frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

At Age 25, Wikipedia Refuses to Evolve

https://spectrum.ieee.org/wikipedia-at-25
1•asdefghyk•24s ago•1 comments

Show HN: ReviewReact – AI review responses inside Google Maps ($19/mo)

https://reviewreact.com
1•sara_builds•49s ago•0 comments

Why AlphaTensor Failed at 3x3 Matrix Multiplication: The Anchor Barrier

https://zenodo.org/records/18514533
1•DarenWatson•1m ago•0 comments

Ask HN: How much of your token use is fixing the bugs Claude Code causes?

1•laurex•5m ago•0 comments

Show HN: Agents – Sync MCP Configs Across Claude, Cursor, Codex Automatically

https://github.com/amtiYo/agents
1•amtiyo•6m ago•0 comments

Hello

1•otrebladih•7m ago•0 comments

FSD helped save my father's life during a heart attack

https://twitter.com/JJackBrandt/status/2019852423980875794
2•blacktulip•10m ago•0 comments

Show HN: Writtte – Draft and publish articles without reformatting, anywhere

https://writtte.xyz
1•lasgawe•12m ago•0 comments

Portuguese icon (FROM A CAN) makes a simple meal (Canned Fish Files) [video]

https://www.youtube.com/watch?v=e9FUdOfp8ME
1•zeristor•13m ago•0 comments

Brookhaven Lab's RHIC Concludes 25-Year Run with Final Collisions

https://www.hpcwire.com/off-the-wire/brookhaven-labs-rhic-concludes-25-year-run-with-final-collis...
2•gnufx•16m ago•0 comments

Transcribe your aunts post cards with Gemini 3 Pro

https://leserli.ch/ocr/
1•nielstron•19m ago•0 comments

.72% Variance Lance

1•mav5431•21m ago•0 comments

ReKindle – web-based operating system designed specifically for E-ink devices

https://rekindle.ink
1•JSLegendDev•22m ago•0 comments

Encrypt It

https://encryptitalready.org/
1•u1hcw9nx•22m ago•1 comments

NextMatch – 5-minute video speed dating to reduce ghosting

https://nextmatchdating.netlify.app/
1•Halinani8•23m ago•1 comments

Personalizing esketamine treatment in TRD and TRBD

https://www.frontiersin.org/articles/10.3389/fpsyt.2025.1736114
1•PaulHoule•25m ago•0 comments

SpaceKit.xyz – a browser‑native VM for decentralized compute

https://spacekit.xyz
1•astorrivera•25m ago•0 comments

NotebookLM: The AI that only learns from you

https://byandrev.dev/en/blog/what-is-notebooklm
1•byandrev•25m ago•1 comments

Show HN: An open-source starter kit for developing with Postgres and ClickHouse

https://github.com/ClickHouse/postgres-clickhouse-stack
1•saisrirampur•26m ago•0 comments

Game Boy Advance d-pad capacitor measurements

https://gekkio.fi/blog/2026/game-boy-advance-d-pad-capacitor-measurements/
1•todsacerdoti•26m ago•0 comments

South Korean crypto firm accidentally sends $44B in bitcoins to users

https://www.reuters.com/world/asia-pacific/crypto-firm-accidentally-sends-44-billion-bitcoins-use...
2•layer8•27m ago•0 comments

Apache Poison Fountain

https://gist.github.com/jwakely/a511a5cab5eb36d088ecd1659fcee1d5
1•atomic128•29m ago•2 comments

Web.whatsapp.com appears to be having issues syncing and sending messages

http://web.whatsapp.com
1•sabujp•29m ago•2 comments

Google in Your Terminal

https://gogcli.sh/
1•johlo•31m ago•0 comments

Shannon: Claude Code for Pen Testing: #1 on Github today

https://github.com/KeygraphHQ/shannon
1•hendler•31m ago•0 comments

Anthropic: Latest Claude model finds more than 500 vulnerabilities

https://www.scworld.com/news/anthropic-latest-claude-model-finds-more-than-500-vulnerabilities
2•Bender•36m ago•0 comments

Brooklyn cemetery plans human composting option, stirring interest and debate

https://www.cbsnews.com/newyork/news/brooklyn-green-wood-cemetery-human-composting/
1•geox•36m ago•0 comments

Why the 'Strivers' Are Right

https://greyenlightenment.com/2026/02/03/the-strivers-were-right-all-along/
1•paulpauper•37m ago•0 comments

Brain Dumps as a Literary Form

https://davegriffith.substack.com/p/brain-dumps-as-a-literary-form
1•gmays•37m ago•0 comments

Agentic Coding and the Problem of Oracles

https://epkconsulting.substack.com/p/agentic-coding-and-the-problem-of
1•qingsworkshop•38m ago•0 comments
Open in hackernews

Life of an inference request (vLLM V1): How LLMs are served efficiently at scale

https://www.ubicloud.com/blog/life-of-an-inference-request-vllm-v1
175•samaysharma•7mo ago

Comments

0xjunhao•7mo ago
Hi, I'm the author of this post. Writing it was a great learning experience. I gained a lot of insight into vLLM. If you have any feedback or questions, feel free to drop a comment below!
criemen•7mo ago
Thanks for writing the article!

I didn't quite get

Note that during the prefill phase, all prompt tokens from a request can be processed in one batch. This is possible because the query (Q) tensors, calculated from the tokens immediately before them, are available for each prompt token position.

I know that in practice prefill is much faster than inference. Would watching the 2h video from Karpathy help me understand why?

criemen•7mo ago
And on the topic of prefill: Do you know what the role of GPUs is vs. in inference?
animan•7mo ago
Prefill is part of Inference. It's the first major step where you calculate all the keys and values for the input tokens.

Decode is the next major step where you start generating output tokens one at a time.

Both run on GPUs but have slightly different workloads

1. Prefill has very little I/o from VRAM to HBM and more compute 2. Decode is light on compute but have to I/o the keys and values computed in the prefill stage for every output token

dist-epoch•7mo ago
Doesn't decode also need to stream in the whole of the model weights, thus very I/O heavy?
0xjunhao•7mo ago
Yes, decoding is very I/O heavy. It has to stream in the whole of the model weights from HBM for every token decoded. However, that cost can be shared between the requests in the same batch. So if the system has more GPU RAM to hold larger batches, the I/O cost per request can be lowered.
animan•7mo ago
That snippet is trying to say that you can calculate KV for all the input tokens at once, and you don't need to loop over them since you have them all available.

Instead for decode, you need to sequentially generate each token.

longbeachbass•7mo ago
Thanks for this! Learnt a lot.

Curious to understand how do we ensure that the same model instance gets requests from the same client/user? Since conversations are stateful and the model needs context from previous turns of the conversation.

Is this happening at the load balancer layer?

cyanf•7mo ago
It's either sticky sessions or an lb that keeps track of prior sequences and route to the instance with the largest match. https://docs.sglang.ai/router/router.html
hhh•7mo ago
They’re not stateful, you submit the entire history with every call. Caching of prompts etc makes it important for performance to have sticky sessions or smth at the load balancer layer
0xjunhao•7mo ago
Yes, typically users send the newest user message and the full conversation history. These combined become the prompt.

Our API endpoint will try to route requests that has the same prefix to the same vLLM instance (similar to longest prefix matching in networking), and hopefully there are still some KV caches for part of the prompt there.

3abiton•7mo ago
Great write up, it would be interesting to see a lot of those covered features in comparison to other frameworks!
zackangelo•7mo ago
In your forward pass section you give a lot of emphasis to FlashAttention, but it might be worth mentioning Paged Attention as well (which was the paper written by the vLLM authors and I believe was the genesis of the project). PA-style block tables are now supported in most fused attention kernels, but vLLM originally came up with it and it's the main reason why vLLM has such high throughput!
0xjunhao•7mo ago
Thank you! We have incorporated your suggestion.
mhlakhani•7mo ago
Thanks for writing this up! I learnt a bunch from it. I noticed this didn’t discuss additional layers of caching - I can see how it would fit in, but is prompt caching out of the scope of this system?
gdiamos•7mo ago
Great write up. We use vLLM kv cache and continuous batching as a foundation for requests in ScalarLM and also add batching optimizations in a centralized queue and by adding explicit batching support in our client.

https://www.scalarlm.com

There is more perf you can sqeeuze out of vLLM

r0b05•7mo ago
Great write up!

Does batching add data from multiple requests into the same context, potentially decreasing perplexity? If so, are we trading off perplexity for lower operating costs?

ethan_smith•7mo ago
Batching in vLLM doesn't combine prompts into the same context - it processes separate requests in parallel while sharing compute resources, so there's no perplexity tradeoff, just efficiency gains.
zettabomb•7mo ago
It's worth noting that reason this works is because basically every LLM architecture currently in use is severely limited by memory bandwidth, not by compute. So it's trivial to run several requests at a time, while waiting for the next weights to arrive from VRAM.
StochasticLi•7mo ago
I would like to know what inference speeds they are achieving exactly on what hardware. I skimmed and searched the article and didn't find that info.
geoffbp•7mo ago
Thanks, good read!