#!/bin/bash
# Download and setup Omni Senter 3B
# This script downloads the base Qwen2.5-Omni GGUF and applies our LoRA adapter

set -e

MODEL_DIR="$HOME/.cache/llama.cpp/models"
mkdir -p "$MODEL_DIR"

echo "=== Omni Senter 3B Setup ==="

# Download base model (Qwen2.5-Omni GGUF)
BASE_MODEL="$MODEL_DIR/Qwen2.5-Omni-3B-Q4_K_M.gguf"
if [ ! -f "$BASE_MODEL" ]; then
    echo "Downloading Qwen2.5-Omni-3B GGUF model..."
    huggingface-cli download ggml-org/Qwen2.5-Omni-3B-GGUF Qwen2.5-Omni-3B-Q4_K_M.gguf \
        --local-dir "$MODEL_DIR" \
        --token HF_TOKEN
else
    echo "Base model already exists: $BASE_MODEL"
fi

# Download mmproj
MMPROJ="$MODEL_DIR/mmproj-Qwen2.5-Omni-3B-Q8_0.gguf"
if [ ! -f "$MMPROJ" ]; then
    echo "Downloading mmproj..."
    huggingface-cli download ggml-org/Qwen2.5-Omni-3B-GGUF mmproj-Qwen2.5-Omni-3B-Q8_0.gguf \
        --local-dir "$MODEL_DIR" \
        --token HF_TOKEN
else
    echo "mmproj already exists: $MMPROJ"
fi

# Download LoRA adapter
LORA_DIR="$HOME/.cache/llama.cpp/omni-senter"
mkdir -p "$LORA_DIR"
echo "Downloading Omni Senter LoRA adapter..."
huggingface-cli download sovthpaw/Omni-Senter-3B senter-lora-500.gguf \
    --local-dir "$LORA_DIR" \
    --token HF_TOKEN

LORA_FILE="$LORA_DIR/senter-lora-500.gguf"

# Apply LoRA and create merged model
MERGED_MODEL="$MODEL_DIR/omni-senter-merged.gguf"
echo "Applying LoRA adapter..."
./llama-export-lora -m "$BASE_MODEL" -o "$MERGED_MODEL" --lora "$LORA_FILE"

echo ""
echo "=== Setup Complete! ==="
echo "Run with:"
echo "  llama-server -m $MERGED_MODEL --mmproj $MMPROJ --port 8107"