Deploy Edge AI Model

See Extended Examples for complete configuration files, quantization scripts, and benchmark templates.

ML → edge devices. Optimized inference, HW accel, on-device mgmt.

Use When

• LLMs (Gemma 4, Phi, Llama) → mobile via Google AI Edge Gallery
• Convert → TFLite/ONNX for on-device
• Quantize → INT8/INT4, less mem + faster
• Android/iOS apps w/ local AI
• HW delegate select (GPU, NPU, DSP, Hexagon, CoreML)
• Bench latency + mem on target
• MediaPipe tasks → mobile/embedded

In

• Required: Trained model (SavedModel, PyTorch, ONNX, HF checkpoint)
• Required: Target platform (Android, iOS, Linux embedded, browser)
• Required: Device constraints (RAM, storage, compute)
• Optional: Calibration dataset → post-training quant
• Optional: AI Edge Gallery config → LLM deploy
• Optional: HW delegate prefs

Do

Step 1: Eval model → edge

Size, latency, device cap.

# 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

Decision matrix:

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

→ Assessment done, size/RAM ratios, quant recommendation by constraints.

If err: SavedModel path valid (ls saved_model/), TF installed (python -c "import tensorflow"), disk space OK, format supported.

Step 2: LLMs via Google AI Edge Gallery

Gemma 4 + LLMs → 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

Gemma 4 config:

{
  "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
      }
    }
  ]
}

Programmatic inference w/ LLM Inference API:

# 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)

→ App builds+installs, Gemma 4 downloads, coherent responses, GPU delegate active.

If err: SDK ≥ 26 (adb shell getprop ro.build.version.sdk), device storage OK, GPU delegate supported (adb logcat | grep -i delegate), HF access, ADB connection (adb devices).

Step 3: Convert + quantize w/ TFLite

Standard → TFLite w/ post-training quant.

# 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 quant alt:

# 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)

→ TFLite gen'd, size -2-4x w/ INT8, accuracy within 1-2%, ONNX quant valid.

If err: TF ≥ 2.15, rep dataset matches input shape, all ops supported (converter.allow_custom_ops = True fallback), ONNX opset compat.

Step 4: HW delegates

Select + config.

# 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

Delegate guide:

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

→ Delegate loads, inference on accel, latency 2-10x vs CPU-only.

If err: Lib on device, delegate supported (adb shell cat /proc/cpuinfo), fall back XNNPACK, OpenCL for GPU, NNAPI ver.

Step 5: Bench on-device

Latency, mem, power.

# 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 bench:

# 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")

→ Latency percentiles + mem + throughput. GPU 2-5x vs CPU. Gemma 4 2B → 10-30 tok/sec flagship.

If err: Bench binary matches arch (arm64-v8a), model pushed (adb shell ls /data/local/tmp/), storage OK, kill bg apps, thermal throttle check (adb shell cat /sys/class/thermal/thermal_zone*/temp).

Step 6: Package → prod

Mobile app w/ embedded/downloadable model.

// 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
    }
}

Download + cache:

// 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
    }
}

→ App builds w/ MediaPipe, model loads first launch, latency OK, cached after download, fallback on unsupported.

If err: minSdk ≥ 26, MediaPipe dep ver, model SHA256, storage, ProGuard preserves MediaPipe classes, test multi-device.

Check

• Model → TFLite/ONNX w/o op errs
• Quant accuracy < 2% degrade
• HW delegate loads + accels
• Latency meets target (< 100ms vision, < 50ms/tok LLM)
• Mem within budget
• AI Edge Gallery runs Gemma 4
• On-device LLM coherent
• App handles download/cache/update
• Graceful degrade on unsupported
• Battery acceptable

Traps

• Unsupported TFLite ops: Custom ops fail → converter.allow_custom_ops = True or replace, check compat list
• Quant accuracy loss: INT4 degrades sensitive → mixed precision, calibrate w/ rep data
• Delegate init fail: GPU crashes old devices → CPU fallback, check compat
• Mem pressure: Model + app > RAM → memory-mapped, unload, batch=1
• Thermal throttle: Sustained inference → overheat → duty cycle, reduce freq, monitor zones
• Download size: Large over cellular → Wi-Fi-only, resumable, progressive
• Version fragmentation: Works some not others → device matrix test, NNAPI ver checks, compat DB

→

• deploy-ml-model-serving — cloud serving (complement to edge)
• monitor-model-drift — quality over time
• register-ml-model — register before edge deploy
• create-dockerfile — containerize conversion pipeline
• create-multistage-dockerfile — multi-stage builds