GPU-Accelerated LLM Serving on Kubernetes

We integrate Kubernetes with GPU nodes for dozens of projects — this reduces time-to-production by 2–3 times compared to bare metal. The problem of manual LLM management: GPU idle time during failures, complex scaling, chaotic updates. Our approach provides autoscaling, rolling updates, health checks, and resource isolation.

Recently, we deployed LLaMA-3-8B for a chatbot with latency requirements of p99 < 200 ms. We used vLLM with PagedAttention on two A100s. After optimization, we achieved 45 req/s and 120 ms p99. Details below.

Why Kubernetes with GPUs Is Critical for LLMs

LLMs consume up to 320 GB of memory per model. A single pod failure should not break the service. Kubernetes ensures resource isolation and automatic recovery. In our experience, a cluster with GPU nodes pays for itself in 2–3 months due to reduced downtime, saving roughly $2,000 per node monthly. The NVIDIA GPU Operator automates drivers, and the Device Plugin manages virtualization.

Cluster Preparation: NVIDIA Device Plugin and GPU Operator

The NVIDIA Device Plugin is mandatory. Install via Helm:

helm repo add nvdp https://nvidia.github.io/k8s-device-plugin
helm repo update
helm upgrade -i nvdp nvdp/nvidia-device-plugin \
  --namespace nvidia-device-plugin --create-namespace \
  --set gfd.enabled=true \
  --set devicePlugin.config.sharing.timeSlicing.resources[0].name=nvidia.com/gpu \
  --set devicePlugin.config.sharing.timeSlicing.resources[0].replicas=4

Time-slicing gives up to 4 virtual GPUs per physical GPU — saving up to 40% of costs under low load, which for a 4-GPU node amounts to roughly $2,000 per month. For production, use dedicated GPUs with MIG (Multi-Instance GPU) — isolation is higher and latency more stable.

How vLLM Compares with Competitors

vLLM is 2× faster than LMDeploy on A100 with LLaMA-3-8B under the same hardware. The reason is PagedAttention and KV-cache optimization, which reduces memory fragmentation and improves continuous batching. According to vLLM documentation, PagedAttention improves memory efficiency by 2–4×. Let's compare key metrics:

Example deployment of LLaMA-3-8B with streaming and monitoring:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-llama3-8b
  namespace: ai-serving
spec:
  replicas: 2
  selector:
    matchLabels:
      app: vllm-llama3-8b
  template:
    metadata:
      labels:
        app: vllm-llama3-8b
      annotations:
        prometheus.io/scrape: "true"
        prometheus.io/port: "8080"
    spec:
      nodeSelector:
        nvidia.com/gpu.product: "A100-SXM4-80GB"
      tolerations:
        - key: nvidia.com/gpu
          operator: Exists
          effect: NoSchedule
      containers:
        - name: vllm
          image: vllm/vllm-openai:v0.5.0
          command:
            - python3
            - -m
            - vllm.entrypoints.openai.api_server
          args:
            - --model=/models/llama-3-8b-instruct
            - --tensor-parallel-size=1
            - --max-model-len=8192
            - --max-num-seqs=256
            - --gpu-memory-utilization=0.90
            - --port=8000
          ports:
            - containerPort: 8000
              name: http
          resources:
            limits:
              nvidia.com/gpu: "1"
              memory: "32Gi"
              cpu: "8"
            requests:
              nvidia.com/gpu: "1"
              memory: "24Gi"
              cpu: "4"
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 60
            periodSeconds: 10
            failureThreshold: 10
          livenessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 120
            periodSeconds: 30
      volumes:
        - name: model-storage
          persistentVolumeClaim:
            claimName: model-storage-pvc

How to Scale LLMs Under Load

The standard HPA based on CPU is useless. We use custom metrics — the vLLM queue size (vllm_queue_size). Example HPA with scaling policies:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: vllm-hpa
  namespace: ai-serving
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: vllm-llama3-8b
  minReplicas: 1
  maxReplicas: 8
  metrics:
    - type: Pods
      pods:
        metric:
          name: vllm_queue_size
        target:
          type: AverageValue
          averageValue: "10"
  behavior:
    scaleUp:
      stabilizationWindowSeconds: 60
      policies:
        - type: Pods
          value: 1
          periodSeconds: 120
    scaleDown:
      stabilizationWindowSeconds: 300

This configuration saves up to 30% of GPU hours under low load.

What If the Model Doesn't Fit on a Single GPU?

For models 70B+, use tensor parallelism and node affinity to ensure GPUs are on the same node. MIG partitions the GPU at the hardware level, providing deterministic compute and memory isolation for models up to 13B. Example with 4 GPUs:

resources:
  limits:
    nvidia.com/gpu: "4"
    memory: "320Gi"
    cpu: "32"
affinity:
  podAntiAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:
      - topologyKey: kubernetes.io/hostname

Comparison of GPU sharing modes:

Step-by-Step Deployment Guide

1. Install NVIDIA Device Plugin via Helm as described above.
2. Create a PVC for model storage with sufficient size (at least 50 GB for a 7B model).
3. Deploy vLLM using the provided manifest, specifying the correct model and resources.
4. Set up a Service and Ingress for API access (e.g., via Istio or Nginx).
5. Configure HPA with custom metrics by scraping them with Prometheus and the adapter.
6. Test scaling by running load tests (e.g., with Locust).

What's Included in Our Work

• Audit of current infrastructure and recommendations for GPU nodes.
• Installation and configuration of NVIDIA Device Plugin / GPU Operator.
• Deployment of vLLM with streaming and monitoring (Prometheus + metrics).
• Configuration of HPA based on custom metrics.
• Documentation and team training (2 days).
• Guaranteed stable operation — 24/7 support for the first month.
• Turnkey solution: we handle everything from cluster setup to production rollout within 2 weeks.
• Free project evaluation — contact us for a preliminary assessment.

Implementation Timeline

1–2 weeks — turnkey deployment of a single model. From 1 month — multi-model cluster with CI/CD, disaster recovery, and cost optimization.

Get a Consultation

Contact our certified engineers — we will help you deploy an LLM infrastructure that can handle production loads. Experience: 50+ projects in ML infrastructure. We offer free project evaluation: just write to us. Our service includes full setup, monitoring, and ongoing support. Evaluate your project with us — we will select the optimal configuration for your tasks.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.