Inside vLLM: Anatomy of a High-Throughput LLM Inference System (2025)

(aleksagordic.com)

137 points | by sebg 20 hours ago

4 comments

  • miki123211 18 hours ago
    Another great way to understand how vllm works is to read the code of nano-vllm[1]. It's basically "vllm but cut down to size. It's ~5kloc, supports just one model, disposes of some of the abstraction layers that vllm needs due to its codebase size, but contains all the major pieces that make an inference engine fast.

    [1] https://github.com/GeeeekExplorer/nano-vllm

    • yu3zhou4 7 hours ago
      If you prefer C++ and CUDA, then there's also tiny-vllm of mine [0] - recently we broke 1k gh stars

      [0] https://github.com/jmaczan/tiny-vllm

      • bytepoet 2 hours ago
        Thanks for tiny-vllm! The documentation is amazing. I'm using it as a resource for a university course I'm teaching.
  • BinRoo 19 hours ago
    Love that this goes beyond paged attention. Curious how this compares with Radix Attention [1]?

    [1] https://sgl-project-sglang-93.mintlify.app/concepts/radix-at...

  • gdiamos 15 hours ago
    vLLM is originally marketed as paged attention, but in hindsight, separating the web server and GPU process, continuous batching, kv caching / chunking, and a huge model library including low precision mattered more.

    I wonder how much it would cost to vibe code the whole thing from scatch?

    I wonder how much better models need to get before such a thing wouldn't look like code vomit?

    • boredatoms 12 hours ago
      Somewhat related, this vibe translation of vllm

      https://old.reddit.com/r/LocalLLaMA/comments/1vh9lx4/i_porte...

    • mmastrac 14 hours ago
      I've been working on a fresh, AI assisted port of DiffusionGemma from scratch and it takes a significant amount of time to deslop. I've spend a nonzero amount of time on refactoring and comment-vomit cleanup.

      https://github.com/mmastrac/diffgemma

      • brainless 6 hours ago
        Any plans to support smaller models? I have a M4 Mac Mini with 16GB unified memory and an RTX 3060 (Laptop) with 6GB VRAM. My own product experiments all revolve around small models and harness around them. Happy to contribute.
        • mmastrac 5 hours ago
          I've been pondering a smaller quantization and experts swapping for this! Happy to take on PRs if you want to experiment as well.
  • TokenLat 9 hours ago
    [flagged]