Download benchmark/run.sh from Tinghong-Ye22/code-HFD: direct link, hf CLI and curl.
- Browser
- Download file 2.95 kB
-
https://huggingface.co/Tinghong-Ye22/code-HFD/resolve/main/benchmark/run.sh
- Command line
-
hf download hf://Tinghong-Ye22/code-HFD/benchmark/run.sh
-
curl -L -o run.sh https://huggingface.co/Tinghong-Ye22/code-HFD/resolve/main/benchmark/run.sh
2.95 kB
| set -euo pipefail | |
| BENCH=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd) | |
| ROOT=$(dirname "$BENCH") | |
| PEER=$(dirname "$(dirname "$(dirname "$ROOT")")")/buwendong/user | |
| RUN=${RUN:-$ROOT/runs/posttrain_full_duplex/hfd.gpus4.2610100010-thirdparty} | |
| STEP=${STEP:-1800000} | |
| COUNT=${COUNT:-2048} | |
| GPUS=${GPUS:-0,1,2,3} | |
| WORKERS_PER_GPU=${WORKERS_PER_GPU:-3} | |
| COHORT=${COHORT:-$BENCH/outputs/sessions2048} | |
| EXTERNAL_ASSETS=${EXTERNAL_ASSETS:-$BENCH/outputs/external_visual/overlay_assets.json} | |
| OUTPUT=${OUTPUT:-$BENCH/outputs/session_ckpt_$STEP} | |
| HFD_PYTHON=${HFD_PYTHON:-$ROOT/.venv-psi/bin/python} | |
| if [[ -z ${EVAL_PYTHON:-} ]]; then | |
| EVAL_PYTHON=$BENCH/.venv-eval/bin/python | |
| [[ -x "$EVAL_PYTHON" ]] || EVAL_PYTHON=$PEER/envs/OMG/bin/python | |
| [[ -x "$EVAL_PYTHON" ]] || EVAL_PYTHON=python | |
| fi | |
| export PYTHONDONTWRITEBYTECODE=1 | |
| export OMP_NUM_THREADS=4 | |
| export MKL_NUM_THREADS=4 | |
| cd "$ROOT" | |
| mkdir -p "$OUTPUT/logs" | |
| exec 9>"$OUTPUT/.run.lock" | |
| flock -n 9 || { echo 'This benchmark output is already running'; exit 1; } | |
| [[ -f "$BENCH/vendor/manifest.json" ]] || { echo 'Missing evaluator bundle: run benchmark/vendor.py first'; exit 1; } | |
| if [[ ! -f "$COHORT/manifest.json" ]]; then | |
| [[ -f "$EXTERNAL_ASSETS" ]] || { echo "External QA is not published yet: $EXTERNAL_ASSETS"; exit 1; } | |
| "$HFD_PYTHON" -u "$BENCH/prepare.py" --run "$RUN" --output "$COHORT" --count "$COUNT" \ | |
| --external-assets "$EXTERNAL_ASSETS" | tee "$OUTPUT/logs/prepare.log" | |
| fi | |
| "$HFD_PYTHON" -c 'import json,sys,hashlib; d=json.load(open(sys.argv[1])); assert len(d["rows"])==int(sys.argv[2]), "Frozen cohort count mismatch"; a=open(sys.argv[3],"rb").read(); assert d.get("external_assets_sha256")==hashlib.sha256(a).hexdigest(), "Cohort must be revised for updated QA admission"; assert json.loads(a)["provenance"].get("direct_inspection_sha256"), "Direct QA inspection required"' "$COHORT/manifest.json" "$COUNT" "$EXTERNAL_ASSETS" | |
| IFS=',' read -r -a devices <<< "$GPUS" | |
| pids=() | |
| workers=$(( ${#devices[@]} * WORKERS_PER_GPU )) | |
| for ((shard=0; shard<workers; shard++)); do | |
| device=${devices[$((shard % ${#devices[@]}))]} | |
| CUDA_VISIBLE_DEVICES=$device "$HFD_PYTHON" -u "$BENCH/generate.py" \ | |
| --run "$RUN" --step "$STEP" --cohort "$COHORT" --output "$OUTPUT/generated" \ | |
| --shard "$shard" --shards "$workers" >"$OUTPUT/logs/generate_$shard.log" 2>&1 & | |
| pids+=("$!") | |
| done | |
| status=0 | |
| for pid in "${pids[@]}"; do wait "$pid" || status=1; done | |
| (( status == 0 )) || exit 1 | |
| "$HFD_PYTHON" -B "$BENCH/audit.py" --cohort "$COHORT" --generated "$OUTPUT/generated" | |
| if [[ ! -f "$OUTPUT/eval/motion/paper_metrics.json" ]]; then | |
| CUDA_VISIBLE_DEVICES=${devices[0]} "$EVAL_PYTHON" -B -u "$BENCH/evaluate.py" \ | |
| --cohort "$COHORT" --generated "$OUTPUT/generated" --output "$OUTPUT/eval" | tee "$OUTPUT/logs/eval.log" | |
| fi | |
| "$EVAL_PYTHON" -B "$BENCH/evaluate.py" --cohort "$COHORT" --generated "$OUTPUT/generated" \ | |
| --output "$OUTPUT/eval" --verify-only | |