runpod-deployment（runpod 部署）

1. Serverless Workers - Scale-to-zero handlers with pay-per-second billing
2. vLLM Endpoints - OpenAI-compatible LLM serving with 2-3x throughput
3. Pod Management - Dedicated GPU instances for development/training
4. Cost Optimization - GPU selection, spot instances, budget controls

Key deliverables:

• Production-ready serverless handlers with streaming
• vLLM deployment with OpenAI API compatibility
• Cost-optimized GPU selection for any model size
• Health monitoring and auto-scaling configuration

<quick_start> Minimal Serverless Handler (v1.8.1):

import runpod

def handler(job):
    """Basic handler - receives job, returns result."""
    job_input = job["input"]
    prompt = job_input.get("prompt", "")

    # Your inference logic here
    result = process(prompt)

    return {"output": result}

runpod.serverless.start({"handler": handler})

Streaming Handler:

import runpod

def streaming_handler(job):
    """Generator for streaming responses."""
    for chunk in generate_chunks(job["input"]):
        yield {"token": chunk, "finished": False}
    yield {"token": "", "finished": True}

runpod.serverless.start({
    "handler": streaming_handler,
    "return_aggregate_stream": True
})

vLLM OpenAI-Compatible Client:

from openai import OpenAI

client = OpenAI(
    api_key="RUNPOD_API_KEY",
    base_url="https://api.runpod.ai/v2/ENDPOINT_ID/openai/v1",
)

response = client.chat.completions.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=100,
)

</quick_start>

<success_criteria> A RunPod deployment is successful when:

• Handler processes requests without errors
• Endpoint scales appropriately (0 → N workers)
• Cold start time is acceptable for use case
• Cost stays within budget projections
• Health checks pass consistently </success_criteria>

<m1_mac_critical>

M1/M2 Mac: Cannot Build Docker Locally

ARM architecture is incompatible with RunPod's x86 GPUs.

Solution: GitHub Actions builds for you:

# Push code - Actions builds x86 image
git add . && git commit -m "Deploy" && git push

See reference/cicd.md for complete GitHub Actions workflow.

Never run docker build locally for RunPod on Apple Silicon. </m1_mac_critical>

<gpu_selection>

GPU Selection Matrix (January 2025)

Quick Selection:

def select_gpu(model_params_b: float, quantized: bool = False) -> str:
    effective = model_params_b * (0.5 if quantized else 1.0)
    if effective <= 3: return "RTX_A4000"      # $0.36/hr
    if effective <= 8: return "RTX_4090"       # $0.44/hr
    if effective <= 30: return "A40"           # $0.65/hr
    if effective <= 70: return "A100_80GB"     # $1.89/hr
    return "H100_80GB"                         # $4.69/hr

See reference/cost-optimization.md for detailed pricing and budget controls. </gpu_selection>

<handler_patterns>

Handler Patterns

Progress Updates (Long-Running Tasks)

import runpod

def long_task_handler(job):
    total_steps = job["input"].get("steps", 10)

    for step in range(total_steps):
        process_step(step)
        runpod.serverless.progress_update(
            job_id=job["id"],
            progress=int((step + 1) / total_steps * 100)
        )

    return {"status": "complete", "steps": total_steps}

runpod.serverless.start({"handler": long_task_handler})

Error Handling

import runpod
import traceback

def safe_handler(job):
    try:
        # Validate input
        if "prompt" not in job["input"]:
            return {"error": "Missing required field: prompt"}

        result = process(job["input"])
        return {"output": result}

    except torch.cuda.OutOfMemoryError:
        return {"error": "GPU OOM - reduce input size", "retry": False}
    except Exception as e:
        return {"error": str(e), "traceback": traceback.format_exc()}

runpod.serverless.start({"handler": safe_handler})

See reference/serverless-workers.md for async patterns, batching, and advanced handlers. </handler_patterns>

<vllm_deployment>

vLLM Deployment

Note: vLLM uses OpenAI-compatible API FORMAT but connects to YOUR RunPod endpoint, NOT OpenAI servers. Models run on your GPU (Llama, Qwen, Mistral, etc.)

Environment Configuration

vllm_env = {
    "MODEL_NAME": "meta-llama/Llama-3.1-70B-Instruct",
    "HF_TOKEN": "${HF_TOKEN}",
    "TENSOR_PARALLEL_SIZE": "2",        # Multi-GPU
    "MAX_MODEL_LEN": "16384",
    "GPU_MEMORY_UTILIZATION": "0.95",
    "QUANTIZATION": "awq",              # Optional: awq, gptq
}

OpenAI-Compatible Streaming

from openai import OpenAI

client = OpenAI(
    api_key="RUNPOD_API_KEY",
    base_url="https://api.runpod.ai/v2/ENDPOINT_ID/openai/v1",
)

stream = client.chat.completions.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    messages=[{"role": "user", "content": "Write a poem"}],
    stream=True,
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

Direct RunPod Streaming

import requests

url = "https://api.runpod.ai/v2/ENDPOINT_ID/run"
headers = {"Authorization": "Bearer RUNPOD_API_KEY"}

response = requests.post(url, headers=headers, json={
    "input": {"prompt": "Hello", "stream": True}
})
job_id = response.json()["id"]

# Stream results
stream_url = f"https://api.runpod.ai/v2/ENDPOINT_ID/stream/{job_id}"
with requests.get(stream_url, headers=headers, stream=True) as r:
    for line in r.iter_lines():
        if line: print(line.decode())

See reference/model-deployment.md for HuggingFace, TGI, and custom model patterns. </vllm_deployment>

<auto_scaling>

Auto-Scaling Configuration

Scaler Types

Configuration Patterns

configs = {
    "interactive_api": {
        "workers_min": 1,      # Always warm
        "workers_max": 5,
        "idle_timeout": 120,
        "scaler_type": "QUEUE_DELAY",
        "scaler_value": 1,     # 1s latency target
    },
    "batch_processing": {
        "workers_min": 0,
        "workers_max": 20,
        "idle_timeout": 30,
        "scaler_type": "REQUEST_COUNT",
        "scaler_value": 5,
    },
    "cost_optimized": {
        "workers_min": 0,
        "workers_max": 3,
        "idle_timeout": 15,    # Aggressive scale-down
        "scaler_type": "QUEUE_DELAY",
        "scaler_value": 5,
    },
}

See reference/pod-management.md for pod lifecycle and scaling details. </auto_scaling>

<health_monitoring>

Health & Monitoring

Quick Health Check

import runpod

async def check_health(endpoint_id: str):
    endpoint = runpod.Endpoint(endpoint_id)
    health = await endpoint.health()

    return {
        "status": health.status,
        "workers_ready": health.workers.ready,
        "queue_depth": health.queue.in_queue,
        "avg_latency_ms": health.metrics.avg_execution_time,
    }

GraphQL Metrics Query

query GetEndpoint($id: String!) {
    endpoint(id: $id) {
        status
        workers { ready running pending throttled }
        queue { inQueue inProgress completed failed }
        metrics {
            requestsPerMinute
            avgExecutionTimeMs
            p95ExecutionTimeMs
            successRate
        }
    }
}

See reference/monitoring.md for structured logging, alerts, and dashboards. </health_monitoring>

<dockerfile_pattern>

Dockerfile Template

FROM runpod/pytorch:2.1.0-py3.10-cuda12.1.1-devel-ubuntu22.04

WORKDIR /app

# Install dependencies (cached layer)
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

# Copy application
COPY . .

# RunPod entrypoint
CMD ["python", "-u", "handler.py"]

See reference/templates.md for runpod.toml, requirements.txt patterns. </dockerfile_pattern>

<file_locations>

Reference Files

Core Patterns:

• reference/serverless-workers.md - Handler patterns, streaming, async
• reference/model-deployment.md - vLLM, TGI, HuggingFace deployment
• reference/pod-management.md - GPU types, scaling, lifecycle

Operations:

• reference/cost-optimization.md - Budget controls, right-sizing
• reference/monitoring.md - Health checks, logging, GraphQL
• reference/troubleshooting.md - Common issues and solutions

DevOps:

• reference/cicd.md - GitHub Actions for M1 Mac builds
• reference/templates.md - Dockerfile, runpod.toml configs
• templates/runpod-worker.py - Production handler template </file_locations>

User wants serverless deployment: → Provide handler pattern, Dockerfile, deployment steps → Reference: reference/serverless-workers.md

User wants vLLM endpoint: → Provide vLLM env config, OpenAI client setup → Reference: reference/model-deployment.md

User wants cost optimization: → Provide GPU selection matrix, spot pricing, budget controls → Reference: reference/cost-optimization.md

User on M1/M2 Mac: → CRITICAL: Must use GitHub Actions for builds → Reference: reference/cicd.md

User has deployment issues: → Check health endpoint, review logs → Reference: reference/troubleshooting.md

<cost_quick_ref>

Cost Quick Reference

Storage Costs:

• Container disk: $0.10/GB/month (billed in 5-min intervals)
• Network volumes: $0.07/GB/month (first 1TB), $0.05/GB after

Cost Estimation:

def estimate_monthly_cost(gpu_type, daily_requests, avg_time_s):
    rates = {"RTX_4090": 0.44, "A40": 0.65, "A100_80GB": 1.89}
    daily_hours = (daily_requests * avg_time_s) / 3600
    return daily_hours * 30 * rates.get(gpu_type, 1.0)

# Example: 1000 requests/day, 5s each, RTX 4090
# = (1000 * 5) / 3600 * 30 * 0.44 = $18.33/month

</cost_quick_ref>

<cli_commands>

Quick CLI Commands

# Install
pip install runpod

# Deploy endpoint
runpodctl project deploy --name my-endpoint --gpu-type "NVIDIA RTX 4090"

# Health check
runpod endpoint health ENDPOINT_ID

# View logs
runpod endpoint logs ENDPOINT_ID

# Scale workers
runpod endpoint scale ENDPOINT_ID --min 1 --max 10

# Local testing
python handler.py --rp_serve_api

</cli_commands>

<example_session>

Example Session

User: "Deploy a Llama 3.1 8B model to RunPod serverless"

Claude:

1. First, select GPU - 8B model fits on RTX 4090 ($0.44/hr):

gpu_type = "NVIDIA GeForce RTX 4090"

1. Use vLLM for 2-3x throughput. Create endpoint with:

env = {
    "MODEL_NAME": "meta-llama/Llama-3.1-8B-Instruct",
    "MAX_MODEL_LEN": "8192",
    "GPU_MEMORY_UTILIZATION": "0.95",
}

1. Access via OpenAI-compatible API:

from openai import OpenAI
client = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://api.runpod.ai/v2/ENDPOINT_ID/openai/v1",
)

1. Cost estimate: ~$0.44/hr compute, scale-to-zero when idle. </example_session>

Emit Outcome Sidecar

As the final step, write to ~/.claude/skill-analytics/last-outcome-runpod-deployment.json:

{"ts":"[UTC ISO8601]","skill":"runpod-deployment","version":"1.0.0","variant":"default",
 "status":"[success|partial|error]","runtime_ms":[estimated ms from start],
 "metrics":{"pods_configured":[n],"deployments_created":[n]},
 "error":null,"session_id":"[YYYY-MM-DD]"}

Use status "partial" if some stages failed but results were produced. Use "error" only if no output was generated.