#!/bin/bash
# 2× GPU DDP restart
# curl -fsSL "https://huggingface.co/datasets/datamatters24/scriptwriter-runpod/resolve/main/restart-train.sh" | bash
set -euo pipefail

echo "=== Force-stopping old training ==="
tmux kill-session -t scriptwriter-train 2>/dev/null || true
tmux kill-server 2>/dev/null || true
pkill -9 -f 'train_runpod.py' 2>/dev/null || true
pkill -9 -f 'scripts/train_runpod' 2>/dev/null || true
pkill -9 -f torchrun 2>/dev/null || true
sleep 2
echo "Old session cleared."

DEST=/workspace/scriptwriter-runpod
TGZ=/tmp/scriptwriter-runpod-ready.tgz
curl -fsSL "https://huggingface.co/datasets/datamatters24/scriptwriter-runpod/resolve/main/scriptwriter-runpod-ready.tgz" -o "$TGZ"
rm -rf "$DEST"
mkdir -p "$DEST"
tar xzf "$TGZ" -C "$DEST" --strip-components=1 --no-same-owner --no-same-permissions

export WORKDIR="$DEST"
export DATA_DIR=/workspace/data/processed
export OUTPUT_DIR=/workspace/models/lora
export HF_DATASET_REPO="${HF_DATASET_REPO:-datamatters24/scriptwriter-corpus-ia}"
export HF_MODEL_REPO="${HF_MODEL_REPO:-datamatters24/scriptwriter-lora-ia}"
export BASE_MODEL="${BASE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}"
export RUNPOD_SKIP_FORCE_PIP=1
unset CUDA_VISIBLE_DEVICES
export MAX_SEQ_LENGTH=2048
export MAX_SEQ_LENGTH_CAP=2048
export PER_DEVICE_TRAIN_BATCH_SIZE=1
export GRADIENT_ACCUMULATION_STEPS=4
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
export TOKENIZERS_PARALLELISM=false
export RESUME_FROM_CHECKPOINT="${RESUME_FROM_CHECKPOINT:-1}"
export NPROC="${NPROC:-2}"

if [[ ! -f /workspace/data/processed/train.jsonl ]]; then
  echo "train.jsonl missing — use fix-and-train.sh instead"
  exit 1
fi

cd "$DEST"
mkdir -p /workspace/logs /workspace/models/lora
mkdir -p "$DEST/data"
ln -sfn /workspace/data/processed "$DEST/data/processed"

apt-get update -qq && DEBIAN_FRONTEND=noninteractive apt-get install -y -qq tmux >/dev/null || true

WORKER=/workspace/logs/_train_worker.sh
cat > "$WORKER" <<'INNER'
#!/bin/bash
set -euo pipefail
cd /workspace/scriptwriter-runpod
export HF_TOKEN="${HF_TOKEN:-}"
export HUGGING_FACE_HUB_TOKEN="${HF_TOKEN:-}"
export HF_HOME="${HF_HOME:-/workspace/.cache/huggingface}"
export CONFIG_PATH=/workspace/scriptwriter-runpod/config/training.yaml
export BASE_MODEL="${BASE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}"
export OUTPUT_DIR=/workspace/models/lora
export HF_MODEL_REPO="${HF_MODEL_REPO:-datamatters24/scriptwriter-lora-ia}"
unset CUDA_VISIBLE_DEVICES
export MAX_SEQ_LENGTH=2048
export MAX_SEQ_LENGTH_CAP=2048
export PER_DEVICE_TRAIN_BATCH_SIZE=1
export GRADIENT_ACCUMULATION_STEPS=4
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
export TOKENIZERS_PARALLELISM=false
export RESUME_FROM_CHECKPOINT="${RESUME_FROM_CHECKPOINT:-1}"
NPROC="${NPROC:-2}"
exec > >(tee -a /workspace/logs/train_live.log) 2>&1
echo "========== TRAIN START $(date -Is) =========="
echo "DDP on ${NPROC} GPUs: seq=2048 batch=1 accum=4 (NCCL_P2P/IB disabled)"
ls /workspace/models/lora/checkpoint-* 2>/dev/null | tail -5 || echo "no prior checkpoint"
nvidia-smi || true
torchrun --standalone --nproc_per_node="$NPROC" scripts/train_runpod.py
echo "========== UPLOAD $(date -Is) =========="
python3 scripts/sync_hf.py upload-model --repo "$HF_MODEL_REPO" --folder /workspace/models/lora
echo "========== DONE $(date -Is) =========="
echo "STOP THE POD"
INNER
chmod +x "$WORKER"

tmux new-session -d -s scriptwriter-train -n train "bash $WORKER"
tmux new-window -t scriptwriter-train -n monitor
tmux send-keys -t scriptwriter-train:monitor "watch -n 15 'nvidia-smi --query-gpu=index,memory.used,utilization.gpu --format=csv; echo; tail -25 /workspace/logs/train_live.log'" Enter
echo "Started 2-GPU DDP — attach: tmux attach -t scriptwriter-train"
echo "Expect both GPUs ~6–11 GiB and high util."