¿Por qué CoreML en Mac Mini M4?
CoreML es el framework nativo de machine learning de Apple, diseñado específicamente para extraer el máximo rendimiento de Apple Silicon. A diferencia de frameworks ML genéricos que tratan la GPU como un único dispositivo de cómputo, CoreML distribuye inteligentemente las cargas de trabajo entre la CPU, GPU y el Neural Engine dedicado de 16 núcleos, ejecutando a menudo diferentes partes de un modelo en diferentes unidades de cómputo simultáneamente.
Neural Engine de 16 núcleos
El Neural Engine del M4 ofrece hasta 38 TOPS (billones de operaciones por segundo) para cargas de trabajo int8 cuantizadas. CoreML enruta automáticamente las capas compatibles (convoluciones, multiplicaciones de matrices, normalización) al Neural Engine para máximo rendimiento con mínimo consumo energético.
Arquitectura de memoria unificada
Todas las unidades de cómputo comparten el mismo pool de memoria con hasta 120 GB/s de ancho de banda. No hay cuello de botella PCIe ni copia de datos entre la memoria de CPU y GPU. Un Mac Mini M4 de 24GB da a cada unidad de cómputo acceso directo a los 24GB completos, permitiendo modelos más grandes que configuraciones equivalentes con GPU discreta.
Despacho automático de cómputo
El compilador de CoreML analiza el grafo de su modelo y asigna cada operación a la unidad de cómputo óptima. Las convoluciones se ejecutan en el Neural Engine, las operaciones personalizadas recurren a GPU o CPU, y todo se ejecuta como un pipeline unificado. Obtiene optimización a nivel de hardware sin esfuerzo manual.
Eficiencia energética a escala
Un Mac Mini M4 ejecutando inferencia CoreML consume 5-20W de potencia total del sistema. Compare eso con 300-450W para un servidor GPU NVIDIA A100. Para inferencia de producción siempre activa, esto se traduce en costes de electricidad drásticamente menores y sin necesidad de infraestructura de refrigeración especializada.
Punto clave: CoreML no es solo para apps iOS. Con bindings de Python a través de coremltools, puede convertir modelos de cualquier framework principal, ejecutar inferencia desde scripts Python y construir APIs de producción, todo aprovechando el Neural Engine que la mayoría de los frameworks del lado del servidor no pueden acceder.
Convertir modelos PyTorch a CoreML
La biblioteca coremltools de Apple proporciona un camino directo de conversión desde modelos PyTorch al formato .mlpackage de CoreML. La conversión traza su modelo con entradas de ejemplo y traduce cada operación a la representación interna de CoreML.
Paso 1: Instalar dependencias
# Create a virtual environment
python3 -m venv ~/coreml-env
source ~/coreml-env/bin/activate
# Install coremltools and PyTorch
pip install coremltools torch torchvision
# Verify installation
python3 -c "import coremltools as ct; print(ct.__version__)"
# 9.0
Paso 2: Convertir un clasificador de imágenes PyTorch
import torch
import torchvision
import coremltools as ct
# Load a pretrained ResNet50 model
model = torchvision.models.resnet50(weights=torchvision.models.ResNet50_Weights.DEFAULT)
model.eval()
# Create a sample input (batch=1, channels=3, height=224, width=224)
example_input = torch.randn(1, 3, 224, 224)
# Trace the model with TorchScript
traced_model = torch.jit.trace(model, example_input)
# Convert to CoreML
coreml_model = ct.convert(
traced_model,
inputs=[ct.ImageType(
name="image",
shape=(1, 3, 224, 224),
scale=1.0 / (255.0 * 0.226),
bias=[-0.485 / 0.226, -0.456 / 0.226, -0.406 / 0.226],
color_layout=ct.colorlayout.RGB
)],
classifier_config=ct.ClassifierConfig("imagenet_classes.txt"),
compute_units=ct.ComputeUnit.ALL, # Use Neural Engine + GPU + CPU
minimum_deployment_target=ct.target.macOS15,
)
# Save the model
coreml_model.save("ResNet50.mlpackage")
print("Model saved: ResNet50.mlpackage")
Paso 3: Convertir un modelo PyTorch personalizado
import torch
import torch.nn as nn
import coremltools as ct
# Define a custom text embedding model
class TextEncoder(nn.Module):
def __init__(self, vocab_size=30000, embed_dim=512, num_heads=8, num_layers=6):
super().__init__()
self.embedding = nn.Embedding(vocab_size, embed_dim)
self.pos_encoding = nn.Parameter(torch.randn(1, 512, embed_dim))
encoder_layer = nn.TransformerEncoderLayer(
d_model=embed_dim, nhead=num_heads, batch_first=True
)
self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
self.output_proj = nn.Linear(embed_dim, 256)
def forward(self, input_ids):
x = self.embedding(input_ids) + self.pos_encoding[:, :input_ids.shape[1], :]
x = self.transformer(x)
x = x.mean(dim=1) # Global average pooling
return self.output_proj(x)
# Initialize and load your trained weights
model = TextEncoder()
# model.load_state_dict(torch.load("text_encoder_weights.pt"))
model.eval()
# Trace with example input
example_input = torch.randint(0, 30000, (1, 128))
traced_model = torch.jit.trace(model, example_input)
# Convert to CoreML
coreml_model = ct.convert(
traced_model,
inputs=[ct.TensorType(name="input_ids", shape=(1, 128), dtype=int)],
outputs=[ct.TensorType(name="embedding")],
compute_units=ct.ComputeUnit.ALL,
minimum_deployment_target=ct.target.macOS15,
)
# Add metadata
coreml_model.author = "My Remote Mac"
coreml_model.short_description = "Text embedding model for semantic search"
coreml_model.version = "1.0.0"
coreml_model.save("TextEncoder.mlpackage")
print("Model saved: TextEncoder.mlpackage")
Paso 4: Verificar el modelo convertido
import coremltools as ct
import numpy as np
# Load the converted model
model = ct.models.MLModel("ResNet50.mlpackage")
# Inspect model metadata
spec = model.get_spec()
print(f"Model type: {spec.WhichOneof('Type')}")
print(f"Inputs: {[inp.name for inp in spec.description.input]}")
print(f"Outputs: {[out.name for out in spec.description.output]}")
# Run a test prediction
from PIL import Image
img = Image.open("test_image.jpg").resize((224, 224))
prediction = model.predict({"image": img})
print(f"Top prediction: {prediction}")
# Check which compute units are available
print(f"Compute units: {model.compute_unit}")
Convertir modelos TensorFlow a CoreML
CoreML convierte modelos de TensorFlow: directorios SavedModel y modelos Keras (un modelo tf.keras o un archivo .h5). Los archivos TensorFlow Lite .tflite no se pueden convertir: parta del modelo TensorFlow original.
Convertir un TensorFlow SavedModel
import coremltools as ct
import tensorflow as tf
# Load a TensorFlow SavedModel (e.g., EfficientNet trained on your data)
tf_model = tf.keras.applications.EfficientNetV2S(
weights="imagenet",
input_shape=(384, 384, 3)
)
# Convert to CoreML
coreml_model = ct.convert(
tf_model,
inputs=[ct.ImageType(
name="image",
shape=(1, 384, 384, 3),
scale=1.0 / 255.0,
color_layout=ct.colorlayout.RGB
)],
compute_units=ct.ComputeUnit.ALL,
minimum_deployment_target=ct.target.macOS15,
)
coreml_model.save("EfficientNetV2S.mlpackage")
print("Saved EfficientNetV2S.mlpackage")
Convertir un modelo Keras H5
import coremltools as ct
import tensorflow as tf
# Load your custom Keras model
model = tf.keras.models.load_model("my_custom_model.h5")
# Print model summary to understand input/output shapes
model.summary()
# Convert with explicit input/output specifications
coreml_model = ct.convert(
model,
inputs=[ct.TensorType(name="features", shape=(1, 128))],
outputs=[ct.TensorType(name="prediction")],
compute_units=ct.ComputeUnit.ALL,
minimum_deployment_target=ct.target.macOS15,
)
# Add model metadata
coreml_model.author = "ML Team"
coreml_model.license = "Proprietary"
coreml_model.short_description = "Customer churn prediction model v2.1"
coreml_model.version = "2.1.0"
coreml_model.save("ChurnPredictor.mlpackage")
Modelos ONNX: convertir desde el framework de origen
coremltools ya no convierte archivos ONNX: su conversor onnx-coreml está congelado y ct.convert solo acepta modelos de TensorFlow y PyTorch (una ruta .onnx falla con "Unable to determine the type of the model"). Si un modelo le llega en ONNX, conviértalo desde el framework que lo generó. Dos casos habituales:
Exportar YOLOv8 directamente a CoreML
# Ultralytics exports its YOLO models to CoreML itself (it runs coremltools for you)
pip install ultralytics
# Writes yolov8n.mlpackage, with non-maximum suppression built in
yolo export model=yolov8n.pt format=coreml imgsz=640 nms=True
Convertir el modelo PyTorch original en lugar de su exportación ONNX
import torch
import torchvision
import coremltools as ct
# Start from the PyTorch model you would have exported to ONNX
model = torchvision.models.mobilenet_v3_large(
weights=torchvision.models.MobileNet_V3_Large_Weights.DEFAULT
).eval()
example_input = torch.rand(1, 3, 224, 224)
# Trace it (ct.convert also accepts a torch.export ExportedProgram)
traced_model = torch.jit.trace(model, example_input)
# Add your model's input normalization as scale/bias, as in Step 2
coreml_model = ct.convert(
traced_model,
inputs=[ct.ImageType(name="image", shape=(1, 3, 224, 224), scale=1.0 / 255.0)],
compute_units=ct.ComputeUnit.ALL,
minimum_deployment_target=ct.target.macOS15,
)
coreml_model.save("MobileNetV3.mlpackage")
print("Saved MobileNetV3.mlpackage")
Optimizar para Neural Engine
El Neural Engine ofrece el máximo rendimiento con modelos cuantizados. Aplicar cuantización post-entrenamiento, paletización y poda puede reducir el tamaño del modelo en 4-8x y mejorar el rendimiento del Neural Engine en 2-4x, a menudo con una pérdida de precisión insignificante.
Float16: ya es la precisión por defecto de los ML Programs
import os
import torch
import torchvision
import coremltools as ct
# ML Programs (.mlpackage) are typed in float16 by default: the ResNet50 from
# Step 2 already stores float16 weights, so no extra quantization pass is needed.
# To see what float16 saves, convert the same model once in float32:
model = torchvision.models.resnet50(weights=torchvision.models.ResNet50_Weights.DEFAULT).eval()
traced_model = torch.jit.trace(model, torch.randn(1, 3, 224, 224))
model_fp32 = ct.convert(
traced_model,
inputs=[ct.ImageType(name="image", shape=(1, 3, 224, 224))],
compute_precision=ct.precision.FLOAT32,
minimum_deployment_target=ct.target.macOS15,
)
model_fp32.save("ResNet50_fp32.mlpackage")
# Check file sizes
def size_on_disk(path):
return sum(
os.path.getsize(os.path.join(dp, f))
for dp, dn, fn in os.walk(path) for f in fn
)
fp32_size = size_on_disk("ResNet50_fp32.mlpackage")
fp16_size = size_on_disk("ResNet50.mlpackage")
print(f"Float32: {fp32_size / 1e6:.1f} MB")
print(f"Float16 (default): {fp16_size / 1e6:.1f} MB")
print(f"Reduction: {(1 - fp16_size / fp32_size) * 100:.1f}%")
Cuantización post-entrenamiento Int8
import coremltools as ct
import coremltools.optimize as cto
from pathlib import Path
from PIL import Image
# Load the model
model = ct.models.MLModel("ResNet50.mlpackage")
# Configure linear (int8) quantization
op_config = cto.coreml.OpLinearQuantizerConfig(
mode="linear_symmetric",
dtype="int8",
granularity="per_channel"
)
config = cto.coreml.OptimizationConfig(global_config=op_config)
# Weight-only int8 quantization: no calibration data needed
model_int8 = cto.coreml.linear_quantize_weights(model, config=config)
model_int8.save("ResNet50_int8.mlpackage")
print("Saved int8 quantized model")
# Optional: quantize activations too (W8A8), calibrated on real samples.
# Samples use the same format as model.predict(): here, the "image" input.
def load_calibration_data():
"""Load 100-200 representative images for calibration."""
return [
{"image": Image.open(path).convert("RGB").resize((224, 224))}
for path in sorted(Path("calibration_images").glob("*.jpg"))[:200]
]
activation_config = cto.coreml.OptimizationConfig(
global_config=cto.coreml.OpLinearQuantizerConfig(mode="linear_symmetric")
)
model_a8 = cto.coreml.linear_quantize_activations(
model, activation_config, load_calibration_data()
)
model_w8a8 = cto.coreml.linear_quantize_weights(model_a8, config=config)
model_w8a8.save("ResNet50_w8a8.mlpackage")
print("Saved W8A8 quantized model")
Paletización (Clustering de pesos)
import coremltools as ct
import coremltools.optimize as cto
# Palettization clusters weights into a small lookup table
# 4-bit palettization = 16 unique weight values per tensor
# Achieves ~4x compression with minimal accuracy loss
model = ct.models.MLModel("ResNet50.mlpackage")
# Configure palettization
op_config = cto.coreml.OpPalettizerConfig(
mode="kmeans",
nbits=4, # 4-bit = 16 clusters, 2-bit = 4 clusters
granularity="per_tensor"
)
config = cto.coreml.OptimizationConfig(global_config=op_config)
# Apply palettization
model_palettized = cto.coreml.palettize_weights(model, config=config)
model_palettized.save("ResNet50_palettized_4bit.mlpackage")
print("4-bit palettized model saved")
print("This model runs optimally on the Neural Engine")
Poda (Esparcidad)
import coremltools as ct
import coremltools.optimize as cto
# Pruning sets small weights to zero, enabling sparse computation
# The Neural Engine can skip zero-weight operations for speed gains
model = ct.models.MLModel("ResNet50.mlpackage")
# Configure magnitude-based pruning
op_config = cto.coreml.OpMagnitudePrunerConfig(
target_sparsity=0.75, # Remove 75% of smallest weights
block_size=None # Unstructured pruning
)
config = cto.coreml.OptimizationConfig(global_config=op_config)
# Apply pruning
model_pruned = cto.coreml.prune_weights(model, config=config)
model_pruned.save("ResNet50_pruned_75.mlpackage")
print("75% sparse model saved")
Pipeline de optimización combinada
import coremltools as ct
import coremltools.optimize as cto
# For maximum optimization, combine pruning + palettization + quantization
# This can achieve 8-16x compression with 1-2% accuracy loss
model = ct.models.MLModel("ResNet50.mlpackage")
# Step 1: Prune (set small weights to zero)
prune_config = cto.coreml.OptimizationConfig(
global_config=cto.coreml.OpMagnitudePrunerConfig(target_sparsity=0.5)
)
model = cto.coreml.prune_weights(model, config=prune_config)
print("Step 1: Pruning complete (50% sparsity)")
# Step 2: Palettize (cluster remaining weights)
palette_config = cto.coreml.OptimizationConfig(
global_config=cto.coreml.OpPalettizerConfig(mode="kmeans", nbits=4)
)
model = cto.coreml.palettize_weights(model, config=palette_config)
print("Step 2: Palettization complete (4-bit)")
# Save the fully optimized model
model.save("ResNet50_optimized.mlpackage")
print("Fully optimized model saved -- ready for Neural Engine deployment")
Consejo: Siempre haga benchmarks de precisión después de la optimización. Comience con float16 (más seguro), luego pruebe cuantización int8, luego paletización. Use un conjunto de validación reservado y defina un umbral de precisión aceptable antes de aplicar optimizaciones agresivas.
Construir una API REST para inferencia CoreML
Envolver su modelo CoreML en una API REST lo hace accesible para cualquier cliente: apps web, apps móviles, microservicios o pipelines de procesamiento por lotes. A continuación se muestran ejemplos listos para producción usando tanto Flask como FastAPI.
Opción A: Servidor API Flask
# flask_coreml_server.py
# pip install flask pillow coremltools gunicorn
import io
import time
import coremltools as ct
from flask import Flask, request, jsonify
from PIL import Image
app = Flask(__name__)
# Load the CoreML model at startup (runs on Neural Engine)
print("Loading CoreML model...")
model = ct.models.MLModel(
"ResNet50_optimized.mlpackage",
compute_units=ct.ComputeUnit.ALL
)
print("Model loaded successfully")
@app.route("/health", methods=["GET"])
def health():
return jsonify({"status": "healthy", "model": "ResNet50"})
@app.route("/predict", methods=["POST"])
def predict():
if "image" not in request.files:
return jsonify({"error": "No image file provided"}), 400
# Read and preprocess the image
image_file = request.files["image"]
image = Image.open(io.BytesIO(image_file.read())).resize((224, 224))
# Run inference with timing
start = time.perf_counter()
prediction = model.predict({"image": image})
latency_ms = (time.perf_counter() - start) * 1000
return jsonify({
"prediction": prediction,
"latency_ms": round(latency_ms, 2),
"compute_unit": "neural_engine+gpu+cpu"
})
@app.route("/predict/batch", methods=["POST"])
def predict_batch():
"""Process multiple images in a single request."""
if "images" not in request.files:
return jsonify({"error": "No image files provided"}), 400
results = []
files = request.files.getlist("images")
start = time.perf_counter()
for image_file in files:
image = Image.open(io.BytesIO(image_file.read())).resize((224, 224))
prediction = model.predict({"image": image})
results.append(prediction)
total_ms = (time.perf_counter() - start) * 1000
return jsonify({
"predictions": results,
"total_latency_ms": round(total_ms, 2),
"images_processed": len(results),
"avg_latency_ms": round(total_ms / len(results), 2)
})
if __name__ == "__main__":
app.run(host="0.0.0.0", port=5000)
# Production: gunicorn flask_coreml_server:app -w 2 -b 0.0.0.0:5000 --timeout 120
Opción B: Servidor FastAPI (Async + Documentación OpenAPI)
# fastapi_coreml_server.py
# pip install fastapi uvicorn python-multipart pillow coremltools
import io
import time
import asyncio
from concurrent.futures import ThreadPoolExecutor
import coremltools as ct
from fastapi import FastAPI, File, UploadFile, HTTPException
from fastapi.responses import JSONResponse
from PIL import Image
from typing import List
app = FastAPI(
title="CoreML Inference API",
description="Production CoreML model serving on Mac Mini M4",
version="1.0.0"
)
# Load model at startup
model = ct.models.MLModel(
"ResNet50_optimized.mlpackage",
compute_units=ct.ComputeUnit.ALL
)
# Thread pool for blocking CoreML calls
executor = ThreadPoolExecutor(max_workers=4)
def run_prediction(image_bytes: bytes) -> dict:
"""Run CoreML prediction in a thread (blocking call)."""
image = Image.open(io.BytesIO(image_bytes)).resize((224, 224))
start = time.perf_counter()
result = model.predict({"image": image})
latency = (time.perf_counter() - start) * 1000
return {"prediction": result, "latency_ms": round(latency, 2)}
@app.get("/health")
async def health():
return {"status": "healthy", "model": "ResNet50_optimized", "engine": "CoreML"}
@app.post("/predict")
async def predict(image: UploadFile = File(...)):
if not image.content_type.startswith("image/"):
raise HTTPException(status_code=400, detail="File must be an image")
image_bytes = await image.read()
loop = asyncio.get_event_loop()
result = await loop.run_in_executor(executor, run_prediction, image_bytes)
return result
@app.post("/predict/batch")
async def predict_batch(images: List[UploadFile] = File(...)):
loop = asyncio.get_event_loop()
tasks = []
for img in images:
image_bytes = await img.read()
tasks.append(loop.run_in_executor(executor, run_prediction, image_bytes))
results = await asyncio.gather(*tasks)
return {
"predictions": list(results),
"total_images": len(results)
}
# Run: uvicorn fastapi_coreml_server:app --host 0.0.0.0 --port 8000 --workers 2
Probar la API
# Test single prediction
curl -X POST http://localhost:8000/predict \
-F "image=@test_image.jpg"
# Test batch prediction
curl -X POST http://localhost:8000/predict/batch \
-F "images=@image1.jpg" \
-F "images=@image2.jpg" \
-F "images=@image3.jpg"
# Health check
curl http://localhost:8000/health
# Python client example
import requests
with open("test_image.jpg", "rb") as f:
response = requests.post(
"http://your-mac-mini:8000/predict",
files={"image": f}
)
print(response.json())
# {"prediction": {"classLabel": "golden_retriever", "confidence": 0.94}, "latency_ms": 3.2}
Servicio tipo Systemd con launchd
# Create a LaunchAgent: the API starts with your user session and restarts
# if it crashes (enable automatic login so it also returns after a reboot)
mkdir -p /Users/admin/coreml-api/logs
cat <<EOF > ~/Library/LaunchAgents/com.coreml.api.plist
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
"http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key>
<string>com.coreml.api</string>
<key>ProgramArguments</key>
<array>
<string>/Users/admin/coreml-env/bin/uvicorn</string>
<string>fastapi_coreml_server:app</string>
<string>--host</string>
<string>0.0.0.0</string>
<string>--port</string>
<string>8000</string>
<string>--workers</string>
<string>2</string>
</array>
<key>WorkingDirectory</key>
<string>/Users/admin/coreml-api</string>
<key>RunAtLoad</key>
<true/>
<key>KeepAlive</key>
<true/>
<key>StandardOutPath</key>
<string>/Users/admin/coreml-api/logs/coreml-api.log</string>
<key>StandardErrorPath</key>
<string>/Users/admin/coreml-api/logs/coreml-api-error.log</string>
</dict>
</plist>
EOF
# Load the service
launchctl load ~/Library/LaunchAgents/com.coreml.api.plist
# Verify
curl http://localhost:8000/health
Benchmarks de rendimiento
Estos benchmarks comparan la inferencia CoreML en el Mac Mini M4 contra PyTorch MPS (Metal Performance Shaders) y ejecución solo en CPU. Todas las pruebas usan inferencia de imagen única con batch size 1.
Clasificación de imágenes (ResNet50, entrada 224x224)
| Runtime | Precisión | Latencia (ms) | Rendimiento (img/s) | Potencia (W) |
|---|---|---|---|---|
| CoreML (Neural Engine) | Int8 | 1.2 ms | ~833 | ~3W |
| CoreML (Neural Engine) | Float16 | 2.1 ms | ~476 | ~4W |
| CoreML (GPU only) | Float16 | 3.8 ms | ~263 | ~8W |
| PyTorch MPS (GPU) | Float32 | 5.4 ms | ~185 | ~10W |
| PyTorch CPU | Float32 | 18.6 ms | ~54 | ~12W |
Detección de objetos (YOLOv8n, entrada 640x640)
| Runtime | Precisión | Latencia (ms) | Rendimiento (img/s) | mAP@0.5 |
|---|---|---|---|---|
| CoreML (All Units) | Float16 | 4.8 ms | ~208 | 37.2% |
| CoreML (All Units) | Int8 | 3.5 ms | ~286 | 36.8% |
| PyTorch MPS (GPU) | Float32 | 12.3 ms | ~81 | 37.3% |
| PyTorch CPU | Float32 | 45.7 ms | ~22 | 37.3% |
Ejecute sus propios benchmarks
import coremltools as ct
import numpy as np
import time
model = ct.models.MLModel("ResNet50_optimized.mlpackage", compute_units=ct.ComputeUnit.ALL)
# Warmup (first inference compiles the model for the Neural Engine)
from PIL import Image
dummy = Image.new("RGB", (224, 224))
for _ in range(10):
model.predict({"image": dummy})
# Benchmark
latencies = []
for _ in range(1000):
start = time.perf_counter()
model.predict({"image": dummy})
latencies.append((time.perf_counter() - start) * 1000)
latencies = np.array(latencies)
print(f"Mean latency: {latencies.mean():.2f} ms")
print(f"Median latency: {np.median(latencies):.2f} ms")
print(f"P95 latency: {np.percentile(latencies, 95):.2f} ms")
print(f"P99 latency: {np.percentile(latencies, 99):.2f} ms")
print(f"Throughput: {1000 / latencies.mean():.0f} images/sec")
Conclusión clave: CoreML con Neural Engine ofrece 3-4x mejor rendimiento que PyTorch MPS en el mismo hardware, y 10-15x mejor que inferencia solo en CPU. La ruta cuantizada int8 es el punto óptimo: inferencia más rápida con menos del 0.5% de pérdida de precisión para la mayoría de los modelos.
Escalado con múltiples modelos
Los despliegues de producción a menudo requieren servir múltiples modelos o manejar alta concurrencia. Puede usar nginx como proxy inverso y balanceador de carga entre múltiples instancias de Mac Mini M4, o servir múltiples modelos desde una sola máquina.
Servidor multi-modelo
# multi_model_server.py
import io
import time
import coremltools as ct
from fastapi import FastAPI, File, UploadFile, HTTPException
from PIL import Image
app = FastAPI(title="Multi-Model CoreML Server")
# Load multiple models at startup
models = {}
@app.on_event("startup")
async def load_models():
print("Loading models...")
models["resnet50"] = ct.models.MLModel(
"ResNet50_optimized.mlpackage", compute_units=ct.ComputeUnit.ALL
)
models["yolov8"] = ct.models.MLModel(
"YOLOv8n.mlpackage", compute_units=ct.ComputeUnit.ALL
)
models["efficientnet"] = ct.models.MLModel(
"EfficientNetV2S.mlpackage", compute_units=ct.ComputeUnit.ALL
)
print(f"Loaded {len(models)} models: {list(models.keys())}")
@app.get("/models")
async def list_models():
return {"models": list(models.keys())}
@app.post("/predict/{model_name}")
async def predict(model_name: str, image: UploadFile = File(...)):
if model_name not in models:
raise HTTPException(404, f"Model '{model_name}' not found. Available: {list(models.keys())}")
image_data = Image.open(io.BytesIO(await image.read()))
# Resize based on model requirements
input_sizes = {"resnet50": (224, 224), "yolov8": (640, 640), "efficientnet": (384, 384)}
image_data = image_data.resize(input_sizes.get(model_name, (224, 224)))
start = time.perf_counter()
result = models[model_name].predict({"image": image_data})
latency = (time.perf_counter() - start) * 1000
return {"model": model_name, "prediction": result, "latency_ms": round(latency, 2)}
# Run: uvicorn multi_model_server:app --host 0.0.0.0 --port 8000
Balanceador de carga Nginx entre múltiples Mac Minis
# /opt/homebrew/etc/nginx/servers/coreml.conf
# Install nginx: brew install nginx
# (Homebrew's nginx.conf loads every file in servers/ inside its http block)
# Rate limiting zone (defined at http level)
limit_req_zone $binary_remote_addr zone=api:10m rate=100r/s;
upstream coreml_backend {
# Round-robin across multiple Mac Mini M4 instances
server 10.0.1.10:8000 weight=1; # Mac Mini M4 #1
server 10.0.1.11:8000 weight=1; # Mac Mini M4 #2
server 10.0.1.12:8000 weight=1; # Mac Mini M4 #3
# Health check: remove unhealthy backends
keepalive 32;
}
server {
listen 80;
server_name api.yourdomain.com;
location / {
limit_req zone=api burst=50 nodelay;
proxy_pass http://coreml_backend;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
# Timeout settings for ML inference
proxy_connect_timeout 10s;
proxy_send_timeout 30s;
proxy_read_timeout 30s;
# Enable keepalive to backend
proxy_http_version 1.1;
proxy_set_header Connection "";
}
location /health {
proxy_pass http://coreml_backend;
access_log off;
}
}
# Test config and start
# nginx -t
# nginx
Docker-Compose para desarrollo local
# docker-compose.yml
# Note: CoreML requires macOS -- Docker containers run CPU-only inference
# For production, use launchd services directly on macOS
version: "3.8"
services:
nginx:
image: nginx:alpine
ports:
- "80:80"
volumes:
# coreml.conf from above; for local runs, point its upstream at coreml-api:8000
- ./coreml.conf:/etc/nginx/conf.d/default.conf:ro
depends_on:
- coreml-api
coreml-api:
build: .
# No host port: nginx reaches both replicas on the compose network
volumes:
- ./models:/app/models
environment:
- MODEL_PATH=/app/models/ResNet50_optimized.mlpackage
- WORKERS=2
deploy:
replicas: 2
Monitorización y observabilidad
Los sistemas ML de producción necesitan monitorización de latencia de inferencia, rendimiento, tasas de error y uso de recursos del sistema. Aquí le mostramos cómo instrumentar su API CoreML con métricas de Prometheus y monitorización a nivel de sistema.
Añadir métricas de Prometheus a FastAPI
# pip install prometheus-client prometheus-fastapi-instrumentator
import io
import time
import coremltools as ct
from fastapi import FastAPI, File, UploadFile
from PIL import Image
from prometheus_client import Counter, Histogram, Gauge, generate_latest
from starlette.responses import Response
app = FastAPI(title="CoreML API with Monitoring")
# Prometheus metrics
PREDICTIONS_TOTAL = Counter(
"coreml_predictions_total",
"Total number of predictions",
["model", "status"]
)
PREDICTION_LATENCY = Histogram(
"coreml_prediction_latency_seconds",
"Prediction latency in seconds",
["model"],
buckets=[0.001, 0.002, 0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1.0]
)
MODEL_LOAD_TIME = Gauge(
"coreml_model_load_time_seconds",
"Time taken to load the model",
["model"]
)
ACTIVE_REQUESTS = Gauge(
"coreml_active_requests",
"Number of currently active requests"
)
# Load model with timing
load_start = time.perf_counter()
model = ct.models.MLModel("ResNet50_optimized.mlpackage", compute_units=ct.ComputeUnit.ALL)
MODEL_LOAD_TIME.labels(model="resnet50").set(time.perf_counter() - load_start)
@app.get("/metrics")
async def metrics():
return Response(content=generate_latest(), media_type="text/plain")
@app.post("/predict")
async def predict(image: UploadFile = File(...)):
ACTIVE_REQUESTS.inc()
try:
img = Image.open(io.BytesIO(await image.read())).resize((224, 224))
start = time.perf_counter()
result = model.predict({"image": img})
latency = time.perf_counter() - start
PREDICTION_LATENCY.labels(model="resnet50").observe(latency)
PREDICTIONS_TOTAL.labels(model="resnet50", status="success").inc()
return {"prediction": result, "latency_ms": round(latency * 1000, 2)}
except Exception as e:
PREDICTIONS_TOTAL.labels(model="resnet50", status="error").inc()
raise
finally:
ACTIVE_REQUESTS.dec()
# Run: uvicorn monitored_server:app --host 0.0.0.0 --port 8000
Configuración de Prometheus
# prometheus.yml
global:
scrape_interval: 15s
evaluation_interval: 15s
scrape_configs:
- job_name: "coreml-api"
static_configs:
- targets:
- "10.0.1.10:8000" # Mac Mini #1
- "10.0.1.11:8000" # Mac Mini #2
- "10.0.1.12:8000" # Mac Mini #3
metrics_path: /metrics
scrape_interval: 5s
- job_name: "node-exporter"
static_configs:
- targets:
- "10.0.1.10:9100"
- "10.0.1.11:9100"
- "10.0.1.12:9100"
Script de monitorización a nivel de sistema
#!/bin/bash
# monitor_coreml.sh -- System health monitoring for CoreML inference servers
# Run with: ./monitor_coreml.sh
echo "=== CoreML Server Health Monitor ==="
echo "$(date)"
echo ""
# Memory usage (critical for CoreML model loading)
echo "--- Memory Usage ---"
vm_stat | head -10
echo ""
memory_pressure
echo ""
# CPU and GPU utilization
echo "--- CPU Usage ---"
top -l 1 -n 5 -stats pid,command,cpu,mem | head -10
echo ""
# GPU/Neural Engine power (indicates compute unit activity)
echo "--- GPU/Neural Engine Power ---"
sudo powermetrics --samplers gpu_power,ane_power -n 1 -i 2000 2>/dev/null | grep -E "(GPU|ANE|Neural)"
echo ""
# Disk usage (model files can be large)
echo "--- Disk Usage ---"
df -h / | tail -1
echo ""
# Network connections to API
echo "--- Active API Connections ---"
netstat -an | grep ":8000" | wc -l | xargs echo "Active connections on port 8000:"
echo ""
# API health check
echo "--- API Health Check ---"
curl -s -w "\nHTTP Status: %{http_code}\nResponse Time: %{time_total}s\n" \
http://localhost:8000/health 2>/dev/null || echo "API is DOWN"
Consultas para dashboard de Grafana
# Useful PromQL queries for your Grafana dashboard:
# Average prediction latency (last 5 minutes)
rate(coreml_prediction_latency_seconds_sum[5m]) / rate(coreml_prediction_latency_seconds_count[5m])
# Predictions per second
rate(coreml_predictions_total[1m])
# P99 latency
histogram_quantile(0.99, rate(coreml_prediction_latency_seconds_bucket[5m]))
# Error rate percentage
rate(coreml_predictions_total{status="error"}[5m]) / rate(coreml_predictions_total[5m]) * 100
# Active concurrent requests
coreml_active_requests
Preguntas frecuentes
¿Puedo usar CoreML desde Python sin un proyecto Xcode?
coremltools proporciona capacidades completas de inferencia. Puede cargar modelos .mlpackage y ejecutar predicciones directamente desde scripts Python, servidores Flask/FastAPI o notebooks Jupyter. No se requiere Xcode, Swift ni Objective-C.¿CoreML realmente usa el Neural Engine en Mac Mini M4?
compute_units=ct.ComputeUnit.ALL, el compilador de CoreML enruta automáticamente las operaciones compatibles al Neural Engine. Puede verificarlo monitoreando el consumo de energía con sudo powermetrics --samplers ane_power: verá el ANE (Apple Neural Engine) consumiendo energía durante la inferencia.