Meetily：Rust写的本地AI会议助手，数据永不离设备

GUI层（Next.js + Tauri WebView）
   ↓ IPC bridge（Tauri Command API，零序列化开销）
Rust核心层（audio capture → onnx-runtime → diarization → llm orchestration）
   ↓ FFI/HAL抽象层（metal-rs / vulkanalia / cuda-driver-sys）
硬件层（Apple Silicon Neural Engine / NVIDIA GPU / AMD GPU）

#!/bin/bash
## ./build-gpu.sh —— Meetily的GPU直连开关
set -e

## 自动探测GPU能力
if command -v nvidia-smi &> /dev/null; then
  echo "[INFO] NVIDIA GPU detected → building with CUDA backend"
  cargo build --release --features cuda
elif vulkaninfo &> /dev/null; then
  echo "[INFO] Vulkan device found → building with Vulkan backend"
  cargo build --release --features vulkan
else
  echo "[INFO] Falling back to CPU inference (ONNX Runtime CPU)"
  cargo build --release --features cpu
fi

pnpm run build:tauri

## crates/audio-engine/Cargo.toml
[dependencies.onnxruntime-rs]
version = "0.15"
default-features = false
features = [
  "cuda",      # 启用CUDA Driver API裸调
  "tensorrt",  # 可选，若系统装有TensorRT
  "parallel"   # 多流pipeline并行
]

// crates/audio-engine/src/transcriber.rs
pub struct ParakeetTranscriber {
    session: ort::Session, // ONNX Runtime Session，已绑定GPU device
    input_name: ort::AllocatedString,
}

impl ParakeetTranscriber {
    pub fn new(model_path: &Path) -> Result<Self> {
        let mut opts = ort::SessionOptions::new()?;
        opts.set_intra_op_num_threads(8)?;
        // ⚠️ 关键：显式设置GPU执行提供者
        #[cfg(feature = "cuda")]
        opts.add_cuda_provider(ort::CUDAProviderOptions {
            device_id: 0,
            arena_extend_strategy: ort::ArenaExtendStrategy::kSameAsRequested,
        })?;
        
        let session = ort::Session::new_with_options(model_path, opts)?;
        Ok(Self {
            session,
            input_name: session.input_names()[0].to_owned(),
        })
    }

    pub fn transcribe_stream(&mut self, pcm_chunk: &[f32]) -> Result<Vec<String>> {
        // 零拷贝：pcm_chunk直接映射为GPU内存视图
        let tensor = ort::Tensor::from_array(
            &self.session,
            pcm_chunk,
            &[1, pcm_chunk.len() as i64],
        )?;
        
        // 同步推理，但延迟<80ms（实测M2 Ultra下44.1kHz单通道）
        let outputs = self.session.run(ort::inputs! {
            self.input_name => tensor
        }?)?;
        
        self.parse_outputs(&outputs)
    }
}

git clone https://github.com/Zackriya-Solutions/meeting-minutes
cd meeting-minutes/frontend
pnpm install
## 自动识别Apple Silicon → 使用metal-rs
./build-gpu.sh
## 输出：target/release/app.app 已就绪

## 启动本地Qwen2服务（使用llama.cpp）
./server -m qwen2-7b.Q4_K_M.gguf -c 4096 --port 8080

## 启动Meetily并注入环境变量
export MEETILY_LLM_ENDPOINT="http://localhost:8080/v1"
export MEETILY_LLM_API_KEY="sk-no-key-required"
export MEETILY_LLM_MODEL="qwen2:7b"
open target/release/app.app

## Ubuntu 24.04
sudo apt install vulkan-tools mesa-vulkan-drivers vulkan-validationlayers
## 验证
vulkaninfo --summary | grep "deviceName\|apiVersion"
## 若输出含"AMD RADV"或"Intel ANV"，即可启用vulkan feature