Orientation

You can explain what an inference server does that a model.generate() loop doesn't, and say out loud why you're choosing vLLM. · ~2 days

Nothing in this stage requires a GPU. It's the stage where you build the mental model that makes every flag in Stage 4 obvious instead of arbitrary — and, just as usefully, the stage where you learn the three situations in which the honest answer is "don't use vLLM."

The one idea to carry out of here: an inference server is not a faster model. It is a scheduler and a memory allocator wrapped around the same weights. Everything else follows from that.

Pages

#	Topic	What it covers	Status
1	Why Inference Servers Exist	The naive serving loop and its five failures; static batching; KV cache waste; the throughput-versus-latency trade-off stated precisely	🚧
2	Where vLLM Sits	vLLM against TGI, TensorRT-LLM, SGLang, llama.cpp and Ollama; what vLLM optimises for; the three situations where it's the wrong choice	🚧
3	Anatomy of a vLLM Setup	Engine versus server; what pip install vllm actually pulls in; where model weights live; the GPU, driver and CUDA prerequisites, honestly stated	🚧

What you need to have working by the end

• A Python 3.10+ environment, local or Colab.
• Either a GPU you can reach, or a Colab account — the free T4 tier is enough for everything in Stages 0–3.
• Nothing installed yet. Installing vLLM is deliberately in Stage 3, because installing a tool before you know what it does is how people end up with a --gpu-memory-utilization value they can't defend.

---

← Back to the introduction · Next: Why Inference Servers Exist →