Purpose

Deploy LLMs for production inference with high throughput, low latency, and efficient GPU utilization. Covers serving framework selection (vLLM, TGI, TensorRT-LLM), memory management via PagedAttention, continuous batching, speculative decoding for latency reduction, and horizontal scaling strategies.

When to use this skill

Use this skill when:

• deploying a model behind an OpenAI-compatible API using vLLM or TGI
• configuring KV cache management, PagedAttention, or prefix caching
• implementing continuous batching (in-flight batching) for throughput optimization
• setting up speculative decoding with a draft model for latency reduction
• configuring streaming token output via server-sent events (SSE)
• scaling inference across multiple GPUs or nodes with load balancing

Do not use this skill when

• the task is quantizing a model — prefer

  quantization-research

• the task is training or fine-tuning — prefer

  training-infrastructure

• the task is building custom CUDA kernels — prefer

  inference-kernel-optimization

Operating procedure

1. Select the serving framework.
   • vLLM: Best throughput via PagedAttention. Supports continuous batching, prefix caching, and quantized models. Start with:

     python -m vllm.entrypoints.openai.api_server \
       --model meta-llama/Llama-3.1-8B-Instruct \
       --tensor-parallel-size 2 \
       --max-model-len 8192 \
       --gpu-memory-utilization 0.90
     

   • TGI (Text Generation Inference): HuggingFace's solution, Docker-native. Good for quick deployment:

     docker run --gpus all -p 8080:80 \
       ghcr.io/huggingface/text-generation-inference:latest \
       --model-id meta-llama/Llama-3.1-8B-Instruct \
       --max-batch-prefill-tokens 4096 \
       --max-input-tokens 2048
     

   • TensorRT-LLM: NVIDIA-optimized, best single-request latency. Requires model compilation step.
2. Configure KV cache and memory. PagedAttention (vLLM) manages KV cache as virtual memory pages, eliminating fragmentation. Key settings:

   • gpu-memory-utilization

     : fraction of GPU memory for KV cache (0.85–0.95)

   • max-model-len

     : maximum sequence length (determines max KV cache per request)
   • Enable prefix caching (

     --enable-prefix-caching

     ) for workloads with shared system prompts
3. Enable continuous batching. Unlike static batching, continuous batching inserts new requests into the batch as existing ones complete, maximizing GPU utilization. vLLM and TGI enable this by default. Monitor

   batch_size

   and

   queue_depth

   metrics.
4. Set up speculative decoding for latency-sensitive applications. A small draft model generates candidate tokens verified in parallel by the target model:

   python -m vllm.entrypoints.openai.api_server \
     --model meta-llama/Llama-3.1-70B-Instruct \
     --speculative-model meta-llama/Llama-3.1-8B-Instruct \
     --num-speculative-tokens 5
   

   Expect 1.5–2× latency reduction for greedy decoding when acceptance rate is >70%.
5. Configure streaming output. Expose SSE endpoints for token-by-token delivery. Both vLLM and TGI support

   stream=True

   in the OpenAI-compatible API. Set appropriate timeouts (30–120s) and handle client disconnections gracefully.
6. Serve quantized models. Load GPTQ/AWQ models directly:

   python -m vllm.entrypoints.openai.api_server \
     --model TheBloke/Llama-2-13B-GPTQ \
     --quantization gptq --dtype float16
   

7. Scale horizontally. Use tensor parallelism across GPUs on a single node (

   --tensor-parallel-size N

   ). For multi-node, use pipeline parallelism or deploy multiple replicas behind a load balancer (round-robin or least-connections). Monitor GPU utilization, TTFT (time to first token), and TPS (tokens per second).

Decision rules

• Prefer vLLM for throughput-critical workloads (batch inference, high-concurrency APIs).
• Prefer TGI for quick HuggingFace model deployment with minimal configuration.
• Prefer TensorRT-LLM when single-request latency is the primary constraint and NVIDIA GPUs are available.
• Enable prefix caching when >50% of requests share a common system prompt.
• Use speculative decoding only when the draft model acceptance rate exceeds 60%; otherwise the overhead negates latency gains.
• Set

  gpu-memory-utilization

  to 0.90 as default; reduce to 0.85 if OOM errors occur under load spikes.
• Always monitor TTFT separately from total generation time — they have different optimization levers.

Output requirements

1. Serving Configuration

   — framework, model, parallelism settings, memory config, and launch command

2. Performance Baseline

   — TTFT (p50/p95), TPS, max concurrent requests, and GPU utilization under load

3. Scaling Plan

   — horizontal scaling strategy, load balancer config, and auto-scaling triggers

4. API Specification

   — endpoint URLs, streaming support, rate limits, and error response format

References

Read these only when relevant:

• vLLM documentation: https://docs.vllm.ai/
• Kwon et al., "Efficient Memory Management for Large Language Model Serving with PagedAttention"
• HuggingFace TGI: https://huggingface.co/docs/text-generation-inference
• Leviathan et al., "Fast Inference from Transformers via Speculative Decoding"
• NVIDIA TensorRT-LLM: https://github.com/NVIDIA/TensorRT-LLM

Related skills

• quantization-research

  — preparing quantized models for serving

• inference-kernel-optimization

  — custom kernels for attention and sampling

• training-infrastructure

  — GPU cluster management (shared concerns)

• benchmark-design

  — designing latency and throughput benchmarks

Failure handling

• If OOM errors occur during serving, reduce

  gpu-memory-utilization

  , decrease

  max-model-len

  , or enable quantized serving (GPTQ/AWQ).
• If TTFT is too high under load, check batch queue depth — continuous batching can delay new requests when the batch is full. Increase replicas or reduce

  max-batch-size

  .
• If speculative decoding shows no latency improvement, profile the acceptance rate — if below 60%, disable it or use a better draft model.
• If streaming connections time out, verify keep-alive settings on the reverse proxy (nginx/envoy) and increase SSE timeout to match max generation time.