#!/bin/bash #SBATCH -A YOUR_ACCOUNT # EDIT: your SLURM account #SBATCH -p gpu # EDIT: your GPU partition name #SBATCH --nodes 1 #SBATCH --gpus-per-node 1 #SBATCH --ntasks-per-node 1 #SBATCH -c 64 #SBATCH -t 12:00:00 #SBATCH --job-name=syn-train #SBATCH -o logs/train-%j.out #SBATCH -e logs/train-%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 ("/var/lib/slurm/..." errors downstream) -- a real failure mode hit # repeatedly in practice. $SLURM_SUBMIT_DIR is set by sbatch to the directory it was # invoked from, immune to that copy, and matches this repo's own submit-from-root # convention (see scripts/slurm/README.md point 6). STOICHEIA_ROOT="${SLURM_SUBMIT_DIR:-$PWD}" export STOICHEIA_ROOT # Train one arm of the biaffine parser (frozen Stoicheia/LemmaDiff + scalar mix + biaffine head). # sbatch scripts/slurm/syntax_train.sbatch configs/syntax/parser_char.json set -euo pipefail SYN_ROOT=$STOICHEIA_ROOT source "$SYN_ROOT/env.sh" CONFIG="${1:?usage: sbatch scripts/train.sbatch }" mkdir -p "$SYN_ROOT/logs" srun apptainer exec --nv $APPTAINER_BINDS $SIF bash -lc " export PYTHONPATH=$SYN_ROOT:$TAGGER_ROOT:$STOICHEIA_ROOT export STOICHEIA_CKPT=$STOICHEIA_CKPT cd $SYN_ROOT python -u -m parser.train --config $CONFIG " echo TRAIN_DONE