vllm

Included with Lifetime

$97 forever

You are an expert in vLLM, the high-throughput LLM serving engine. You help developers deploy open-source models (Llama, Mistral, Qwen, Phi, Gemma) with PagedAttention for efficient memory management, continuous batching, tensor parallelism for multi-GPU, OpenAI-compatible API, and quantization support — achieving 2-24x higher throughput than HuggingFace Transformers for production LLM serving.

Backend & APIs

What this skill does


# vLLM — High-Throughput LLM Inference Engine

You are an expert in vLLM, the high-throughput LLM serving engine. You help developers deploy open-source models (Llama, Mistral, Qwen, Phi, Gemma) with PagedAttention for efficient memory management, continuous batching, tensor parallelism for multi-GPU, OpenAI-compatible API, and quantization support — achieving 2-24x higher throughput than HuggingFace Transformers for production LLM serving.

## Core Capabilities

### Server Deployment

```bash
# Start OpenAI-compatible API server
vllm serve meta-llama/Llama-3.1-8B-Instruct \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 1 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.9 \
  --quantization awq \
  --api-key my-secret-key

# Multi-GPU (tensor parallelism)
vllm serve meta-llama/Llama-3.1-70B-Instruct \
  --tensor-parallel-size 4 \
  --pipeline-parallel-size 1 \
  --max-num-seqs 256

# With Docker
docker run --runtime nvidia --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  vllm/vllm-openai:latest \
  --model meta-llama/Llama-3.1-8B-Instruct
```

### OpenAI-Compatible Client

```typescript
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:8000/v1",
  apiKey: "my-secret-key",
});

// Chat completion
const response = await client.chat.completions.create({
  model: "meta-llama/Llama-3.1-8B-Instruct",
  messages: [
    { role: "system", content: "You are a helpful coding assistant." },
    { role: "user", content: "Write a Python fibonacci function" },
  ],
  temperature: 0.7,
  max_tokens: 1024,
});

// Streaming
const stream = await client.chat.completions.create({
  model: "meta-llama/Llama-3.1-8B-Instruct",
  messages: [{ role: "user", content: "Explain quantum computing" }],
  stream: true,
});
for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content || "");
}

// Embeddings
const embeddings = await client.embeddings.create({
  model: "BAAI/bge-large-en-v1.5",
  input: ["Your text here"],
});
```

### Python Offline Inference

```python
from vllm import LLM, SamplingParams

llm = LLM(
    model="meta-llama/Llama-3.1-8B-Instruct",
    quantization="awq",
    gpu_memory_utilization=0.9,
    max_model_len=8192,
)

sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=512,
)

# Batch inference (processes all prompts efficiently)
prompts = [
    "Explain machine learning in simple terms",
    "Write a haiku about programming",
    "What is the capital of France?",
]

outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    print(f"Prompt: {output.prompt[:50]}...")
    print(f"Output: {output.outputs[0].text}")
    print(f"Tokens/sec: {len(output.outputs[0].token_ids) / output.metrics.finished_time:.1f}")
```

## Installation

```bash
pip install vllm
# Requires: CUDA 12.1+, PyTorch 2.4+
# GPU: NVIDIA A100, H100, L40S, RTX 4090 recommended
```

## Best Practices

1. **PagedAttention** — vLLM's core innovation; manages KV cache like OS virtual memory, eliminates waste
2. **Continuous batching** — Processes new requests immediately without waiting; maximizes GPU utilization
3. **Quantization** — Use AWQ or GPTQ for 4-bit inference; 2-3x more throughput, minimal quality loss
4. **Tensor parallelism** — Split model across GPUs with `--tensor-parallel-size`; serve 70B+ models
5. **OpenAI compatibility** — Drop-in replacement for OpenAI API; any OpenAI SDK client works unchanged
6. **GPU memory** — Set `--gpu-memory-utilization 0.9` for max throughput; leave 10% for overhead
7. **Max sequences** — Tune `--max-num-seqs` based on your workload; higher = more concurrent requests
8. **Prefix caching** — Enable for shared system prompts; reuses KV cache across requests with same prefix

Files: 2

Size: 5.9 KB

Complexity: 18/100

Category: Backend & APIs

Source: https://github.com/terminalskills/skills/tree/main/skills/vllm

Related in Backend & APIs

jfrog

Included

Interact with the JFrog Platform via the JFrog CLI and REST/GraphQL APIs. Use this skill when the user wants to manage Artifactory repositories, upload or download artifacts, manage builds, configure permissions, manage users and groups, work with access tokens, configure JFrog CLI servers, search artifacts, manage properties, set up replication, manage JFrog Projects, run security audits or scans, look up CVE details, query exposures scan results from JFrog Advanced Security, manage release bundles and lifecycle operations, aggregate or export platform data, or perform any JFrog Platform administration task. Also use when the user mentions jf, jfrog, artifactory, xray, distribution, evidence, apptrust, onemodel, graphql, workers, mission control, curation, advanced security, exposures, or any JFrog product name.

Backend & APIsscripts

cupynumeric-migration-readiness

Included

Pre-migration readiness assessor for porting NumPy to cuPyNumeric. Use BEFORE substantial porting work begins when the user asks whether code will scale on GPU, whether they should migrate to cuPyNumeric, which NumPy patterns transfer cleanly, what must be refactored before porting, or mentions pre-port assessment, scaling analysis, or refactor planning. Inspect the user's source code, look up NumPy usage, cross-reference the cuPyNumeric API support manifest, and distinguish distributed-scaling-friendly patterns from blockers such as unsupported APIs, scalar synchronization, host round-trips, Python/object-heavy control flow, shape/data-dependent branching, and in-place mutation hazards. Produce a verdict of READY, LIGHT REFACTOR, SIGNIFICANT REFACTOR, or NOT RECOMMENDED, with concrete refactor pointers.

Backend & APIsscripts

alibabacloud-data-agent-skill

Included

Invoke Alibaba Cloud Apsara Data Agent for Analytics via CLI to perform natural language-driven data analysis on enterprise databases. Data Agent for Analytics is an intelligent data analysis agent developed by Alibaba Cloud Database team for enterprise users. It automatically completes requirement analysis, data understanding, analysis insights, and report generation based on natural language descriptions. This tool supports: discovering data resources (instances/databases/tables) managed in DMS, initiating query or deep analysis sessions, real-time progress tracking, and retrieving analysis conclusions and generated reports. Use this Skill when users need to query databases, analyze data trends, generate data reports, ask questions in natural language, or mention "Data Agent", "data analysis", "database query", "SQL analysis", "data insights".

Backend & APIsscripts

token-optimizer

Included

Reduce OpenClaw token usage and API costs through smart model routing, heartbeat optimization, budget tracking, and native 2026.2.15 features (session pruning, bootstrap size limits, cache TTL alignment). Use when token costs are high, API rate limits are being hit, or hosting multiple agents at scale. The 4 executable scripts (context_optimizer, model_router, heartbeat_optimizer, token_tracker) are local-only — no network requests, no subprocess calls, no system modifications. Reference files (PROVIDERS.md, config-patches.json) document optional multi-provider strategies that require external API keys and network access if you choose to use them. See SECURITY.md for full breakdown.

Backend & APIsscripts

resend-cli

Included

Use this skill when the task is specifically about operating Resend from an AI agent, terminal session, or CI job via the official resend CLI: installing/authenticating the CLI, sending/listing/updating/cancelling emails, batch sends, domains and DNS, webhooks and local listeners, inbound receiving, contacts, topics, segments, broadcasts, templates, API keys, profiles, or debugging Resend CLI/API failures. Trigger on mentions of Resend CLI, `resend`, `resend doctor`, `resend emails send`, `resend domains`, `resend webhooks listen`, `resend emails receiving`, or agent-friendly terminal automation.

Backend & APIsscripts

alibabacloud-odps-maxframe-coding

Included

Use this skill for MaxFrame SDK development and documentation navigation on Alibaba Cloud MaxCompute (ODPS). Helps answer MaxFrame API, concept, official example, and supported pandas API questions; create data processing programs; read/write MaxCompute tables; debug jobs (remote or local); and build custom DPE runtime images. Trigger when users mention MaxFrame, MaxCompute with MaxFrame, ODPS table processing, DPE runtime, MaxFrame docs/examples, DataFrame/Tensor operations, or GPU runtime setup. Works for both English and Chinese queries about Alibaba Cloud data processing with MaxFrame.

Backend & APIsscripts