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

# One-command recommended training run for a single RTX 6000 Ada 48/50GB server.
# Usage:
#   export HF_TOKEN=...
#   export TRACKIO_SPACE_ID=nraptisss/tmf921-trackio   # optional
#   bash scripts/run_rtx6000ada.sh

python -m pip install -U pip
python -m pip install -r requirements.txt

# Optional throughput improvement. Uncomment only if compatible with your CUDA/PyTorch build.
# python -m pip install flash-attn --no-build-isolation

export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
export TOKENIZERS_PARALLELISM=false
export PYTHONPATH="$PWD/src:${PYTHONPATH:-}"

python scripts/train_qlora.py \
  --config configs/rtx6000ada_qwen3_8b_qlora.yaml

python scripts/evaluate_model.py \
  --model Qwen/Qwen3-8B \
  --adapter outputs/qwen3-8b-tmf921-qlora \
  --dataset nraptisss/TMF921-intent-to-config-research-sota \
  --output_dir outputs/qwen3-8b-tmf921-qlora/eval \
  --load_in_4bit