#!/bin/bash #SBATCH -A YOUR_ACCOUNT #SBATCH -p gpu #SBATCH --nodes 4 #SBATCH --gpus-per-node 4 #SBATCH --ntasks-per-node 1 #SBATCH -c 128 #SBATCH -t 10:00:00 #SBATCH --job-name=insc-ftw2 #SBATCH -o logs/ftwhole-%j.out #SBATCH -e logs/ftwhole-%j.out # SLURM copies the submitted script into a per-job spool dir before running it on this # cluster, so locating the repo via ${BASH_SOURCE[0]} resolves to that spool path, not # the real one. $SLURM_SUBMIT_DIR is set by sbatch to the directory it was invoked from, # immune to that copy (see scripts/slurm/README.md point 6). STOICHEIA_ROOT="${SLURM_SUBMIT_DIR:-$PWD}" export STOICHEIA_ROOT # 4-node (16-GPU) whole-document documentary finetune (torchrun) -- finetune_whole.py # already has full DDP support (ddp_setup()/DDP wrap, the same pattern as pretraining), this # just launches it under torchrun instead of plain python. Use attn=sdpa in the configs # (the compiled flex_attention block-sparse kernel can fall back to a memory-exploding # dense computation at these sequence lengths). # Usage: sbatch scripts/slurm/insc_finetune_whole_4node.sbatch set -euo pipefail CONFIG="${1:?usage: sbatch scripts/slurm/insc_finetune_whole_4node.sbatch }" INS_ROOT=$STOICHEIA_ROOT source "$INS_ROOT/env.sh" MASTER=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -1) export MASTER_ADDR=$MASTER MASTER_PORT=$((29000 + SLURM_JOB_ID % 1000)) NGPU=${SLURM_GPUS_PER_NODE:-4} echo "config=$CONFIG nnodes=$SLURM_NNODES ngpu=$NGPU $(date)" srun --ntasks="$SLURM_NNODES" --ntasks-per-node=1 \ apptainer exec --nv $APPTAINER_BINDS $SIF bash -lc " set -e export PYTHONPATH=$STOICHEIA_ROOT export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True export OMP_NUM_THREADS=16 cd $INS_ROOT torchrun \ --nnodes=$SLURM_NNODES --nproc_per_node=$NGPU \ --rdzv_id=$SLURM_JOB_ID --rdzv_backend=c10d \ --rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT \ --node_rank=\$SLURM_PROCID \ insc/train/finetune_whole.py --config $INS_ROOT/$CONFIG " echo FT_JOB_END