#!/usr/bin/env bash
# Interactive 2xH100 LoRA SFT. Run this on the compute node, not the login node.
# Effective batch = 1 * 16 * 2 = 32
set -euo pipefail
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
cd "$ROOT"

export PYTHONPATH="$ROOT/src${PYTHONPATH:+:$PYTHONPATH}"
export HF_HOME="${HF_HOME:-$ROOT/.cache/huggingface}"
export TOKENIZERS_PARALLELISM=false

if [[ -n "${PHYSICS_SFT_CONDA_ENV:-}" ]]; then
  # shellcheck disable=SC1091
  source ~/.bashrc
  conda activate "$PHYSICS_SFT_CONDA_ENV"
elif [[ -x "$ROOT/.venv/bin/python" ]]; then
  # shellcheck disable=SC1091
  source "$ROOT/.venv/bin/activate"
fi

PYTHON="$(command -v python)"
echo "cwd=$(pwd)  python=$PYTHON  CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-all}"
nvidia-smi -L || true

# DDP, not ZeRO-3: LoRA 7B fits on one H100; ZeRO-3 made the first step take many minutes.
"$PYTHON" -m accelerate.commands.launch --num_processes 2 \
  --config_file configs/accelerate_ddp_2gpu.yaml \
  src/physics_model/sft.py --config configs/sft_qwen25vl_7b_2gpu.yaml \
  --skip-length-scan