic evidence.

Example job script

#!/bin/bash -l
#SBATCH -A ${PAWSEY_PROJECT}-gpu
#SBATCH --nodes=1
#SBATCH --partition=gpu
#SBATCH --time=1:00:00
#SBATCH --gres=gpu:1

# Load required module
module load singularity/3.11.4-nompi

# Example command
srun -N 1 -n 1 -c 8 --gres=gpu:1 --gpus-per-task=1 \
  singularity exec docker://quay.io/pawsey/tiberius:v2.0.6_rocm7.0.2 \
  tiberius \
  --genome $MYSCRATCH/Tiberius/test_data/Panthera_pardus/inp/genome.fa \
  --model_cfg mammalia_softmasking_v2 \
  --out test.gtf \
  --batch_size 32 \
  --seq_len 259992

Important notes:

The batch_size 32 and seq_len 259992 make the best use of the GPU memory, and provide a 32% speedup compared to the default settings.
Each Tiberius process can only use one GPU at a time, so you want to always set --gres=gpu:1 --gpus-per-task=1. To be very specific, Tiberius is actually using one GCD per GPU.
Consider downloading the container ahead of time so you’re not wasting expensive GPU time pulling the container.