Skip to main content ホーム クリエイター damionrashford mlx serve
serve Compress, deploy, and serve trained ML models in production. Covers model compression (quantization, pruning, distillation, ONNX export), inference APIs, containerization, CI/CD pipelines, monitoring, health endpoints, model versioning, and reproducibility packaging. Use when the user has a trained model and wants to reduce its size, deploy it, serve it, containerize it, build an inference API, set up monitoring, write a model card, create a CI/CD pipeline, or package for reproducibility.
インストールへ移動 Skills Marketplace コミュニティが作成したAIスキルを発見・探索
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/damionrashford/mlx --skill serveコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Zipをダウンロード ダウンロード中... このリポジトリの他の Skills Comprehensive Claude Code knowledge base — plugins, hooks, skills, agents, MCP, channels, headless mode, permissions, settings, and all extensibility features. Use when building, configuring, debugging, or extending Claude Code.
Use when building a skill, creating a SKILL.md, packaging a workflow, making a slash command, or asked "how do I make a skill". Scaffolds the folder, generates SKILL.md from a template, validates against spec. Produces a complete ready-to-deploy skill folder: scripts, references, assets. Also use to review or improve an existing skill.
Statistical analysis, hypothesis testing, A/B testing, cohort analysis, segmentation, trend detection, business metrics, pre-delivery validation, and data visualization. Use when the user asks to "analyze this data", "run a statistical test", "compare groups", "find trends", "do A/B test analysis", "segment customers", "calculate KPIs", "validate this analysis", "check my work", "sanity check", "review my numbers", "make a chart", "create a dashboard", "plot the data", "visualize results", or mentions hypothesis testing, cohort analysis, business analytics, data validation, bar charts, line charts, heatmaps, scatter plots, or data storytelling.
name serve description Compress, deploy, and serve trained ML models in production. Covers model compression (quantization, pruning, distillation, ONNX export), inference APIs, containerization, CI/CD pipelines, monitoring, health endpoints, model versioning, and reproducibility packaging. Use when the user has a trained model and wants to reduce its size, deploy it, serve it, containerize it, build an inference API, set up monitoring, write a model card, create a CI/CD pipeline, or package for reproducibility.
allowed-tools Bash(uv run * scripts/benchmark_model.py *) Bash, Read, Write, Edit, Glob, Grep
disable-model-invocation true argument-hint path to model file or project root (e.g. "model.joblib" or ".") model opus effort high compatibility >=1.0 metadata {"category":"mlops","tags":["deployment","serving","docker","fastapi","ci-cd","monitoring","model-card","onnx","quantization","compression"],"phase":"deploy"}
Model Serving & Deployment
Reference for taking trained models to production. Follow phases in order.
Phase 0: Compression (before deploying)
Reduce model size and latency before serving. Run benchmarks to confirm tradeoff is acceptable.
uv run ${CLAUDE_SKILL_DIR} /scripts/benchmark_model.py model.joblib model_compressed.joblib data/test.csv
Quantization
Dynamic (no calibration): torch.quantization.quantize_dynamic — easiest, CPUStatic (calibration required): torch.quantization.prepare + convert — faster inference4-bit (LLMs): bitsandbytes load_in_4bit=True — 4× memory reductionGPTQ : AutoGPTQ — accurate 4-bit for LLMs
Export
ONNX : skl2onnx for sklearn, torch.onnx.export for PyTorchONNX Runtime : InferenceSession — cross-platform optimized inferenceGGUF : convert_hf_to_gguf.py for local LLM inference
Pruning
torch.nn.utils.prune.l1_unstructured — remove smallest weightsStructured pruning: remove entire filters/heads for real speedup
Knowledge distillation
Student model learns from teacher's soft labels
Loss = α × CE(hard labels) + (1−α) × KD(soft labels, τ)
Benchmark targets
Metric Goal Latency p50/p95 < 2× original Throughput > 1.5× original Accuracy delta < 1% absolute Memory RSS < 50% original
See references/compression-guide.md
Phase 1: Pre-deployment checklist
Before writing any serving code:
python3 -c "import joblib, tracemalloc; tracemalloc.start(); joblib.load('model.joblib')")
requirements.txt pinned (pip freeze > requirements.txt)
Phase 2: Inference API patterns
Standard FastAPI template
from contextlib import asynccontextmanager
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, Field
import joblib, numpy as np, logging, time
logger = logging.getLogger(__name__)
model = None
@asynccontextmanager
async def lifespan (app: FastAPI ):
global model
model = joblib.load("model.joblib" )
logger.info("Model loaded" )
yield
app = FastAPI(title="Model API" , version="1.0.0" , lifespan=lifespan)
class PredictRequest (BaseModel ):
features: list [float ] = Field(..., min_length=1 )
class PredictResponse (BaseModel ):
prediction: float
probability: float | None = None
latency_ms: float
@app.post("/predict" , response_model=PredictResponse
def predict (req: PredictRequest ):
start = time.perf_counter()
X = np.array(req.features).reshape(1 , -1 )
pred = float (model.predict(X)[0 ])
proba = float (model.predict_proba(X).max ()) if hasattr (model, 'predict_proba' ) else None
return PredictResponse(prediction=pred, probability=proba,
latency_ms=(time.perf_counter() - start) * 1000 )
@app.post("/predict/batch"
def predict_batch (requests: list [PredictRequest] ):
X = np.array([r.features for r in requests])
preds = model.predict(X).tolist()
return {"predictions" : preds, "count" : len (preds)}
@app.get("/health"
def health ():
return {"status" : "healthy" , "model_loaded" : model is not None }
@app.get("/ready"
def ready ():
if model is None :
raise HTTPException(503 , "Model not loaded" )
return {"status" : "ready" }
Async with background batching (high throughput) import asyncio
from collections import deque
request_queue: deque = deque()
async def batch_processor ():
while True :
if len (request_queue) >= 32 or (request_queue and await asyncio.sleep(0.005 ) is None ):
batch = [request_queue.popleft() for _ in range (min (32 , len (request_queue)))]
features = np.array([b['features' ] for b in batch])
preds = model.predict(features)
for item, pred in zip (batch, preds):
item['future' ].set_result(float (pred))
await asyncio.sleep(0.001 )
Model caching (multiple models) from functools import lru_cache
@lru_cache(maxsize=4 )
def load_model (version: str ):
return joblib.load(f"models/{version} /model.joblib" )
@app.post("/predict/{version}"
def predict_versioned (version: str , req: PredictRequest ):
m = load_model(version)
return {"prediction" : float (m.predict([req.features])[0 ])}
Phase 3: Containerization
Dockerfile FROM python:3.11-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
COPY model.joblib .
COPY serve.py .
EXPOSE 8000
HEALTHCHECK --interval=30s --timeout=5s CMD curl -f http://localhost:8000/health || exit 1
CMD ["uvicorn", "serve:app", "--host", "0.0.0.0", "--port", "8000", "--workers", "2"]
docker-compose.yml services:
model-api:
build: .
ports:
- "8000:8000"
environment:
- MODEL_PATH=/app/model.joblib
volumes:
- ./models:/app/models:ro
restart: unless-stopped
healthcheck:
test: ["CMD" , "curl" , "-f" , "http://localhost:8000/health" ]
interval: 30s
retries: 3
Build and run docker build -t model-api:v1.0 .
docker run -p 8000:8000 model-api:v1.0
docker compose up -d
Phase 4: CI/CD (GitHub Actions)
name: Build & Deploy
on:
push:
branches: [main ]
paths: ['serve.py' , 'requirements.txt' , 'model.joblib' ]
jobs:
test:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with: { python-version: '3.11' }
- run: pip install -r requirements.txt pytest httpx
- run: pytest tests/ -v
build:
needs: test
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Build Docker image
run: |
docker build -t model-api:${{ github.sha }} .
docker tag model-api:${{ github.sha }} model-api:latest
- name: Smoke test
run: |
docker run -d -p 8000:8000 --name test-api model-api:latest
sleep 5
curl -f http://localhost:8000/health
docker stop test-api
Phase 5: Monitoring
Structured request logging import json, time
from fastapi import Request
@app.middleware("http"
async def log_requests (request: Request, call_next ):
start = time.perf_counter()
response = await call_next(request)
duration_ms = (time.perf_counter() - start) * 1000
logger.info(json.dumps({
"path" : request.url.path,
"method" : request.method,
"status" : response.status_code,
"duration_ms" : round (duration_ms, 2 ),
}))
return response
Prometheus metrics from prometheus_client import Counter, Histogram, generate_latest, CONTENT_TYPE_LATEST
from fastapi import Response
predict_counter = Counter('predictions_total' , 'Total predictions' , ['status' ])
predict_latency = Histogram('prediction_latency_seconds' , 'Prediction latency' ,
buckets=[.005 , .01 , .025 , .05 , .1 , .25 , .5 ])
@app.get("/metrics"
def metrics ():
return Response(generate_latest(), media_type=CONTENT_TYPE_LATEST)
Data drift detection (simple) import numpy as np
class DriftMonitor :
def __init__ (self, reference_stats: dict ):
self .ref = reference_stats
self .window = []
def record (self, features: list [float ] ):
self .window.append(features)
if len (self .window) >= 1000 :
self ._check_drift()
self .window = []
def _check_drift (self ):
arr = np.array(self .window)
for i, (feat, stats) in enumerate (self .ref.items()):
z = abs (arr[:, i].mean() - stats['mean' ]) / (stats['std' ] + 1e-8 )
if z > 3 :
logger.warning(f"Drift detected in {feat} : z-score={z:.2 f} " )
Phase 6: Model card template # Model Card: {model_name}
## Model details
- **Type** : {classification|regression}
- **Algorithm** : {e.g. XGBoost, RandomForest}
- **Version** : {semantic version}
- **Training date** : {date}
## Intended use
- **Primary use** : {task description}
- **Out-of-scope** : {what this model should NOT be used for}
## Performance
| Metric | Validation | Test |
|--------|-----------|------|
| {metric} | {val_ score} | {test_score} |
## Data
- **Training data** : {description, size, date range}
- **Features** : {N} input features
- **Target** : {description}
## Limitations
- {List known limitations}
## Ethical considerations
- {Bias considerations, fairness assessment}
Phase 7: Reproducibility package {project}/
├── model.joblib # trained model
├── config.json # hyperparameters + training config
├── data_manifest.json # dataset checksums + version
├── requirements.txt # pinned dependencies
├── serve.py # inference API
├── Dockerfile
├── README.md # how to reproduce training + how to serve
└── tests/
└── test_serve.py # API smoke tests
config.json {
"model" : "xgboost" ,
"experiment_id" : "exp004" ,
"hyperparameters" : { "max_depth" : 6 , "learning_rate" : 0.1 } ,
"training_date" : "2026-03-07" ,
"python_version" : "3.11" ,
"val_score" : 0.8634 ,
"test_score" : 0.8601 ,
"feature_count" : 42
}
Rules
Pin ALL dependencies in requirements.txt before containerizing
Always have /health and /ready endpoints — they are required for load balancers
Log every prediction with input hash, output, and latency
Never load models on every request — load once at startup
Test the Docker image locally before pushing to registry
Save config.json alongside every model artifact
