alt Hacker NewsThe whole thing feels AI written, generated from the codebase.*
*this is incorrect per the author’s response, my apologies.
For instance, it goes into (nano)vLLM internals and doesn’t mention PagedAttention once (one of the core ideas that vLLM is based on)[1].
Also mentions that Part 2 will cover dense vs MoE’s, which is weird because nanovllm hardcodes a dense Qwen3 into the source.
Here are better (imo) explainers about how vLLM works:
- https://hamzaelshafie.bearblog.dev/paged-attention-from-firs...
- https://www.aleksagordic.com/blog/vllm
- https://huggingface.co/blog/continuous_batching
Aleksa’s blog is a bit in the weeds for my taste but it’s really worth working through.
A lot of the magic of vLLM happens in the PagedAttention kernels, which are really succinctly implanted in nanovllm. And the codebase is great and readable by itself!
—
Replies
Actually I thought it was a great example clarity, focus, and economy of words that AI is not capable of at this point in time.
Not really in the PagedAttention kernels. Paged attention was integrated into FlashAttention so that FlashAttention kernels can be used both for prefill and decoding with paged KV. The only paged attention specific kernels are for copying KV blocks (device to device, device to host and host to device). At least for FA2 and FA3, vLLM maintained a fork of FA with paged attention patches.
Hi jbarrow, thanks for your feedback and the links you shared—they're great readings for me (and likely others too).
That said, I need to clarify: the content was not written by AI, and certainly not generated from a database in one shot. If there's some agent + prompt that can produce what I wrote, I'd love to learn it—it would've saved me two weekends :)
Before addressing your questions further, some context: I'm a developer with no ML background but plenty of Cloud Infra experience. I'm currently building an open-source AI Infra project, which is why I studied nano-vllm. So my writing reflects some gaps in ML knowledge.
To your specific points:
> it goes into (nano)vLLM internals and doesn't mention PagedAttention once
I didn't find any explicit "paged attention" naming in nano-vllm. After reading the first article you linked—specifically the "Paged KV Caching" section—I believe the block management logic and CPU/GPU block mapping it describes is exactly what I covered in both posts. It may not be the full picture of paged attention, but I interpreted what I saw in the code and captured the core idea. I think that's a reasonable outcome.
> Part 2 will cover dense vs MoE's, which is weird because nanovllm hardcodes a dense Qwen3 into the source
This reflects my learning approach and background. Same as point 1—I may not have realized the block design was the famous PagedAttention implementation, so I didn't name it as such. For point 2, seeing a dense Qwen3 naturally made me wonder how it differs from the xx-B-A-yy-B MoE models I'd seen on Hugging Face—specifically what changes in the decoder layers. That curiosity led me to learn about MoE and write it up for others with the same questions.
---
I completely understand that in this era, people care more about whether what they're reading is AI-generated—no one wants to waste time on low-effort slop with no human involvement.
But as I explained above—and as my hand-drawn Excalidraw diagrams show (I haven't seen an LLM produce diagrams with logic that satisfies me)—this is the result of learning shaped by my own knowledge background and preferences.