Serves LLMs with vLLM PagedAttention and continuous batching: OpenAI-compatible /v1/chat/completions, GPTQ/AWQ/FP8 quantization, tensor parallelism, and offline LLM.generate. Use when deploying high-throughput GPU inference endpoints or fitting 30B-70B models. Not for TRL training, llama.cpp CPU/edge, or using vLLM only as an inspect-ai/lighteval backend.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Serves LLMs with vLLM PagedAttention and continuous batching: OpenAI-compatible /v1/chat/completions, GPTQ/AWQ/FP8 quantization, tensor parallelism, and offline LLM.generate. Use when deploying high-throughput GPU inference endpoints or fitting 30B-70B models. Not for TRL training, llama.cpp CPU/edge, or using vLLM only as an inspect-ai/lighteval backend.
vLLM achieves up to 24x higher throughput than standard HuggingFace transformers through PagedAttention (block-based KV cache) and continuous batching (mixing prefill/decode requests). It provides an OpenAI-compatible API server, offline batch inference, quantization support (GPTQ/AWQ/FP8), and tensor parallelism for multi-GPU deployments.
When to Use
Use vLLM when:
Deploying production LLM APIs targeting 100+ req/sec
Compare quantized vs non-quantized responses to confirm task-specific performance is unchanged.
Load references/quantization.md when you need detailed AWQ/GPTQ/FP8 setup instructions, model preparation steps, or accuracy comparison benchmarks.
Pitfalls
Out of Memory During Model Loading
Reduce memory usage:
vllm serve MODEL \
--gpu-memory-utilization 0.7 \
--max-model-len 4096
Or use quantization:
vllm serve MODEL --quantization awq
Slow First Token (TTFT > 1 second)
Enable prefix caching for repeated prompts:
vllm serve MODEL --enable-prefix-caching
For long prompts, enable chunked prefill:
vllm serve MODEL --enable-chunked-prefill
Model Not Found Error
Use --trust-remote-code for custom models:
vllm serve MODEL --trust-remote-code
Low Throughput (< 50 req/sec)
Increase concurrent sequences:
vllm serve MODEL --max-num-seqs 512
Check GPU utilization with nvidia-smi — should be > 80%.
Inference Slower Than Expected
Tensor parallelism must use power-of-2 GPU counts:
vllm serve MODEL --tensor-parallel-size 4 # Not 3
Enable speculative decoding for faster generation:
vllm serve MODEL --speculative-model DRAFT_MODEL
Load references/troubleshooting.md when you encounter detailed error messages, need debugging steps, or require performance diagnostics beyond the common issues above.
Load references/optimization.md when you need PagedAttention tuning details, continuous batching configuration, or benchmark results for your specific hardware.
