| # LR sweep at half-length with best WD from prior sweep (wd=0.05). | |
| set -euo pipefail | |
| cd "$(dirname "$0")/../../modded-nanogpt" | |
| SWEEP_SCRIPT="../artifacts/adamw_baseline_v2_cmpatino-0/train_gpt_adamw_v2_cmpatino-0.py" | |
| RESULTS_DIR="../artifacts/adamw_sweep_cmpatino-0" | |
| mkdir -p "$RESULTS_DIR" | |
| export TRAIN_STEPS=2812 | |
| export WARMUP_STEPS=125 | |
| export COOLDOWN_FRAC=0.7 | |
| export BLOCK_WD=0.05 | |
| export BLOCK_BETA1=0.9 | |
| export BLOCK_BETA2=0.95 | |
| for LR in 0.0010 0.0020 0.0030; do | |
| TAG="lr${LR}_wd0.05_steps${TRAIN_STEPS}_cmpatino-0" | |
| LOG="${RESULTS_DIR}/log_${TAG}.txt" | |
| echo "=== Running BLOCK_LR=${LR} -> ${LOG} ===" | |
| BLOCK_LR=$LR RUN_TAG="$TAG" \ | |
| ../.venv/bin/torchrun --standalone --nproc_per_node=2 "$SWEEP_SCRIPT" 1 \ | |
| 2>&1 | tee "$LOG" | |
| done | |
| echo "=== LR sweep complete ===" | |
Xet Storage Details
- Size:
- 804 Bytes
- Xet hash:
- 2b0dc2da127395466a11c3dffb2f06eb162260ebb5868c15268206e05f146fea
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.