Why CoreML on Mac Mini M4?
CoreML is Apple's native machine learning framework, purpose-built to extract maximum performance from Apple Silicon. Unlike generic ML frameworks that treat the GPU as a single compute device, CoreML intelligently distributes workloads across the CPU, GPU, and the dedicated 16-core Neural Engine, often running different parts of a model on different compute units simultaneously.
16-Core Neural Engine
The M4's Neural Engine delivers up to 38 TOPS (trillion operations per second) for quantized int8 workloads. CoreML automatically routes compatible layers (convolutions, matrix multiplications, normalization) to the Neural Engine for maximum throughput at minimal power consumption.
Unified Memory Architecture
All compute units share the same memory pool with up to 120 GB/s bandwidth. There is no PCIe bottleneck or data copying between CPU and GPU memory. A 24GB Mac Mini M4 gives every compute unit direct access to the full 24GB, enabling larger models than equivalent discrete GPU setups.
Automatic Compute Dispatch
CoreML's compiler analyzes your model graph and assigns each operation to the optimal compute unit. Convolutions run on the Neural Engine, custom ops fall back to GPU or CPU, and everything executes as a unified pipeline. You get hardware-level optimization with zero manual effort.
Power Efficiency at Scale
A Mac Mini M4 running CoreML inference consumes 5-20W total system power. Compare that to 300-450W for an NVIDIA A100 GPU server. For always-on production inference, this translates to dramatically lower electricity costs and no need for specialized cooling infrastructure.
Key Insight: CoreML is not just for iOS apps. With Python bindings via coremltools, you can convert models from any major framework, run inference from Python scripts, and build production APIs, all while leveraging the Neural Engine that most server-side frameworks cannot access.
Convert PyTorch Models to CoreML
Apple's coremltools library provides a direct conversion path from PyTorch models to CoreML's .mlpackage format. The conversion traces your model with sample inputs and translates each operation to CoreML's internal representation.
Step 1: Install Dependencies
# 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
Step 2: Convert a PyTorch Image Classifier
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")
Step 3: Convert a Custom PyTorch Model
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")
Step 4: Verify the Converted Model
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}")
Convert TensorFlow Models to CoreML
CoreML converts TensorFlow models: SavedModel directories and Keras models (a tf.keras model or an .h5 file). TensorFlow Lite .tflite files cannot be converted: start from the original TensorFlow model instead.
Convert a 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")
Convert a Keras H5 Model
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")
ONNX Models: Convert from the Source Framework
coremltools no longer converts ONNX files: its onnx-coreml converter is frozen, and ct.convert accepts only TensorFlow and PyTorch models (a .onnx path fails with "Unable to determine the type of the model"). If a model reaches you as ONNX, convert it from the framework that produced it. Two common cases:
Export YOLOv8 Straight to 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
Convert the Original PyTorch Model, Not Its ONNX Export
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")
Optimize for Neural Engine
The Neural Engine delivers peak performance with quantized models. Applying post-training quantization, palettization, and pruning can reduce model size by 4-8x and improve Neural Engine throughput by 2-4x, often with negligible accuracy loss.
Float16: Already the Default for 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}%")
Int8 Post-Training Quantization
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")
Palettization (Weight Clustering)
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")
Pruning (Sparsity)
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")
Combined Optimization Pipeline
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")
Tip: Always benchmark accuracy after optimization. Start with float16 (safest), then try int8 quantization, then palettization. Use a held-out validation set and define an acceptable accuracy threshold before applying aggressive optimizations.
Build a REST API for CoreML Inference
Wrapping your CoreML model in a REST API makes it accessible to any client: web apps, mobile apps, microservices, or batch processing pipelines. Below are production-ready examples using both Flask and FastAPI.
Option A: Flask API Server
# 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
Option B: FastAPI Server (Async + OpenAPI Docs)
# 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
Test the 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}
Systemd-Style Service with 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
Performance Benchmarks
These benchmarks compare CoreML inference on the Mac Mini M4 against PyTorch MPS (Metal Performance Shaders) and CPU-only execution. All tests use single-image inference with batch size 1.
Image Classification (ResNet50, 224x224 input)
| Runtime | Precision | Latency (ms) | Throughput (img/s) | Power (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 |
Object Detection (YOLOv8n, 640x640 input)
| Runtime | Precision | Latency (ms) | Throughput (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% |
Run Your Own 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")
Key Takeaway: CoreML with Neural Engine delivers 3-4x better throughput than PyTorch MPS on the same hardware, and 10-15x better than CPU-only inference. The int8 quantized path is the sweet spot: fastest inference with less than 0.5% accuracy loss for most models.
Scaling with Multiple Models
Production deployments often require serving multiple models or handling high concurrency. You can use nginx as a reverse proxy and load balancer across multiple Mac Mini M4 instances, or serve multiple models from a single machine.
Multi-Model Server
# 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
Nginx Load Balancer Across Multiple 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 for Local Development
# 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
Monitoring & Observability
Production ML systems need monitoring for inference latency, throughput, error rates, and system resource usage. Here is how to instrument your CoreML API with Prometheus metrics and system-level monitoring.
Add Prometheus Metrics to 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
Prometheus Configuration
# 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"
System-Level Monitoring Script
#!/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"
Grafana Dashboard Queries
# 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
Frequently Asked Questions
Can I use CoreML from Python without an Xcode project?
coremltools Python package provides full inference capabilities. You can load .mlpackage models and run predictions directly from Python scripts, Flask/FastAPI servers, or Jupyter notebooks. No Xcode, Swift, or Objective-C required.Does CoreML actually use the Neural Engine on Mac Mini M4?
compute_units=ct.ComputeUnit.ALL, CoreML's compiler automatically routes compatible operations to the Neural Engine. You can verify this by monitoring power consumption with sudo powermetrics --samplers ane_power: you will see the ANE (Apple Neural Engine) drawing power during inference.