#!/usr/bin/env bash
set -euo pipefail

# CUDA inference for ghana-tts-36k via VoxCPM.cpp
# Requires an NVIDIA GPU + CUDA toolkit (nvcc) on this machine.
# Usage: bash infer_gpu.sh "text" prompt.wav "prompt transcript" out.wav

TEXT="${1:?text required}"
PROMPT_AUDIO="${2:?prompt audio path required}"
PROMPT_TEXT="${3:?prompt transcript required}"
OUTPUT="${4:-out.wav}"

MODEL_PATH="$(dirname "$0")/ghana-tts-36k-q8_0-audiovae-f16.gguf"   # recommended GPU variant
THREADS="${THREADS:-4}"

if ! command -v nvcc >/dev/null 2>&1; then
    echo "nvcc not found — this script requires a CUDA toolkit installed." >&2
    exit 1
fi

if [ ! -d VoxCPM.cpp ]; then
    git clone --depth 1 https://github.com/bluryar/VoxCPM.cpp.git
fi

if [ ! -f VoxCPM.cpp/build-cuda/examples/voxcpm_tts ]; then
    cmake -S VoxCPM.cpp -B VoxCPM.cpp/build-cuda -G Ninja \
        -DCMAKE_BUILD_TYPE=Release \
        -DVOXCPM_CUDA=ON \
        -DVOXCPM_BUILD_TESTS=OFF \
        -DVOXCPM_BUILD_WASM=OFF \
        -DVOXCPM_BUILD_BENCHMARK=OFF
    cmake --build VoxCPM.cpp/build-cuda --target voxcpm_tts -j"$(nproc)"
fi

VoxCPM.cpp/build-cuda/examples/voxcpm_tts \
    --text "$TEXT" \
    --prompt-audio "$PROMPT_AUDIO" \
    --prompt-text "$PROMPT_TEXT" \
    --output "$OUTPUT" \
    --model-path "$MODEL_PATH" \
    --backend cuda \
    --threads "$THREADS"

echo "Wrote $OUTPUT"