deploy-edge-ai-model
À propos
Cette compétence permet aux développeurs de déployer des modèles de ML sur des appareils périphériques tels que les téléphones et les systèmes IoT, en utilisant des frameworks comme TensorFlow Lite et ONNX Runtime. Elle couvre des techniques clés telles que la quantification des modèles et la sélection de délégués matériels pour optimiser les performances sous des contraintes de latence, de coût ou de connectivité. Utilisez-la lorsque l'inférence cloud est impraticable et que vous devez mettre en œuvre une inférence sur l'appareil avec des outils comme Google AI Edge Gallery.
Installation rapide
Claude Code
Recommandénpx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/deploy-edge-ai-modelCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Deploy Edge AI Model
See Extended Examples for complete configuration files, quantization scripts, and benchmark templates.
Deploy ML models to edge devices with optimized inference, hardware acceleration, and on-device model management.
When to Use
- Deploying LLMs (Gemma 4, Phi, Llama) to mobile devices via Google AI Edge Gallery
- Converting models to TensorFlow Lite or ONNX for on-device inference
- Quantizing models to INT8/INT4 for reduced memory and faster inference
- Building Android/iOS apps with local AI capabilities
- Selecting hardware delegates (GPU, NPU, DSP, Hexagon, CoreML)
- Benchmarking inference latency and memory on target devices
- Deploying MediaPipe tasks (vision, text, audio) to mobile or embedded platforms
Inputs
- Required: Trained model (SavedModel, PyTorch, ONNX, or Hugging Face checkpoint)
- Required: Target platform (Android, iOS, Linux embedded, browser)
- Required: Target device constraints (RAM, storage, compute capability)
- Optional: Calibration dataset for post-training quantization
- Optional: Google AI Edge Gallery configuration for LLM deployment
- Optional: Hardware delegate preferences (GPU, NPU, CPU-only)
Procedure
Step 1: Evaluate Model for Edge Deployment
Assess model size, latency requirements, and target device capabilities.
# 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
Edge deployment 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 |
Got: Model assessment completes, size and RAM ratios calculated, quantization recommendation generated based on device constraints.
If fail: Verify SavedModel path is valid (ls saved_model/), check TensorFlow installation (python -c "import tensorflow"), ensure sufficient disk space for model loading, verify model format is supported.
Step 2: Deploy LLMs via Google AI Edge Gallery
Use Google AI Edge Gallery to deploy Gemma 4 and other LLMs to Android devices.
# 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
Configure Gemma 4 model for AI Edge Gallery:
{
"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 on-device inference with 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)
Got: AI Edge Gallery app builds and installs successfully, Gemma 4 model downloads to device, on-device inference produces coherent responses, GPU delegate activates for acceleration.
If fail: Check Android SDK version >= 26 (adb shell getprop ro.build.version.sdk), verify device has sufficient storage for model download, ensure GPU delegate is supported (adb logcat | grep -i delegate), check Hugging Face model access permissions, verify ADB connection (adb devices).
Step 3: Convert and Quantize Models with TFLite
Convert standard models to TFLite format with post-training quantization.
# 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 quantization alternative:
# 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)
Got: TFLite model generated at specified path, model size reduced by 2-4x with INT8, inference accuracy within 1-2% of original, ONNX quantization produces valid model.
If fail: Check TensorFlow version >= 2.15 for latest quantization support, verify representative dataset matches model input shape, ensure all ops are supported in TFLite (converter.allow_custom_ops = True as fallback), check ONNX opset version compatibility.
Step 4: Configure Hardware Delegates
Select and configure hardware acceleration delegates for target devices.
# 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 selection 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 |
Got: Delegate loads without errors, inference runs on target accelerator, latency improves 2-10x over CPU-only depending on model and device.
If fail: Verify delegate library exists on device, check device supports requested delegate (adb shell cat /proc/cpuinfo for CPU features), fall back to XNNPACK if GPU/NPU unavailable, check OpenCL support for GPU delegate, verify NNAPI version (adb shell getprop ro.android.ndk.version).
Step 5: Benchmark On-Device Performance
Measure inference latency, memory usage, and power consumption on target devices.
# 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 benchmarking:
# 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")
Got: Benchmark produces latency percentiles, memory usage, and throughput metrics; GPU delegate shows 2-5x speedup over CPU for vision models; Gemma 4 2B achieves 10-30 tokens/sec on flagship phones.
If fail: Ensure benchmark binary matches device architecture (arm64-v8a), verify model pushed to device (adb shell ls /data/local/tmp/), check sufficient device storage, kill background apps to reduce memory pressure, verify thermal throttling not active (adb shell cat /sys/class/thermal/thermal_zone*/temp).
Step 6: Package for Production Deployment
Build the final mobile application with embedded or 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
}
}
Model download and caching strategy:
// 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
}
}
Got: Android app builds with MediaPipe dependency, model loads on first launch, inference runs within latency budget, model cached after first download, graceful fallback when device is unsupported.
If fail: Check minSdk >= 26 in build.gradle, verify MediaPipe dependency version, ensure model file not corrupted (check SHA256), verify sufficient device storage for model, check ProGuard rules preserve MediaPipe classes, test on multiple device tiers.
Validation
- Model converts to TFLite/ONNX without op compatibility errors
- Quantized model accuracy within acceptable tolerance (< 2% degradation)
- Hardware delegate loads and accelerates inference
- Benchmark latency meets target (e.g., < 100ms for vision, < 50ms/token for LLM)
- Memory usage stays within device budget
- AI Edge Gallery successfully loads and runs Gemma 4 model
- On-device LLM generates coherent responses
- Application handles model download, caching, and updates
- Graceful degradation on unsupported devices
- Battery impact within acceptable range for target use case
Pitfalls
- Unsupported ops in TFLite: Custom ops fail conversion - use
converter.allow_custom_ops = Trueor replace with supported alternatives, check op compatibility list - Quantization accuracy loss: INT4 degrades quality for sensitive tasks - use mixed precision, calibrate with representative data, evaluate on edge-specific test set
- Delegate initialization failure: GPU delegate crashes on older devices - always implement CPU fallback, check delegate compatibility before loading
- Memory pressure on device: Model + app exceeds available RAM - use memory-mapped models, implement model unloading, reduce batch size to 1
- Thermal throttling: Sustained inference causes device overheating - implement duty cycling, reduce inference frequency, monitor thermal zones
- Model download size: Large models over cellular data - offer Wi-Fi-only download, implement resumable downloads, use progressive model loading
- Version fragmentation: Model works on some devices but not others - test on representative device matrix, use NNAPI version checks, maintain device compatibility database
Related Skills
deploy-ml-model-serving- Cloud-based model serving (complement to edge)monitor-model-drift- Monitor model quality over timeregister-ml-model- Register models before edge deploymentcreate-dockerfile- Containerize edge model conversion pipelinecreate-multistage-dockerfile- Multi-stage builds for model conversion pipelines
Dépôt GitHub
Compétences associées
qmd
Développementqmd est un outil CLI de recherche et d'indexation locale qui permet aux développeurs d'indexer et de rechercher dans des fichiers locaux en utilisant une recherche hybride combinant BM25, des embeddings vectoriels et du reranking. Il prend en charge à la fois une utilisation en ligne de commande et un mode MCP (Model Context Protocol) pour l'intégration avec Claude. L'outil utilise Ollama pour les embeddings et stocke les index localement, ce qui le rend idéal pour rechercher dans de la documentation ou des bases de code directement depuis le terminal.
subagent-driven-development
DéveloppementCette compétence exécute des plans de mise en œuvre en déployant un nouveau sous-agent pour chaque tâche indépendante, avec une revue de code entre les tâches. Elle permet une itération rapide tout en maintenant des contrôles de qualité grâce à ce processus de revue. Utilisez-la lorsque vous travaillez sur des tâches principalement indépendantes au sein d'une même session pour assurer une progression continue avec des vérifications de qualité intégrées.
mcporter
DéveloppementLa compétence mcporter permet aux développeurs de gérer et d'appeler des serveurs Model Context Protocol (MCP) directement depuis Claude. Elle fournit des commandes pour lister les serveurs disponibles, appeler leurs outils avec des arguments, et gérer l'authentification ainsi que le cycle de vie du démon. Utilisez cette compétence pour intégrer et tester les fonctionnalités des serveurs MCP dans votre flux de travail de développement.
adk-deployment-specialist
DéveloppementCette compétence déploie et orchestre des agents Vertex AI ADK en utilisant le protocole A2A, gérant la découverte d'AgentCard, la soumission de tâches, et prenant en charge des outils tels que le bac à sable d'exécution de code et la banque de mémoire. Elle permet de construire des systèmes multi-agents avec des modèles d'orchestration séquentiels, parallèles ou en boucle en Python, Java ou Go. Utilisez-la lorsqu'on vous demande de déployer des agents ADK ou d'orchestrer des flux de travail d'agents sur Google Cloud.
