monet_code/script_examples/sft_no_text_online.sh

#!/bin/bash
set -euo pipefail

# ============================================================================
# Monet no-text Stage 2 + Stage 3 with in-process online teacher.
# - Skips the offline Step 1 / Step 3 export passes.
# - Online teacher still writes its outputs to disk under
#     ${teacher_*_dir}/${fingerprint}/...
#   where fingerprint is sha1(abspath + key-file mtime/size)[:12]. This prevents
#   stale latents from a previous Stage 2 ckpt being silently reused.
# - To reuse online-produced files later in pure-offline mode, point
#   --teacher_reps_dir / --teacher_latent_dir at the fingerprinted SUBDIR
#   (see TEACHER_INFO.txt inside it for which ckpt produced the files).
# - Numerics match offline exactly: same teacher inputs (text + image budget) as
#   precompute_teacher_reps.py and precompute_teacher_latents.py.
# ============================================================================

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

export CUDA_VISIBLE_DEVICES

mkdir -p "${CKPT_ROOT}"

CE_EMPHASIZE_FACTOR=4.0
ALIGNMENT_WEIGHT=2.0
EMPHASIZE_LATENT_WEIGHT=2.0

# ─── Stage 2: train latent body. Teacher = frozen base model, lives next to student. ───
SAVE2=stage2_notext_latent${LATENT_SIZE}_ce${CE_EMPHASIZE_FACTOR}_al${ALIGNMENT_WEIGHT}_emph${EMPHASIZE_LATENT_WEIGHT}_online
"${TORCHRUN}" --nproc-per-node=${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 ${CKPT_ROOT}/${SAVE2} \
  --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 \
  --online_teacher \
  --online_teacher_model_path ${BASE_MODEL}

# ─── Stage 3: distill no-aux student. Teacher = frozen Stage 2 ckpt above. ───
STAGE2=${CKPT_ROOT}/${SAVE2}
SAVE3=stage3_notext_latent${LATENT_SIZE}_ce${CE_EMPHASIZE_FACTOR}_al${ALIGNMENT_WEIGHT}_online
"${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 ${CKPT_ROOT}/${SAVE3} \
  --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 \
  --online_teacher \
  --online_teacher_model_path ${STAGE2}