jackAL

Enjoy my free time. 农民🧑‍🌾 山寨价值为0
Fate is not in the heavens, but in your breath and your gaze

[jackAL @jackalxkh]

交易的核心精髓是什么？
你的交易策略：要能接的住大运气。
你的资金管理：要能等的到大运来临。

[AB Kuai.Dong @_FORAB]

社区普遍反馈，小红书已软封禁散帅一词，而集美仍然正常显示推流，如果在该平台上发布带有散帅的内容，则会被限流。

散帅是集美的男性视角对应词，前阵子突然爆火。

[大虾老师（成教授） @yeweiyangfou]

@qin88484 完全没有气象常识的人用AI制作了一个低级的视频

[秦剑波（民权党筹建人） @qin88484]

台湾旅客搭乘MM929航由冲绳回台湾，
在飞机上看到的巴威台风。

上帝保佑

[Lotto @LottoLabs]

Word on the street is new qwen open source models are coming out but they’re skipping 3.7

So it’ll be 3.8 or 4 later this year

Sweet sweet 27b

[max @maxxfuu]

Day 6/90 of Inference Engineering

I wrote a CUDA kernel for 1D Convolution, just getting the reps in for writing unoptimized boring but correct CUDA! I also read PagedAttention (2312.07104), watched 1 short lecture on GPU Memory, and hack squatted 455lb x6 and 495lb x3 as a top set!

Heres what I've learned about PagedAttention:

PagedAttention reduces internal fragmentation and solves external fragmentation.

During the decode phase, the KV cache manager actually hands out one cache block at a time to store output tokens. If the output tokens overflows the cache block, the KV cache manager assigns a new cache block to hold the remaining tokens that didn't fit into the previous cache block. What ends up happening is that the VRAM is used effectively such that each cache block is filled before new cache block is allocated on the VRAM.

In a naive KV cache implementation, memory is reserved up front for the maximum sequence length, which might not ever be used, meaning there is just a massive chunk of VRAM that's not touched.

PagedAttention reduces internal fragmentation by ensuring memory is never wasted between each decode sequence.

To make usage of memory even more efficient, any unique prompts that share the same prefix tokens, the KV cache manager allows them to share the same cache block, however, it writes to a new cache block starting from where the new tokens differ within the two prompts.

While each decode phase effectively writes to random parts of the VRAM, the Block table is able to provide an abstraction that makes KV cache seem like a contiguous memory to the model. Making the allocation process as simple as checking if a cache block is full or not.

This continuous practice of using cache blocks to allocate just enough memory and using a block table to allocate non-contiguous free scattered memory on the VRAM reduces internal fragmentation and solves external fragmentation.

---

Watched a video on GPU Memory: https://t.co/2TySFdj1im

I read this blog that made absolutely cemented everything, written by (@hamzaelshafie): https://t.co/GEzvexo7pl

[keys 🧪 @u1tra_instinct]

🚨BREAKING!!:

Full GLM 5.2 Quantrio INT4/INT8 mixed - now Abliterated closed to 100%. C1 = ~30 toks/sec at 128K CTX. Ran on Four-DGX Spark cluster.

I finally beat the final boss, maxed out all the token burn, and not NVME drive start to fail on one of the machine to bring this for free open source for everyone.

if you like my work please donate to:
https://t.co/hm9sQkQe0P

Here is the Github Repo:
https://t.co/eTmgmAngvl

Here is the Hugginfac repo for weights:
https://t.co/l9KMf0NMX6

[maria @maria_rcks]

I solved the model picker problem, no need to thank me Tibo

[Hikari∣LocalLLM⚡ @Hikari_07_jp]

35B-A6B Model Development Progress Report!!
In A6B, we're expanding the number of experts used from 8 to 32. Previously we treated all 32 experts equally, but I've come up with the idea of prioritizing the top 8 experts. I've updated the research.
https://t.co/yRagBZ2U1e

[h100envy @h100envy]

Ex-vLLM core contributor explained how to make LLM inference 10x cheaper in 34 minutes - better than $3000 inference optimization bootcamps.

request comes in -> check LMCache -> hit? load KV cache from CPU/SSD/remote -> skip prefill -> serve.

That loop is why Bloomberg and other production stacks now push 300 terabytes of KV cache per week.

LMCache + vLLM + CPU/SSD/remote storage + zero-copy CUDA kernels - that's the stack.

Watch and save it, then wire the KV-offload into your inference stack.

[Mia @MiaAI_lab]

Run the new @UnslothAI Qwen3.6-35b-NVFP4 on your @NVIDIAAI DGX Spark ease! ✨

256K context • 24 concurrency

~81 tok/s single session
~350+ tok/s at 24 sessions

This is now the recommended Qwen3.6-35B terms of accuracy/performance.

Get it here 👇
https://t.co/ULJkq3A7rN

[Vincent @Vincent_AINotes]

刚刚，科研圈被一个名为 The Well 的开源项目炸开了锅 🤯。

有人把原本需要国家级实验室耗费数百万美元、用超算跑上几周才能得出的高精度物理仿真结果，足足打包了 15TB，直接丢到了网上，100% 免费开源！

这是人类顶尖算力的“数字双胞胎”合集 👇

[Richard Gage, AIA, Architect @RichardGage_911]

[RG911Team] This fact alone proves 9/11 was a false flag.

The Twin Towers were designed with the lightest, thinnest steel on the top floors, and each lower section used thicker, heavier and stronger steel than the one above it. This made a progressive collapse IMPOSSIBLE.

The idea that a thin, light steel beam would collapse and then crush thick, strong steel below it… and that would in turn crush the even thicker and stronger steel below it… all through the path of GREATEST resistance… all in less than 12 seconds… violates the laws of physics.

This is why FDNY sent hundreds of their best firefighters in those buildings. They knew the planes and fires could not completely collapse the towers because the extremely robust design would easily withstand it.

343 firefighters died on 9/11 because they were caught off guard by something no one expected - every single column being ripped apart and blown in every direction.

And the only force that can do that requires explosives.

[Nora @Heal_within96]

What no one tells you about the Bosnian genocide is how some rich White people from Italy, France and the UK paid upwards of $90k and extra to kill children for 'snipper hunts.'

[Aakash Gupta @aakashgupta]

Tokyo has a building whose entire job is to air condition other buildings. One plant in Shinjuku produces 65,000 tons of cooling capacity, roughly the output of 20,000 home AC units, and pipes chilled water underground to more than 20 skyscrapers at once.

The mechanism is called district cooling. Rather than every tower installing its own chillers and cooling towers, one central plant chills water with massive compressors and steam absorption chillers, then pushes it through insulated pipes beneath the streets. Each skyscraper just runs a heat exchanger. That giant fan visible from the observation deck is one building rejecting the heat of an entire district.

The math is why it wins. A chiller sized for one building has to survive that building's single worst hour. A plant sized for 20 buildings shares capacity, because offices, hotels, and department stores all hit peak load at different times. Central plants also run machines far bigger and more efficient than anything that fits on a rooftop, and they free the top floors of every connected tower, some of the most valuable real estate on the planet.

The plant generates its own electricity with gas turbines too. When the grid fails, it keeps cooling and powering the Tokyo Metropolitan Government building next door. Tokyo's disaster command center rides out blackouts on a neighborhood air conditioner.

The system started in 1971, before most of the skyscrapers it now serves existed. Tokyo laid the cooling grid first, then built the skyline on top of it.

[Adel El Zabayar @Zabayar]

Irán llevaba décadas sin lluvia. Luego que Irán destruyera los radares que algunos resultaron ser sistema de manipulación ambiental HAARP estadounidense e israelí,
ahora en Irán no deja de llover.

[NIK @ns123abc]

🚨OPENAI ALLEGEDLY STOLE APPLE’S ENTIRE PLAYBOOK

“Hundreds of billions of dollars, decades of work”

Hardware Engineering

- circuit designs, component architecture, power management
- unreleased products guarded by internal codenames
- AI/ML integration for hardware
- EMI engineering + testing methods

Manufacturing Secrets

- proprietary processes + custom machinery
- equipment Apple builds and installs IN supplier factories
- proprietary metal alloys + finishing techniques
- DFM (Design for Manufacturability) expertise

Component Technologies

- power chips, battery systems, displays, acoustics, touch
- identities of specialized sub-suppliers
- exact specs Apple demands from components

Testing Data

- failure analyses, lifecycle simulations
- “negative know-how” — everything Apple tried that DIDN’T work

Supply Chain

- supplier contracts, allocation strategies
- global logistics coordination
- systems-level integration

ALL GONE 💀

[arb8020 @arb8020]

i love this visualization of nvpf4 from humans&

[nic @nicdunz]

Terra Light → Luna High: smarter, cheaper.
Terra Medium → Luna High: equally smart, cheaper.
Terra High → Luna XHigh: equally smart, cheaper.
Terra Extra High → Sol Medium: smarter, cheaper.
Terra Max → Sol High: smarter, cheaper.

[iwsfutcmd @iwsfutcmd]

the Chinese are way ahead of you, buddy

we're currently in the 小暑 (Lesser Heat)

[Jukan @jukan05]

One of NVIDIA’s comments during its Morgan Stanley NDR: Anthropic’s share of compute running on NVIDIA has now risen to nearly 50%.