| #SBATCH -J text8-pure-cont-embed | |
| #SBATCH -p gpu # H100 partition :contentReference[oaicite:0]{index=0} | |
| #SBATCH --nodes=1 # keep all GPUs on the same node | |
| #SBATCH --gres=gpu:8 # all 8 H100 | |
| #SBATCH -A ruqiz | |
| #SBATCH --cpus-per-task=112 # give the single task every core | |
| #SBATCH -t 24:00:00 # wall-time (edit as needed) | |
| #SBATCH -o logs/%x-%j.out # capture stdout/stderr | |
| module purge | |
| module load conda | |
| conda activate cont_diff | |
| export SCRATCH_DIR=$RCAC_SCRATCH/cont_diff | |
| CUDA_LAUNCH_BLOCKING=1 torchrun --nproc_per_node=8 main.py scratch_dir=$SCRATCH_DIR \ | |
| data=text8 \ | |
| wandb.name=pure-cont-embed-text8 \ | |
| optim.lr=.0003 \ | |
| lr_scheduler=constant_warmup \ | |
| lr_scheduler.num_warmup_steps=250 \ | |
| wandb.project=candi-owt \ | |
| algo=candi \ | |
| model=small \ | |
| model.length=1024 \ | |
| trainer.max_steps=1000 \ | |
| trainer.accumulate_grad_batches=4 \ | |
| batch_size=512 \ | |
| trainer.val_check_interval=500 \ | |
| algo.mixed_coeff=0 \ | |
| algo.min_percentile=.1 \ | |
| algo.max_percentile=.4 \ | |
| algo.pure_continuous=True \ | |
| algo.is_embed=True; |