Serving Frameworks: vLLM vs TGI vs Triton (and LitServe)

~14 min read

You almost never serve an LLM with a raw model.generate() loop. Purpose-built engines — vLLM, TGI, Triton, LitServe — add continuous batching, paged KV memory, and production plumbing. Which to pick depends on your constraints.

Serving Frameworks: vLLM vs TGI vs Triton (and LitServe) is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Never serve LLMs with a raw generate() loop — purpose-built engines add continuous batching, paged KV memory, streaming, and concurrency
  • vLLM: throughput-focused pure-LLM serving (PagedAttention + continuous batching), OpenAI-compatible API — the common default
  • TGI: Hugging Face's equivalent with a similar feature set and tight HF-ecosystem integration
  • Triton: the generalist — serves any model type / multiple backends (incl. TensorRT-LLM) behind one server, at higher config cost
  • LitServe: lightweight FastAPI-based framework for wrapping any model or custom logic with batching, when you want flexibility over a fixed engine