setup_and_sft.sh

#!/usr/bin/env bash
# =============================================================================
#  LUNA 100M — SFT Cloud Setup & Train Entrypoint
#  Runs on RunPod, Vast.ai, Lambda Labs, or any Linux GPU pod.
#
#  USAGE:
#    bash setup_and_sft.sh [huggingface|local] [HF_REPO_OR_PATH]
#
#  EXAMPLES:
#    # From HuggingFace (recommended):
#    bash setup_and_sft.sh huggingface ASTERIZER/LUNA
#
#    # Already cloned locally:
#    bash setup_and_sft.sh local /workspace/LUNA
# =============================================================================

set -e

DATA_SOURCE="${1:-huggingface}"
REPO_ID="${2:-ASTERIZER/LUNA}"

echo "=========================================="
echo "  LUNA 100M — SFT Setup"
echo "  Source: $DATA_SOURCE"
echo "  Repo  : $REPO_ID"
echo "=========================================="

# ── 1. Python packages ────────────────────────────────────────────────────────
echo ""
echo "[1/4] Installing dependencies..."

pip install -q --upgrade pip
pip install -q \
    torch torchvision \
    psutil \
    pyyaml \
    transformers \
    huggingface_hub \
    datasets 2>/dev/null || true

echo "      Done."

# ── 2. Clone repo / fetch data ────────────────────────────────────────────────
echo ""
echo "[2/4] Fetching repository..."

WORK_DIR="/workspace/LUNA"

if [ "$DATA_SOURCE" = "huggingface" ]; then
    if [ ! -d "$WORK_DIR" ]; then
        # Clone the HF Space as a git repo
        HF_TOKEN="${HF_TOKEN:-}"
        if [ -n "$HF_TOKEN" ]; then
            git clone "https://huggingface.co/spaces/$REPO_ID" "$WORK_DIR" || \
            python -c "
from huggingface_hub import snapshot_download
snapshot_download(repo_id='$REPO_ID', repo_type='space', local_dir='$WORK_DIR',
                  token='$HF_TOKEN' if '$HF_TOKEN' else None)
"
        else
            git clone "https://huggingface.co/spaces/$REPO_ID" "$WORK_DIR" || \
            python -c "
from huggingface_hub import snapshot_download
snapshot_download(repo_id='$REPO_ID', repo_type='space', local_dir='$WORK_DIR')
"
        fi
    else
        echo "  $WORK_DIR already exists, pulling latest..."
        cd "$WORK_DIR" && git pull || echo "  (not a git repo, using existing)"
    fi
elif [ "$DATA_SOURCE" = "local" ]; then
    WORK_DIR="$REPO_ID"
fi

cd "$WORK_DIR"
echo "  Working dir: $(pwd)"

# ── 3. System probe ──────────────────────────────────────────────────────────
echo ""
echo "[3/4] System probe..."
python -c "
import torch, psutil, os
props = torch.cuda.get_device_properties(0) if torch.cuda.is_available() else None
print(f'  GPU  : {props.name if props else \"CPU only\"} ({props.total_memory/1024**3:.1f} GB)' if props else '  GPU: None')
print(f'  RAM  : {psutil.virtual_memory().total/1024**3:.1f} GB')
print(f'  CPUs : {os.cpu_count()}')
"

# Verify files exist
echo ""
echo "  Checking required files..."
for f in sft_train.py sft_config.yaml Base/Datasets/sft_clean/train.json; do
    if [ -f "$f" ]; then
        echo "    ✓ $f"
    else
        echo "    ✗ $f  MISSING!"
    fi
done
echo "  (Pretrained checkpoint will be auto-downloaded from HuggingFace if not present)"

# ── 4. Train SFT ─────────────────────────────────────────────────────────────
echo ""
echo "[4/4] Starting SFT training..."
echo ""

python sft_train.py \
    --config sft_config.yaml

echo ""
echo "=========================================="
echo "  SFT complete!  Output: Base/out/sft/luna_100m_sft"
echo "=========================================="