monet_code/script_examples/sft_no_text.sh

#!/bin/bash
set -euo pipefail

if command -v conda >/dev/null 2>&1; then
  CONDA_BASE=$(conda info --base 2>/dev/null || true)
  if [ -n "${CONDA_BASE}" ] && [ -f "${CONDA_BASE}/etc/profile.d/conda.sh" ]; then
    source "${CONDA_BASE}/etc/profile.d/conda.sh"
    if conda env list | awk '{print $1}' | grep -qx "monet"; then
      conda activate monet
    fi
  fi
fi

REPO_DIR=${REPO_DIR:-/raid/yrl/Monet}
cd "${REPO_DIR}"

DATA_ROOT=${DATA_ROOT:-/raid/yrl/dataset_v1_60k/full_no_tool_paper_strict_v1}
TRAIN_JSON=${TRAIN_JSON:-/raid/yrl/dataset_v1_60k/full_no_tool_paper_strict_v1/no_text/train.jsonl}
BASE_MODEL=${BASE_MODEL:-/raid/yrl/hf_models/Qwen2.5-VL-7B-Instruct}
CKPT_ROOT=${CKPT_ROOT:-/raid/yrl/Monet_no_text_ckpts}
TORCHRUN=${TORCHRUN:-/raid/yrl/.envs/lvr/bin/torchrun}
CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}
NPROC=${NPROC:-8}
LATENT_SIZE=8
TOTAL_SAMPLES=${TOTAL_SAMPLES:-53106}
WAIT_SECONDS=${WAIT_SECONDS:-300}

export CUDA_VISIBLE_DEVICES

count_teacher_reps() {
  find "${CKPT_ROOT}/teacher_reps_pooled" -maxdepth 1 -type f -name 'rep_*.pt' 2>/dev/null | wc -l
}

wait_for_existing_teacher_reps_job() {
  while pgrep -f "src.precompute_teacher_reps.*${CKPT_ROOT}/teacher_reps_pooled" >/dev/null; do
    cur=$(count_teacher_reps)
    echo "[wait] existing teacher reps job running: ${cur}/${TOTAL_SAMPLES}"
    sleep "${WAIT_SECONDS}"
  done
}

mkdir -p "${CKPT_ROOT}"

wait_for_existing_teacher_reps_job

# Step 1: precompute pooled aux-image teacher representations from the base model.
mkdir -p "${CKPT_ROOT}/teacher_reps_pooled"
if [ "$(count_teacher_reps)" -lt "${TOTAL_SAMPLES}" ]; then
  "${TORCHRUN}" --nproc-per-node=${NPROC} -m src.precompute_teacher_reps \
    --bsz 1 \
    --data_path ${TRAIN_JSON} \
    --load_model_path ${BASE_MODEL} \
    --save_model_path ${CKPT_ROOT}/teacher_reps_pooled \
    --dataset_root ${DATA_ROOT} \
    --latent_size ${LATENT_SIZE} \
    --output_hidden_states \
    --alignment_layer all_layers \
    --allow_no_observation \
    --resume
fi

# Step 2: SFT Stage 2 trains the latent body from the base model.
CE_EMPHASIZE_FACTOR=4.0
ALIGNMENT_WEIGHT=2.0
EMPHASIZE_LATENT_WEIGHT=2.0
SAVE=stage2_notext_latent${LATENT_SIZE}_ce${CE_EMPHASIZE_FACTOR}_al${ALIGNMENT_WEIGHT}_emph${EMPHASIZE_LATENT_WEIGHT}
"${TORCHRUN}" --nproc-per-node=${NPROC} -m src.main \
  --epochs 2 --bsz 1 --grad_accum_steps 16 \
  --stage sft_stage2 \
  --data_path ${TRAIN_JSON} \
  --load_model_path ${BASE_MODEL} \
  --save_model_path ${CKPT_ROOT}/${SAVE} \
  --dataset_root ${DATA_ROOT} \
  --deepspeed ./deepspeed/ds_zero2_gpu.json \
  --latent_size ${LATENT_SIZE} \
  --alignment_weight ${ALIGNMENT_WEIGHT} \
  --emphasize_latent_weight ${EMPHASIZE_LATENT_WEIGHT} \
  --ce_emphasize_factor ${CE_EMPHASIZE_FACTOR} \
  --teacher_reps_dir ${CKPT_ROOT}/teacher_reps_pooled \
  --alignment_layer all_layers \
  --allow_no_observation

# Step 3: dump Stage 2 teacher latents.
STAGE2=${CKPT_ROOT}/${SAVE}
"${TORCHRUN}" --nproc-per-node=${NPROC} -m src.precompute_teacher_latents \
  --bsz 1 \
  --data_path ${TRAIN_JSON} \
  --load_model_path ${STAGE2} \
  --save_model_path ${CKPT_ROOT}/teacher_latents \
  --dataset_root ${DATA_ROOT} \
  --latent_size ${LATENT_SIZE} \
  --output_hidden_states \
  --allow_no_observation \
  --resume

# Step 4: SFT Stage 3 distills no-aux-image student latents from the base model.
CE_EMPHASIZE_FACTOR=4.0
ALIGNMENT_WEIGHT=2.0
SAVE=stage3_notext_latent${LATENT_SIZE}_ce${CE_EMPHASIZE_FACTOR}_al${ALIGNMENT_WEIGHT}
"${TORCHRUN}" --nproc-per-node=${NPROC} -m src.main \
  --epochs 2 --bsz 1 --grad_accum_steps 16 \
  --stage sft_stage3 \
  --data_path ${TRAIN_JSON} \
  --load_model_path ${BASE_MODEL} \
  --save_model_path ${CKPT_ROOT}/${SAVE} \
  --dataset_root ${DATA_ROOT} \
  --deepspeed ./deepspeed/ds_zero2_gpu.json \
  --latent_size ${LATENT_SIZE} \
  --alignment_weight ${ALIGNMENT_WEIGHT} \
  --ce_emphasize_factor ${CE_EMPHASIZE_FACTOR} \
  --teacher_latent_dir ${CKPT_ROOT}/teacher_latents \
  --alignment_layer all_layers \
  --allow_no_observation