spb/wpai0002_uqo

1#!/usr/bin/env bash2# Second Lambda instance (gpu_2x_h100_sxm5): GPU 0 runs the HTR-Large chain, GPU 1 a second seed of HTR-Base.3# Same stages as lambda_train_chain.sh (denoising pretraining → pair fine-tuning → val/test/held-out predictions).4set -uo pipefail5cd ~/wpai00026export PATH="$HOME/.local/bin:$PATH"7P=data/pairs/medium_v1_haiku/validated/pairs.jsonl; S=data/splits/medium_v1/splits.jsonl; T=artifacts/tokenizers/medium_v08COR=data/parsed/medium_v0; SP0=data/splits/medium_v0/splits.jsonl9run_size() {  # $1 gpu  $2 size  $3 run suffix  $4 seed  $5 denoise steps  $6 denoise max_tokens  $7 ft lr  $8 ft max_tokens10  local GPU=$1 SZ=$2 SUF=$3 SEED=$4 DS=$5 DT=$6 LR=$7 FT=$811  export CUDA_VISIBLE_DEVICES=$GPU12  .venv/bin/python scripts/pretrain_denoise.py --corpus $COR --splits $SP0 --tokenizer $T --run denoise_m1_${SZ}_${SUF} --arch m1 --size $SZ --seed $SEED --max-steps $DS --eval-every 1000 --warmup 1500 --lr 2e-4 --variants-per-passage 2 --max-tokens $DT --max-batch 128 --patience 6 --device cuda --amp bf16 --num-workers 16 > train_denoise_${SUF}.log 2>&113  .venv/bin/python scripts/train.py --pairs $P --splits $S --tokenizer $T --run medium_m1_${SZ}_${SUF} --preset m1 --size $SZ --seed $SEED --init-from artifacts/runs/denoise_m1_${SZ}_${SUF}/best --lr $LR --max-steps 30000 --eval-every 1000 --warmup 500 --max-tokens $FT --max-batch 128 --patience 6 --device cuda --amp bf16 --num-workers 16 > train_medium_${SUF}.log 2>&114  .venv/bin/python scripts/predict.py --pairs $P --splits $S --split val --checkpoint artifacts/runs/medium_m1_${SZ}_${SUF}/best --system htr-m1-${SZ}-${SUF} --out artifacts/predictions/medium_val --limit 3000 --device cuda --batch-size 64 > predict_${SUF}_val.log 2>&115  for SPL in test heldout_domain heldout_period; do16    .venv/bin/python scripts/predict.py --pairs $P --splits $S --split $SPL --checkpoint artifacts/runs/medium_m1_${SZ}_${SUF}/best --system htr-m1-${SZ}-${SUF} --out artifacts/predictions/medium_${SPL} --device cuda --batch-size 64 > predict_${SUF}_${SPL}.log 2>&117  done18  echo "CHAIN ${SUF} DONE" >> chain_done.txt19}20run_size 0 large lambda 1 40000 12288 6e-5 8192 &21run_size 1 base lambda-s2 2 30000 16384 1e-4 12288 &22wait23echo "ALL CHAINS DONE" >> chain_done.txt24