setup.sh

#!/usr/bin/env bash
# Environment setup for LoRA fine-tuning with Unsloth
# Validated for RTX 5070 Laptop (Blackwell, sm_120) with CUDA 12.8+ and PyTorch 2.11
set -e

GPU=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1 || echo "not detected")
DRIVER_CUDA=$(nvidia-smi | grep -oP "CUDA Version: \K[\d\.]+" || echo "unknown")
echo "=== LoRA fine-tuning setup ==="
echo "GPU:         $GPU"
echo "Driver CUDA: $DRIVER_CUDA"
echo ""

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip --quiet

# cu128 wheels support Blackwell (sm_120) and require CUDA driver >= 12.8
# cu126 wheels top out at sm_90 — will NOT use the RTX 5000 GPU for training
echo "[1/4] Installing PyTorch 2.11 with CUDA 12.8 support (Blackwell)..."
pip install \
  "torch==2.11.0+cu128" \
  "torchvision" \
  "torchaudio" \
  --index-url https://download.pytorch.org/whl/cu128 \
  --quiet

echo "[2/4] Installing Unsloth..."
# Install from PyPI — this version is validated against PyTorch 2.11
# The unsloth-zoo metadata constraint (torch<2.11) is outdated and safe to ignore
pip install unsloth --quiet

echo "[3/4] Installing training and data libraries..."
pip install trl peft accelerate bitsandbytes datasets transformers --quiet

echo "[4/4] Installing utilities..."
pip install pyyaml wandb --quiet

echo ""
echo "=== Verifying installation ==="
python - <<'EOF'
from unsloth import FastLanguageModel
from trl import SFTTrainer, SFTConfig
import torch
assert torch.cuda.is_available(), "CUDA not available"
print(f"  torch:   {torch.__version__}")
print(f"  GPU:     {torch.cuda.get_device_name(0)}")
print(f"  VRAM:    {torch.cuda.get_device_properties(0).total_memory / 1024**3:.1f} GB")
print("  unsloth: OK")
print("  trl:     OK")
EOF

echo ""
echo "=== Setup complete ==="
echo ""
echo "Activate the environment:"
echo "  source .venv/bin/activate"
echo ""
echo "Workflow:"
echo "  make prepare   # process raw data into train/eval JSONL"
echo "  make train     # run fine-tuning"
echo "  make eval      # compare base vs fine-tuned"
echo "  make export    # export to GGUF for Ollama"
echo ""
echo "To enable Ollama export, install llama.cpp:"
echo "  git clone https://github.com/ggerganov/llama.cpp"
echo "  pip install -r llama.cpp/requirements.txt"