setup_lambda.sh

#!/bin/bash
# Setup script for Lambda GPU instance
# Installs Unsloth (Gemma 4) + transformers (Qwen2.5 backup)
#
# Usage:
#   bash hackathon/upload_to_lambda.sh <LAMBDA_IP>
#   ssh ubuntu@<LAMBDA_IP>
#   bash ~/hackathon/setup_lambda.sh

set -euo pipefail

echo "=== AccountantGPT Lambda Setup ==="
echo "GPU: $(nvidia-smi --query-gpu=name --format=csv,noheader)"
echo "VRAM: $(nvidia-smi --query-gpu=memory.total --format=csv,noheader)"
echo "CUDA: $(nvidia-smi --query-gpu=driver_version --format=csv,noheader)"
echo ""

# Create venv
echo "Creating virtual environment..."
python3 -m venv ~/venv
source ~/venv/bin/activate

# Install Unsloth (includes transformers, peft, trl, bitsandbytes)
echo "Installing Unsloth + dependencies..."
pip install --upgrade pip
pip install "unsloth[cu124-torch260] @ git+https://github.com/unslothai/unsloth.git"
pip install wandb huggingface_hub scipy

# Also install standard transformers stack for Qwen2.5 backup
pip install -r ~/hackathon/requirements.txt 2>/dev/null || true

# HuggingFace login (needed for model downloads + pushing)
echo ""
echo "============================================"
echo "HuggingFace token needed for:"
echo "  - Gemma 4 (Apache 2.0, no gate)"
echo "  - Llama 3.1 (gated - pending approval)"
echo "  - Qwen2.5 (Apache 2.0, no gate)"
echo ""
echo "Create token: https://huggingface.co/settings/tokens"
echo "============================================"
echo ""
read -p "Enter your HuggingFace token (or press Enter to skip): " HF_TOKEN

if [ -n "$HF_TOKEN" ]; then
    huggingface-cli login --token "$HF_TOKEN"
    echo "Logged in to HuggingFace"
else
    echo "Skipping HF login - make sure you're already logged in"
fi

# Prepare CPA data
echo ""
echo "Preparing CPA FAR training data..."
cd ~/hackathon
python prepare_cpa_sft.py

# Pre-download models (both download in parallel)
echo ""
echo "Pre-downloading models (this takes a few minutes)..."

python3 -c "
from huggingface_hub import snapshot_download
import concurrent.futures

def download(model_id):
    print(f'Downloading {model_id}...')
    snapshot_download(model_id)
    print(f'  Done: {model_id}')

models = [
    'google/gemma-4-12b-it',        # Primary: Gemma 4 12B
    'Qwen/Qwen2.5-7B-Instruct',     # Backup: Qwen2.5
]

# Try Gemma 4 12B first; if not found, fall back to other sizes
try:
    download('google/gemma-4-12b-it')
except Exception as e:
    print(f'  Gemma 4 12B not available ({e}), trying 31B...')
    try:
        download('google/gemma-4-31b-it')
    except Exception as e2:
        print(f'  Gemma 4 31B failed ({e2}), trying E4B...')
        download('google/gemma-4-E4b-it')

download('Qwen/Qwen2.5-7B-Instruct')
print('All downloads complete!')
"

echo ""
echo "=== Setup complete! ==="
echo ""
echo "TRAINING COMMANDS:"
echo "  source ~/venv/bin/activate"
echo ""
echo "  ┌─────────────────────────────────────────────────────────┐"
echo "  │ PRIMARY: Gemma 4 (via Unsloth)                         │"
echo "  │                                                         │"
echo "  │  # Journal entries (~45 min for 31B, ~15 min for E4B)  │"
echo "  │  python ~/hackathon/train_gemma4.py --model 31b \\      │"
echo "  │      --run combined --no-wandb                          │"
echo "  │                                                         │"
echo "  │  # If OOM, use 4-bit:                                  │"
echo "  │  python ~/hackathon/train_gemma4.py --model 31b \\      │"
echo "  │      --run combined --use-4bit --no-wandb               │"
echo "  │                                                         │"
echo "  │  # If still OOM, use smaller model:                    │"
echo "  │  python ~/hackathon/train_gemma4.py --model e4b \\      │"
echo "  │      --run combined --no-wandb                          │"
echo "  └─────────────────────────────────────────────────────────┘"
echo ""
echo "  ┌─────────────────────────────────────────────────────────┐"
echo "  │ BACKUP: Qwen2.5-7B (via transformers)                  │"
echo "  │                                                         │"
echo "  │  python ~/hackathon/train_qlora.py --run combined \\    │"
echo "  │      --base-model Qwen/Qwen2.5-7B-Instruct --no-wandb  │"
echo "  └─────────────────────────────────────────────────────────┘"
echo ""
echo "  ┌─────────────────────────────────────────────────────────┐"
echo "  │ EVAL (after training)                                   │"
echo "  │                                                         │"
echo "  │  python ~/hackathon/eval_model.py \\                    │"
echo "  │      --model ./models/accountant-gpt-*/final \\         │"
echo "  │      --eval both                                        │"
echo "  └─────────────────────────────────────────────────────────┘"