File size: 1,951 Bytes
e46c127 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | #!/usr/bin/env bash
# Sync scripts to a remote box and run train + eval there.
# ./run.sh sync # just rsync
# ./run.sh train Qwen/Qwen3.5-0.8B # train (full FT); add --lora --grad-ckpt for 9B via EXTRA
# ./run.sh gen # generate GSM8K candidates via llama-server (background-friendly)
# ./run.sh eval ckpt/qwen3.5-0.8b-nli results/qwen0.8b.json
set -euo pipefail
# override for another box, e.g.:
# HOST=mybox REMOTE=~/qwen_nli PY=python ENVS="HF_HOME=/mnt/hf" ./run.sh train ...
HOST=${HOST:-mybox}
REMOTE=${REMOTE:-~/qwen_nli}
PY=${PY:-python}
ENVS=${ENVS:-}
HERE="$(cd "$(dirname "$0")" && pwd)"
sync() { rsync -az --exclude results --exclude data "$HERE/"{train.py,eval.py,eval_image_nli.py,data_mix.py,run_v2_evals.sh,latent_mlp.py,summarize.py,flappy.py,flappy_video.py,sweep_flappy.sh,doom.py,doom_vision.py,minecraft.py,mc_bot.js,mc_record.js,mc_video.py,hf_publish.py,radar.py,modeling_openjev.py,modeling_qwen35_moe_seqcls.py,webql_bench.py,webql_fulldoc.py,webql_gemini_prompt.py,webql_mlp.py,run.sh} "$HOST:$REMOTE/"; }
case "${1:-}" in
sync) sync ;;
train)
sync
MODEL=${2:-Qwen/Qwen3.5-0.8B}; NAME=$(basename "$MODEL" | tr 'A-Z' 'a-z')
ssh "$HOST" "cd $REMOTE && mkdir -p logs && nohup env $ENVS $PY train.py --model $MODEL --out ckpt/${NAME}-nli ${EXTRA:-} > logs/train_${NAME}.log 2>&1 & echo started pid \$!"
;;
gen)
sync
ssh "$HOST" "cd $REMOTE && mkdir -p logs data && nohup env $ENVS $PY eval.py --models x --out /dev/null --gen-only > logs/gen_gsm8k.log 2>&1 & echo started pid \$!"
;;
eval)
sync
CKPT=${2:?ckpt}; OUT=${3:?out}
ssh "$HOST" "cd $REMOTE && mkdir -p logs results && nohup env $ENVS $PY eval.py --models $CKPT dleemiller/ModernCE-large-nli --out $OUT ${EXTRA:-} > logs/eval_$(basename "$OUT" .json).log 2>&1 & echo started pid \$!"
;;
*) echo "usage: $0 {sync|train MODEL|gen|eval CKPT OUT}"; exit 1 ;;
esac
|