Buckets:
| # Second queue: after the 1.5-epoch queue (pid $1) exits, run ModernBERT then Qwen for 2 epochs. | |
| # Each run: smoke test (writes resolved.yaml) -> training -> checkpoint/report upload to W&B. | |
| set -uo pipefail | |
| cd "$(dirname "$0")/.." | |
| export CUDA_VISIBLE_DEVICES=0 | |
| WAIT_PID=${1:-} | |
| if [[ -n "$WAIT_PID" ]]; then | |
| echo "[queue2] waiting for pid $WAIT_PID" | |
| while kill -0 "$WAIT_PID" 2>/dev/null; do sleep 20; done | |
| fi | |
| run_one() { # $1 config, $2 training module, $3 output dir, $4 log name | |
| echo "[queue2] $4 smoke test" | |
| if ! python -m training.smoke --config "$1" > "smoke_$4.log" 2>&1; then | |
| echo "[queue2] $4 smoke FAILED; see smoke_$4.log"; return | |
| fi | |
| echo "[queue2] $4 training" | |
| python -m "$2" --config "$3/resolved.yaml" > "$4.log" 2>&1 | |
| echo "[queue2] $4 exit $?" | |
| if [[ -f "$3/final/adapter_config.json" ]]; then | |
| python scripts/log_checkpoints_wandb.py "$3" || echo "[queue2] $4 artifact upload FAILED" | |
| fi | |
| } | |
| run_one training/modernbert_large_30k_2ep.yaml training.train_modernbert checkpoints/modernbert-large-30k-2ep modernbert_2ep | |
| run_one training/qwen3_06b_30k_2ep.yaml training.train_qwen checkpoints/qwen3-0.6b-30k-2ep qwen_2ep | |
| echo "[queue2] done" | |
Xet Storage Details
- Size:
- 1.2 kB
- Xet hash:
- a3c8a27a5a870676d941534b8ea3f5dabd80eb1897876a869101d950f9c4fca5
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.