#!/usr/bin/env bash
set -euo pipefail

# —––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––
# CONFIGURATION
# —––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––

# GPUs and DDP settings
export CUDA_VISIBLE_DEVICES="4,5,6,7"
export NPROC_PER_NODE=4

# Paths
DATASET_DIR="/data/qingq/PathVLM/baselines/github/SlideChat/dataset/stage_2_vqa_by_tumor/stage2_vqa_tumor_"
SCRIPT_DIR="configs/slidechat/experiments"
DEEPSPEED_CONFIG="configs/deepspeed/deepspeed_zero2.json"
OUTPUT_BASE="/data/qingq/PathVLM/baselines/github/SlideChat/models/outputs"

# Suffix for your work‐dirs
WORKDIR_SUFFIX="original_2048maxlength_train_token_reducer_attn"

# —––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––
# LOOP
# —––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––

for json_path in "${DATASET_DIR}"/*.json; do
  # extract e.g. "BLCA" then lowercase -> "blca"
  tumor="$(basename "${json_path}" .json)"
  tumor_lc="$(echo "${tumor}" | tr '[:upper:]' '[:lower:]')"

  cfg="${SCRIPT_DIR}/stage2_reducer_attn_${tumor_lc}.py"
  workdir="${OUTPUT_BASE}/${tumor_lc}_${WORKDIR_SUFFIX}"

  if [[ ! -f "${cfg}" ]]; then
    echo "⚠️  Config not found for ${tumor_lc}, skipping: ${cfg}"
    continue
  fi

  echo "🚀 Starting SFT on ${tumor_lc}"
  xtuner train \
    "${cfg}" \
    --deepspeed "${DEEPSPEED_CONFIG}" \
    --work-dir "${workdir}" \
    --local_rank 0

  echo "✅ Finished ${tumor_lc}"
  echo
done

echo "🎉 All jobs submitted."