David Wang

gpu perf @modal

[David Wang @_dcw02]

@nanjiangwill 🤝

[David Wang @_dcw02]

me with 0 commits to the modal monorepo 👀

[Erik Bernhardsson @bernhardsson]

@schrockn Yeah agree PR count is slightly better metric than LoC

[David Wang @_dcw02]

@KaichaoYou yes

[David Wang @_dcw02]

@rfleury ah yes the OS created by Linyos Torovoltos

[David Wang @_dcw02]

@charles_irl @tenderizzation I think even my lmod configs blow up if I touch them nowadays

[David Wang @_dcw02]

i use an nvl72 as a devbox btw
https://t.co/1YZhOGVH4y

[Modal @modal]

https://t.co/5DZKxwdGY5

[Charles 🎉 Frye @charles_irl]

Added a fun lil widget to the LLM Engineer's Almanac -- a "Token Timing Simulator" so you can get a visceral feel for what a benchmark perf number means.

Here's @_dcw02's latest work with @zhijianliu_'s DFlash technique in @sgl_project -- ~1k TPS!

https://t.co/iUJm984dq0

[David Wang @_dcw02]

gpu poor? inframaxx
gpu rich? also inframaxx
gpu middle class? believe it or not, still inframaxx

[David Wang @_dcw02]

tried kimi k2.6, really shows the literary heritage gap

[David Wang @_dcw02]

my goat @AhanGupta13 !!!

[PyTorch @PyTorch]

Want to train LLMs on longer contexts without re-engineering your entire systems stack?

Introducing AutoSP — the first compiler-based solution that automatically optimizes LLM training for long contexts. Under the hood, AutoSP applies a series of compiler passes that trigger sequence parallelism, paired with a curated activation-checkpointing scheme tailored for long-context training. It's integrated directly into DeepSpeed, so enabling long-context training is just a config change away.

No more rewiring your stack to push context lengths. Read the blog to learn more 🖇️ https://t.co/TMjWfsO8fy

✍ @AhanGupta13, Zhihao W., Neel Dani, @toh_tana, Tunji Ruwase, @_Minjia_Zhang_

#PyTorch #DeepSpeed #AutoSP #OpenSourceAI

[David Wang @_dcw02]

@tenderizzation Sent from my iPhone

[David Wang @_dcw02]

as AI ascends, so too may we descend back to the asm mines

[Charles 🎉 Frye @charles_irl]

https://t.co/A0LexEUZG1

"Everything is open source if you can read assembly" -- @_dcw02

[David Wang @_dcw02]

@menhguin I have a pretty good guess what rationalists think of Omelas

[David Wang @_dcw02]

@vikhyatk just drink more?

[David Wang @_dcw02]

@ekzhang1 hope you feel better soon!

[Zhijian Liu @zhijianliu_]

DFlash⚡ meets OpenClaw🦞 = FlashClaw

Same Claw. >4X faster or cheaper.
DFlash support for Qwen3.5 is live — outperforming native MTP by up to 2.3X.

More to come! 🔥

[David Wang @_dcw02]

@vikhyatk @realmcore_ blind (codex) leading the blind (me)

[David Wang @_dcw02]

@Dorialexander @AmpCode we have instructions to set up opencode with our glm5 endpoint here: https://t.co/M0u0SjWGWd :)

[Robert Clausecker @FUZxxl]

@lisperati Just write assembly code. That has always been allowed.

[Charles 🎉 Frye @charles_irl]

The GLM models by @Zai_org have been a gamechanger for me. I was reluctant to embrace coding agents before I could run the models myself.

Now, with GLM-5, I have a top-quality self-hosted intelligence endpoint tightly integrated into my engineering work.

https://t.co/95XSG31K0P