Mark Sunner

retired geek happily tinkering. waaay out of my depth - and loving it😜

mainly here for AI / low-key lurking & learning 🔬📚

no followers ≠ fake

[Mark Sunner @MarkSunner]

We’re witnessing the calcification of a two tier society in real time ..and make no mistake, we’re NOT invited because we’re on the menu.

I can’t tell whether Dario is trying to convince others or himself of this verifiable falsehood - it might be both? But either way its ugly, manipulative and reeks of self interest.

Fearing obvious future price/moat erosion Anthropic are clearly choosing to pull up the ladder over innovation. If you weren’t a fan of Open Source before, I urge you to consider supporting it now - no one else is coming to save you.

Ubuntu means “I am because we are” ..definitely feeling that energy right now ✊

[Mark Sunner @MarkSunner]

@ciprianveg Omg. I am in awe 🤯

What is the best (QSFP) switch solution for this gorgeous beast??

[Mark Sunner @MarkSunner]

@unseenmars_ 2 x DSv4F 0731 dual node clusters

Or

1x GLM 5.2 4 node cluster

Decisions decisions 🏗️🤔

(❤️this platform more each day!!)

[Mark Sunner @MarkSunner]

@Tech2Wild Genius metaphor 💡👌

[Wësche @WescheNex1q]

Me this morning

[Mark Sunner @MarkSunner]

@Blackwellboy OMG - On it 🫡

If this works half as well as I hope it going to work then this could get expensive 🤪

[Mark Sunner @MarkSunner]

@FishAudio Amazing - can’t wait to try this 🙌🗣️

[Fish Audio @FishAudio]

Today we’ve raised $52M Seed and we are announcing the public launch of S2.1 Pro.

>It can clone a voice from 5 seconds of audio
>2x faster than Cartesia & 1/6th the cost of Eleven Labs
>most expressive model with word level control over emotion, intonation, pacing etc

We support frontier AI companies including HeyGen, LiveKit, Retell, Sanas, and OpenArt all run our model in production.

If you're a business and we can't cut your voice AI costs by 50%, we'll give you 1 year of Fish Audio for free.
Book a demo: https://t.co/vHkyZf9JoG

To celebrate our first birthday, we'll give you 1 month of S2.1 Pro for free. Like, retweet, and comment “Fish” to get it.

[Mark Sunner @MarkSunner]

I’m using GLM on 4 DGX to drive OpenCode with DSv4F driving Hermes for orchestration (on two separate sparks). But the marriage of GLMs brain with DS’s big context is amazing!

So GLM puts the jigsaw together whilst DS always comprehends the picture on the box. For me at least they’re a great combo.

[Mark Sunner @MarkSunner]

@fogle_ryan @nvidia @NVIDIAAI

[Mia @MiaAI_lab]

You can now run GLM-5.2 on 3x @NVIDIAAI DGX Sparks at home 🤯

• 248k context
• 15-20 tok/s (content-dependent)
• MTP for max context / DSpark optional
• NVFP4+AQLM hybrid

Still a work in progress — will keep improving over time.

→ https://t.co/JAvrJL92h4

[Mark Sunner @MarkSunner]

@mfranz_on I *love* it - but does it have any heat dissipation? 4 sparks (especially if working as a cluster) get HOT 🥵🌡️

A little fan system to pull the heat away from rig and help vent it would be a brilliant addition imho

[Mark Sunner @MarkSunner]

@Tech2Wild Woah - that’s awesome 🤯

Would love to understand the architectural pipeline behind this if you wouldn’t mind sharing?

Does it go as far as posting to SM Itself?

[Mark Sunner @MarkSunner]

Ahh interesting. My setup is slightly different: I'm using Hermes as an agent, but it's running DSv4F for the 1M context and acting as orchestrator for OpenCode, which is where I have GLM wired in, so all my tool calling takes place server side.

But if I'm reading you correctly, Hermes' tool-call validation is rejecting some of GLM's call formats and giving up at its 3-retry cap, which absolutely kills long-horizon momentum. Rather than patching the app side, it's worth checking your vLLM serve config has the right tool-call parser enabled for GLM: the serve side can normalize its tool calls into the OpenAI-style shape harnesses expect, which often calms guardrails with no app changes at all.

On the deadlock: one incident fits the pattern exactly, it's probabilistic and cold-prefill-weighted, so the dice are rolling. I'd apply the attention-route fix from the write-up before it catches you mid-task, it's one flag plus drop-ins. And agreed on Tony, rock-star is the right word 🫡

[Mark Sunner @MarkSunner]

Hi Gene, they're the same lineage! My setup is built directly on tonyd2wild's QuantTrio recipe (that's the int4-int8mix quant and serve config I run): https://t.co/k0FK3j5MKl

What my write-up adds is what we found underneath it during sustained load with OpenCode: the deadlocks trace to a FlashInfer sparse-MLA kernel livelock on GB10 (captured with cuda-gdb, receipts in the repo), and routing the main model's attention to the Triton sparse-MLA path eliminated it, 1000+ context-ceiling sessions clean since, full 200K validated.

So the path I'd suggest: start with Tony's recipe. Then, if you hit any stability issues apply the attention-route fix from our write-up, and you’ll get the fab setup minus the wedges. All credit to Tony and the QuantTrio community, our work just stands on their shoulders.👍

[Mark Sunner @MarkSunner]

@danushman 🧑‍🍳😘

[Mark Sunner @MarkSunner]

@sudoingX DSv4F on dual Sparks needs to be experienced to be believed. It's one of the rare cases where I can say with some confidence that, no matter what your use case, you will be delighted with how well this works / performs. ⚡️⚡️🤖

[Mark Sunner @MarkSunner]

@Tech2Wild Great work as always Tony 💪

Out of interest did you experience any deadlocks during testing?

[Mark Sunner @MarkSunner]

@light_foundry Out of interest did you experience any deadlocks during testing?