CoreWeave Local Dev Loop

Overview

Local development workflow for CoreWeave: build containers, test YAML manifests with dry-run, push to registry, and deploy to CoreWeave CKS.

Prerequisites

• Completed coreweave-install-auth setup
• Docker installed locally
• Container registry access (Docker Hub, GHCR, or CoreWeave registry)

Instructions

Step 1: Project Structure

my-inference-service/
├── Dockerfile
├── src/
│   ├── server.py          # Inference server code
│   └── model_config.py    # Model configuration
├── k8s/
│   ├── deployment.yaml    # GPU deployment manifest
│   ├── service.yaml       # Service and ingress
│   └── hpa.yaml           # Horizontal pod autoscaler
├── scripts/
│   ├── build.sh           # Build and push container
│   └── deploy.sh          # Deploy to CoreWeave
├── .env.local
└── Makefile

Step 2: Build and Push Container

# Build locally
docker build -t my-inference:latest .

# Tag for registry
docker tag my-inference:latest ghcr.io/myorg/my-inference:v1.0.0

# Push
docker push ghcr.io/myorg/my-inference:v1.0.0

Step 3: Validate Manifests Before Deploy

# Dry-run against CoreWeave cluster
kubectl apply -f k8s/deployment.yaml --dry-run=server

# Diff against current state
kubectl diff -f k8s/deployment.yaml

# Check resource requests match available GPU types
kubectl get nodes -l gpu.nvidia.com/class=A100_PCIE_80GB --no-headers | wc -l

Step 4: Deploy and Watch

kubectl apply -f k8s/
kubectl rollout status deployment/my-inference
kubectl logs -f deployment/my-inference

Error Handling

Resources

• CoreWeave CKS Docs
• kubectl dry-run

Next Steps

See coreweave-sdk-patterns for inference client patterns.

CoreWeave Migration Deep Dive

Cost Comparison

Migration Steps

Phase 1: Containerize

# If running on bare EC2/GCE, containerize first
docker build -t inference-server:v1 .
docker push ghcr.io/myorg/inference-server:v1

Phase 2: Adapt YAML for CoreWeave

Key changes from AWS EKS / GKE:

1. Node affinity: Use gpu.nvidia.com/class instead of nvidia.com/gpu.product
2. Storage: Use CoreWeave storage classes (shared-ssd-ord1)
3. Networking: CoreWeave provides flat networking within VPC

Phase 3: Parallel Deploy

Run both old and new infrastructure simultaneously, gradually shift traffic.

Phase 4: Cut Over

Decommission old GPU instances after validation period.

Common Gotchas

Resources

• CoreWeave Pricing
• CoreWeave Documentation

Next Steps

This completes the CoreWeave skill pack. Start with coreweave-install-auth for new deployments.

CoreWeave Multi-Environment Setup

Overview

CoreWeave GPU cloud requires strict environment separation to control infrastructure costs and prevent resource contention. Each environment maps to an isolated Kubernetes namespace with its own GPU quota, scaling policy, and access controls. Development uses cheaper GPU tiers for iteration speed, staging mirrors production GPU types for accurate benchmarking, and production runs full-scale with no scale-to-zero to guarantee inference latency SLAs.

Environment Configuration

const coreweaveConfig = (env: string) => ({
  development: {
    namespace: "app-dev", apiEndpoint: process.env.CW_API_ENDPOINT_DEV!,
    token: process.env.CW_TOKEN_DEV!, gpuType: "L40", scaleToZero: true, replicas: [0, 1],
  },
  staging: {
    namespace: "app-staging", apiEndpoint: process.env.CW_API_ENDPOINT_STG!,
    token: process.env.CW_TOKEN_STG!, gpuType: "A100_PCIE_40GB", scaleToZero: true, replicas: [0, 2],
  },
  production: {
    namespace: "app-prod", apiEndpoint: process.env.CW_API_ENDPOINT_PROD!,
    token: process.env.CW_TOKEN_PROD!, gpuType: "A100_PCIE_80GB", scaleToZero: false, replicas: [2, 10],
  },
}[env]);

Environment Files

# Per-env files: .env.development, .env.staging, .env.production
CW_API_ENDPOINT_{DEV|STG|PROD}=https://k8s.{ord1|ord1|las1}.coreweave.com
CW_TOKEN_{DEV|STG|PROD}=<service-account-token>
CW_NAMESPACE={app-dev|app-staging|app-prod}
CW_GPU_TYPE={L40|A100_PCIE_40GB|A100_PCIE_80GB}

Environment Validation

function validateCoreWeaveEnv(env: string): void {
  const required = ["CW_API_ENDPOINT", "CW_TOKEN", "CW_NAMESPACE", "CW_GPU_TYPE"];
  const suffix = { development: "_DEV", staging: "_STG", production: "_PROD" }[env];
  const missing = required
    .map((k) => (k.includes("NAMESPACE") ? k : `${k}${suffix}`))
    .filter((k) => !process.env[k]);
  if (missing.length) throw new Error(`Missing env vars for ${env}: ${missing.join(", ")}`);
}

Promotion Workflow

# 1. Validate model in dev namespace
kubectl -n app-dev get inferenceservice my-model -o jsonpath='{.status.conditions}'

# 2. Apply staging overlay with production GPU type
kustomize build k8s/overlays/staging | kubectl apply -f -

# 3. Run inference benchmarks against staging endpoint
curl -X POST https://staging.myapp.coreweave.cloud/v1/predict -d @test-payload.json

# 4. Promote to production (blue-green via namespace switch)
kustomize build k8s/overlays/prod | kubectl apply -f -
kubectl -n app-prod rollout status deployment/my-model

Environment Matrix

Instrumentation

async function trackInference(model: string, fn: () => Promise<any>) {
  const start = Date.now();
  try {
    const result = await fn();
    metrics.record('coreweave.inference.latency', Date.now() - start, { model, status: 'ok' });
    metrics.increment('coreweave.inference.completed', { model });
    return result;
  } catch (err) {
    metrics.increment('coreweave.inference.errors', { model, error: err.code });
    throw err;
  }
}

Health Check Dashboard

async function coreweaveHealth(): Promise<Record<string, string>> {
  const gpu = await queryPrometheus('avg(DCGM_FI_DEV_GPU_UTIL)');
  const mem = await queryPrometheus('avg(DCGM_FI_DEV_FB_USED/(DCGM_FI_DEV_FB_USED+DCGM_FI_DEV_FB_FREE))');
  const pods = await queryPrometheus('kube_deployment_status_replicas_available{namespace="inference"}');
  return {
    gpu_utilization: gpu > 20 ? 'healthy' : 'underutilized',
    gpu_memory: mem < 0.9 ? 'healthy' : 'critical',
    inference_pods: pods > 0 ? 'healthy' : 'down',
  };
}

Alerting Rules

const alerts = [
  { metric: 'DCGM_FI_DEV_GPU_UTIL', condition: 'avg < 20', window: '30m', severity: 'warning' },
  { metric: 'gpu_memory_pct', condition: '> 0.95', window: '5m', severity: 'critical' },
  { metric: 'inference_latency_p99', condition: '> 500ms', window: '10m', severity: 'warning' },
  { metric: 'pod_restart_count', condition: '> 3', window: '15m', severity: 'critical' },
];

CoreWeave Performance Tuning

GPU Selection by Workload

Inference Optimization

# Continuous batching with vLLM
containers:
  - name: vllm
    args:
      - "--model=meta-llama/Llama-3.1-8B-Instruct"
      - "--max-num-batched-tokens=8192"
      - "--max-num-seqs=256"
      - "--gpu-memory-utilization=0.90"
      - "--enable-prefix-caching"
      - "--dtype=float16"

Autoscaling Tuning

# HPA based on GPU utilization
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: inference-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: inference-server
  minReplicas: 2
  maxReplicas: 10
  metrics:
    - type: Pods
      pods:
        metric:
          name: DCGM_FI_DEV_GPU_UTIL
        target:
          type: AverageValue
          averageValue: "70"

Performance Benchmarks

Resources

• CoreWeave Inference
• vLLM Documentation

Next Steps

For cost optimization, see coreweave-cost-tuning.

CoreWeave Production Checklist

Inference Services

• [ ] GPU type and count validated for model size
• [ ] Autoscaling configured (KServe or HPA)
• [ ] Health and readiness probes set
• [ ] Resource requests AND limits specified
• [ ] Node affinity targeting correct GPU class
• [ ] minReplicas >= 1 for production (no cold starts)

Storage

• [ ] Model weights in PVC (not downloaded at startup)
• [ ] Checkpoints saved to persistent storage
• [ ] Storage class appropriate (SSD for inference, HDD for archival)

Security

• [ ] Secrets for model tokens and registry access
• [ ] Network policies applied
• [ ] Container images from trusted registries

Monitoring

• [ ] GPU utilization metrics collected
• [ ] Inference latency and throughput tracked
• [ ] Alert on pod restarts and OOM events
• [ ] Log aggregation configured

Rollback

kubectl rollout undo deployment/my-inference
kubectl rollout status deployment/my-inference

Resources

• CoreWeave CKS

Next Steps

For upgrades, see coreweave-upgrade-migration.

CoreWeave Rate Limits

Overview

CoreWeave limits are primarily GPU quota-based rather than API rate limits. Each namespace has allocated GPU quotas per type.

Check GPU Quota

kubectl describe resourcequota -n my-namespace
kubectl get resourcequota -o json | jq '.items[].status'

Inference Request Queuing

import asyncio
from collections import deque

class InferenceQueue:
    def __init__(self, max_concurrent: int = 10):
        self.semaphore = asyncio.Semaphore(max_concurrent)
        self.queue_depth = 0

    async def inference(self, client, prompt: str) -> str:
        self.queue_depth += 1
        async with self.semaphore:
            try:
                return await asyncio.to_thread(client.generate, prompt)
            finally:
                self.queue_depth -= 1

Resources

• CoreWeave Node Pools

Next Steps

For security, see coreweave-security-basics.

CoreWeave Reference Architecture

Architecture Diagram

                    ┌─────────────────────┐
                    │   Load Balancer     │
                    │   (Ingress/LB)      │
                    └──────────┬──────────┘
                               │
              ┌────────────────┼────────────────┐
              │                │                │
     ┌────────▼──────┐ ┌──────▼────────┐ ┌─────▼───────┐
     │ Model A       │ │ Model B       │ │ Model C     │
     │ (vLLM, A100)  │ │ (TGI, H100)  │ │ (SD, L40)   │
     │ 2 replicas    │ │ 1 replica     │ │ 3 replicas  │
     └───────────────┘ └───────────────┘ └─────────────┘
              │                │                │
     ┌────────▼────────────────▼────────────────▼───────┐
     │              Shared Storage (PVC)                │
     │         Models / Checkpoints / Data              │
     └──────────────────────────────────────────────────┘

Project Structure

ml-platform/
├── k8s/
│   ├── base/                    # Shared templates
│   ├── models/
│   │   ├── llama-8b/           # Per-model manifests
│   │   ├── llama-70b/
│   │   └── stable-diffusion/
│   └── infra/
│       ├── storage.yaml         # PVCs
│       ├── secrets.yaml         # Model tokens
│       └── monitoring.yaml      # Prometheus rules
├── containers/
│   ├── vllm/Dockerfile
│   └── custom-server/Dockerfile
├── scripts/
│   ├── deploy.sh
│   └── benchmark.sh
└── monitoring/
    ├── grafana-dashboards/
    └── alert-rules.yaml

Key Design Decisions

Resources

• CoreWeave Documentation
• CoreWeave Examples

Next Steps

For multi-environment setup, see coreweave-multi-env-setup.

CoreWeave SDK Patterns

Overview

CoreWeave is Kubernetes-native -- use kubectl, Kubernetes Python client, or Helm for programmatic management. These patterns cover GPU-aware deployment templates, inference client wrappers, and node affinity configurations.

Instructions

GPU Affinity Helper

# coreweave_helpers.py
from dataclasses import dataclass

@dataclass
class GPUConfig:
    gpu_class: str        # A100_PCIE_80GB, H100_SXM5, L40, etc.
    gpu_count: int = 1
    memory_gb: int = 32
    cpu_cores: int = 4

GPU_CATALOG = {
    "a100-80gb": GPUConfig("A100_PCIE_80GB", memory_gb=48, cpu_cores=8),
    "h100-80gb": GPUConfig("H100_SXM5", memory_gb=64, cpu_cores=12),
    "l40":       GPUConfig("L40", memory_gb=24, cpu_cores=4),
    "a100-8x":   GPUConfig("A100_NVLINK_A100_SXM4_80GB", gpu_count=8, memory_gb=256, cpu_cores=64),
}

def gpu_affinity_block(gpu_class: str) -> dict:
    return {
        "nodeAffinity": {
            "requiredDuringSchedulingIgnoredDuringExecution": {
                "nodeSelectorTerms": [{
                    "matchExpressions": [{
                        "key": "gpu.nvidia.com/class",
                        "operator": "In",
                        "values": [gpu_class],
                    }]
                }]
            }
        }
    }

def gpu_resources(config: GPUConfig) -> dict:
    return {
        "limits": {
            "nvidia.com/gpu": str(config.gpu_count),
            "memory": f"{config.memory_gb}Gi",
            "cpu": str(config.cpu_cores),
        },
        "requests": {
            "nvidia.com/gpu": str(config.gpu_count),
            "memory": f"{config.memory_gb // 2}Gi",
            "cpu": str(config.cpu_cores // 2),
        },
    }

Inference Client Wrapper

# inference_client.py
import requests
from typing import Optional

class CoreWeaveInferenceClient:
    def __init__(self, endpoint: str, timeout: int = 30):
        self.endpoint = endpoint.rstrip("/")
        self.timeout = timeout
        self.session = requests.Session()

    def generate(self, prompt: str, max_tokens: int = 256, **kwargs) -> str:
        resp = self.session.post(
            f"{self.endpoint}/v1/completions",
            json={"prompt": prompt, "max_tokens": max_tokens, **kwargs},
            timeout=self.timeout,
        )
        resp.raise_for_status()
        return resp.json()["choices"][0]["text"]

    def chat(self, messages: list[dict], **kwargs) -> str:
        resp = self.session.post(
            f"{self.endpoint}/v1/chat/completions",