部署邊緣 AI 模型

詳例見 Extended Examples。

部署 ML 模型至邊緣裝置，優化推理、硬件加速、本地模型管理。

用

• LLM（Gemma 4、Phi、Llama）→移動裝置 via AI Edge Gallery
• 模型→TFLite/ONNX 本地推理
• 量化 INT8/INT4 減內存+提速
• 構 Android/iOS 本地 AI 應用
• 選硬件代理（GPU、NPU、DSP、Hexagon、CoreML）
• 測目標設備時延+內存
• 部署 MediaPipe 任務（視、文、音）

入

• 必：訓好模型（SavedModel、PyTorch、ONNX、Hugging Face 檢查點）
• 必：目標平台（Android、iOS、嵌入 Linux、瀏覽器）
• 必：設備限（RAM、存儲、算力）
• 可：校準數據集
• 可：AI Edge Gallery 配
• 可：硬件代理偏好

法

一：評模型堪邊緣

察模大小、時延要求、目標設備力。

# assess_model.py
import os
import tensorflow as tf

def assess_model_for_edge(saved_model_path, target_ram_mb=4096):
    """Evaluate whether a model is suitable for edge deployment."""
    model = tf.saved_model.load(saved_model_path)

    # Check model size on disk
    model_size_mb = sum(
        os.path.getsize(os.path.join(dp, f))
        for dp, _, filenames in os.walk(saved_model_path)
        for f in filenames
    ) / (1024 * 1024)

    print(f"Model size: {model_size_mb:.1f} MB")
    print(f"Target RAM: {target_ram_mb} MB")
    print(f"Size/RAM ratio: {model_size_mb / target_ram_mb:.2%}")

    if model_size_mb > target_ram_mb * 0.25:
        print("WARNING: Model exceeds 25% of device RAM - quantization recommended")
        return False
    return True

邊緣部署決策矩陣：

Model Size	Device RAM	Recommended Action
< 50 MB	2+ GB	Direct TFLite conversion
50-500 MB	4+ GB	INT8 quantization + TFLite
500 MB-2 GB	6+ GB	INT4 quantization + AI Edge Gallery
2-4 GB	8+ GB	Gemma 4 via AI Edge Gallery with INT4
> 4 GB	12+ GB	Weight streaming or cloud-edge hybrid

得： 評估畢，大小+RAM 比計，量化建議生。

敗： 驗 SavedModel 路徑（ls saved_model/），查 TF 裝（python -c "import tensorflow"），確磁盤足，驗格式可支。

二：以 Google AI Edge Gallery 部署 LLM

部 Gemma 4 等至 Android。

# Clone AI Edge Gallery
git clone https://github.com/nickoala/ai-edge-gallery.git
cd ai-edge-gallery

# Build the Android app
./gradlew assembleDebug

# Install on connected device
adb install -r app/build/outputs/apk/debug/app-debug.apk

配 Gemma 4 模型：

{
  "models": [
    {
      "name": "Gemma 4 2B IT",
      "url": "https://huggingface.co/google/gemma-4-2b-it-gpu-int4",
      "format": "tflite",
      "backend": "gpu",
      "config": {
        "max_tokens": 1024,
        "temperature": 0.7,
        "top_k": 40,
        "top_p": 0.95
      }
    },
    {
      "name": "Gemma 4 4B IT",
      "url": "https://huggingface.co/google/gemma-4-4b-it-gpu-int4",
      "format": "tflite",
      "backend": "gpu",
      "config": {
        "max_tokens": 2048,
        "temperature": 0.7
      }
    }
  ]
}

以 LLM Inference API 程式化本地推理：

# gemma_edge_inference.py
from mediapipe.tasks.genai import llm_inference

# Configure the LLM
options = llm_inference.LlmInferenceOptions(
    model_path="/data/local/tmp/gemma-4-2b-it-int4.tflite",
    max_tokens=512,
    temperature=0.7,
    top_k=40,
    supported_lora_ranks=[4, 8, 16]  # Optional LoRA support
)

# Create inference engine
engine = llm_inference.LlmInference(options=options)

# Run inference
response = engine.generate_response("Explain edge computing in one sentence.")
print(response)

# Streaming inference
for chunk in engine.generate_response_async("List three benefits of on-device AI."):
    print(chunk, end="", flush=True)

得： Gallery 構+裝成，Gemma 4 下載至設備，本地推理生連貫應答，GPU 代理啟。

敗： 查 Android SDK ≥ 26（adb shell getprop ro.build.version.sdk），驗存儲足，確 GPU 代理可支（adb logcat | grep -i delegate），查 HF 訪問權，驗 ADB 連（adb devices）。

三：TFLite 轉+量化

標模→TFLite+訓後量化。

# convert_tflite.py
import os
import tensorflow as tf
import numpy as np

def convert_to_tflite(saved_model_path, output_path, quantization="dynamic"):
    """Convert SavedModel to TFLite with quantization."""
    converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_path)

    if quantization == "dynamic":
        converter.optimizations = [tf.lite.Optimize.DEFAULT]

    elif quantization == "int8":
        converter.optimizations = [tf.lite.Optimize.DEFAULT]
        converter.target_spec.supported_ops = [
            tf.lite.OpsSet.TFLITE_BUILTINS_INT8
        ]
        converter.inference_input_type = tf.int8
        converter.inference_output_type = tf.int8

        # Representative dataset for calibration
        def representative_dataset():
            for _ in range(100):
                yield [np.random.randn(1, 224, 224, 3).astype(np.float32)]
        converter.representative_dataset = representative_dataset

    elif quantization == "float16":
        converter.optimizations = [tf.lite.Optimize.DEFAULT]
        converter.target_spec.supported_types = [tf.float16]

    tflite_model = converter.convert()

    with open(output_path, "wb") as f:
        f.write(tflite_model)

    original_size = sum(
        os.path.getsize(os.path.join(dp, f))
        for dp, _, filenames in os.walk(saved_model_path)
        for f in filenames
    ) / (1024 * 1024)
    quantized_size = len(tflite_model) / (1024 * 1024)
    print(f"Original: {original_size:.1f} MB -> Quantized: {quantized_size:.1f} MB")
    print(f"Compression ratio: {original_size / quantized_size:.1f}x")

# Usage
convert_to_tflite("saved_model/", "model_int8.tflite", quantization="int8")

ONNX Runtime 量化替代：

# quantize_onnx.py
from onnxruntime.quantization import quantize_dynamic, quantize_static, QuantType

# Dynamic quantization (no calibration data needed)
quantize_dynamic(
    model_input="model.onnx",
    model_output="model_int8.onnx",
    weight_type=QuantType.QInt8
)

# Static quantization (better accuracy, needs calibration)
# ... (see EXAMPLES.md for complete calibration workflow)

得： TFLite 模生於指定路，INT8 縮 2-4x，精度失 1-2% 內，ONNX 量化有效。

敗： 查 TF ≥ 2.15，驗代表數據集入形匹，確所有 op 可支（converter.allow_custom_ops = True 備），查 ONNX opset 相容。

四：配硬件代理

擇配目標設備之硬件加速。

# configure_delegates.py
import tensorflow as tf

def create_interpreter_with_delegate(model_path, delegate="gpu"):
    """Create TFLite interpreter with hardware delegate."""

    if delegate == "gpu":
        delegate_obj = tf.lite.experimental.load_delegate(
            "libtensorflowlite_gpu_delegate.so",
            options={"precision": "fp16", "allow_quantized_models": "true"}
        )
    elif delegate == "nnapi":
        # Android Neural Networks API - routes to NPU/DSP
        delegate_obj = tf.lite.experimental.load_delegate(
            "libtensorflowlite_nnapi_delegate.so"
        )
    elif delegate == "xnnpack":
        # Optimized CPU inference
        delegate_obj = None  # XNNPACK is default in TFLite

    interpreter = tf.lite.Interpreter(
        model_path=model_path,
        experimental_delegates=[delegate_obj] if delegate_obj else None,
        num_threads=4
    )
    interpreter.allocate_tensors()
    return interpreter

代理擇指南：

Device	Best Delegate	Fallback	Notes
Android (Qualcomm)	NNAPI -> Hexagon DSP	GPU -> XNNPACK	Check nnapi_accelerator_name
Android (MediaTek)	NNAPI -> APU	GPU -> XNNPACK	Dimensity chips have dedicated APU
Android (Samsung)	NNAPI -> NPU	GPU -> XNNPACK	Exynos NPU via NNAPI
iOS	CoreML delegate	Metal GPU	Use coreml_delegate for ANE
Linux embedded	GPU (if available)	XNNPACK	RPi uses XNNPACK CPU
Browser	WebGL / WebGPU	WASM SIMD	Via TensorFlow.js

得： 代理載無錯，推理於目標加速器跑，時延較 CPU-only 快 2-10x。

敗： 驗代理庫存於設備，查設備支持（adb shell cat /proc/cpuinfo），GPU/NPU 不可用→退 XNNPACK，查 OpenCL 支持，驗 NNAPI 版本（adb shell getprop ro.android.ndk.version）。

五：測本地性能

測目標設備推理時延、內存、能耗。

# Use TFLite benchmark tool
adb push model_int8.tflite /data/local/tmp/

# CPU benchmark
adb shell /data/local/tmp/benchmark_model \
  --graph=/data/local/tmp/model_int8.tflite \
  --num_threads=4 \
  --num_runs=50 \
  --warmup_runs=5

# GPU benchmark
adb shell /data/local/tmp/benchmark_model \
  --graph=/data/local/tmp/model_int8.tflite \
  --use_gpu=true \
  --num_runs=50

# NNAPI benchmark
adb shell /data/local/tmp/benchmark_model \
  --graph=/data/local/tmp/model_int8.tflite \
  --use_nnapi=true \
  --nnapi_accelerator_name=google-edgetpu \
  --num_runs=50

Python 測：

# benchmark_edge.py
import time
import numpy as np
import psutil

def benchmark_inference(interpreter, input_data, num_runs=100):
    """Benchmark TFLite model inference."""
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()

    # Warmup
    for _ in range(10):
        interpreter.set_tensor(input_details[0]["index"], input_data)
        interpreter.invoke()

    # Benchmark
    latencies = []
    mem_before = psutil.Process().memory_info().rss / (1024 * 1024)
    for _ in range(num_runs):
        start = time.perf_counter()
        interpreter.set_tensor(input_details[0]["index"], input_data)
        interpreter.invoke()
        latencies.append((time.perf_counter() - start) * 1000)
    mem_after = psutil.Process().memory_info().rss / (1024 * 1024)

    print(f"Latency (p50): {np.percentile(latencies, 50):.1f} ms")
    print(f"Latency (p95): {np.percentile(latencies, 95):.1f} ms")
    print(f"Latency (p99): {np.percentile(latencies, 99):.1f} ms")
    print(f"Memory delta: {mem_after - mem_before:.1f} MB")
    print(f"Throughput: {1000 / np.mean(latencies):.1f} inferences/sec")

得： 測生時延百分位、內存、吞吐；視覺模型 GPU 代理較 CPU 快 2-5x；Gemma 4 2B 旗艦機達 10-30 tokens/sec。

敗： 確測二進制匹設備架構（arm64-v8a），驗模型推至設備（adb shell ls /data/local/tmp/），查存儲足，殺背景應用減內存壓，驗熱節流未啟（adb shell cat /sys/class/thermal/thermal_zone*/temp）。

六：打包生產部署

構終移動應用，含內嵌或可下載模型。

// Android: EdgeAIManager.kt
import com.google.mediapipe.tasks.genai.llminference.LlmInference

class EdgeAIManager(private val context: Context) {
    private var llmInference: LlmInference? = null

    fun initialize(modelPath: String) {
        val options = LlmInference.LlmInferenceOptions.builder()
            .setModelPath(modelPath)
            .setMaxTokens(512)
            .setTemperature(0.7f)
            .setTopK(40)
            .setResultListener { result, done ->
                // Handle streaming tokens
                onTokenReceived(result, done)
            }
            .build()

        llmInference = LlmInference.createFromOptions(context, options)
    }

    fun generateResponse(prompt: String): String {
        return llmInference?.generateResponse(prompt)
            ?: throw IllegalStateException("Model not initialized")
    }

    fun release() {
        llmInference?.close()
        llmInference = null
    }
}

模型下載+緩存策略：

// ModelDownloader.kt
class ModelDownloader(private val context: Context) {
    private val modelDir = File(context.filesDir, "models")

    suspend fun ensureModel(modelName: String, url: String): File {
        val modelFile = File(modelDir, modelName)
        if (modelFile.exists()) return modelFile

        modelDir.mkdirs()
        // Download with progress tracking
        // ... (see EXAMPLES.md for complete implementation)
        return modelFile
    }
}

得： Android 應用以 MediaPipe 依賴構，首啟載模型，推理於時延預算內，首下後緩存，不支設備優雅退。

敗： 查 build.gradle minSdk ≥ 26，驗 MediaPipe 版本，確模檔未損（查 SHA256），驗存儲足，查 ProGuard 規則保 MediaPipe 類，測多級設備。

驗

• 模型轉 TFLite/ONNX 無 op 相容錯
• 量化精度於可容忍內（< 2% 劣化）
• 硬件代理載並加速
• 測時延達目標（視覺 < 100ms，LLM < 50ms/token）
• 內存於設備預算內
• AI Edge Gallery 成載跑 Gemma 4
• 本地 LLM 生連貫應答
• 應用處模下載、緩存、更新
• 不支設備優雅退
• 電耗於用例可容範內

忌

• TFLite op 不支：自定義 op 轉失→converter.allow_custom_ops = True 或替，查相容表
• 量化精度失：INT4 敏感任務劣→混精、代表數據校準、邊緣專測
• 代理初始化失敗：GPU 代理舊機崩→常備 CPU 退，載前查相容
• 設備內存壓：模+應超 RAM→用 mmap 模、卸載、批 1
• 熱節流：續推過熱→工作週期、降頻、監熱區
• 模下載大：蜂窩大模→限 WiFi、續傳、漸進載
• 版本碎片：此機行彼不行→代表機陣測、NNAPI 版檢、維相容庫

參

• deploy-ml-model-serving
• monitor-model-drift
• register-ml-model
• create-dockerfile
• create-multistage-dockerfile