Gustav RLo

3D Artist - Director of Assets - VFX Artist - Pipeline Developer - AI enthusiast.

[Gustav RLo @gustavrlo]

I wrote a song about my GPU refusing to run AI locally.

8 months ago it was a joke.
This week it's the news.
Not sure whether to be proud or scared...

@NVIDIAAI Save us from a future with no VRAM! https://t.co/FXwyG8mJpU

[Gustav RLo @gustavrlo]

@Tech2Wild woooo I need that app

[Gustav RLo @gustavrlo]

@bijanbowen Whaaaaa

[Gustav RLo @gustavrlo]

@0xSero I like omp such a capable one.

[Gustav RLo @gustavrlo]

@llixeey onto, preposition
outdo, verb
overdo, verb
like that?

[Gustav RLo @gustavrlo]

I was wondering what would be the collective of sparks. A "Shimmer" of sparks? or something more like a "Flurry of sparks? I don't quite like "cascade" or "shower".
Here are my favorites what is yours:

a shimmer of sparks
a dazzle of sparks
a mischief of sparks
a riot of sparks

[Gustav RLo @gustavrlo]

@OnlyPromptsxAi A prompt to help me fix graphics in a 2D Sprite game in Unreal engine 5.8 using paper 2D. Prompt for all common issues that can happen when using Paper 2D.

[Gustav RLo @gustavrlo]

The biggest question of this season: "To buy or not to buy?"
GPUs are sold out almost everywhere. And honestly? With the amazing open models dropping right now, it's obvious this hardware has never been more valuable.
But that raises new questions…
The new Mac Studio? 512GB of unified memory is a slap in the face to Nvidia — nothing they offer comes close on memory.
So the looming question hangs in the air: will Nvidia even try to compete? Or are they comfortable resting on their laurels?
Buy Nvidia? Buy Apple Silicon? Or wait?
But here's the catch — if the memory drought doesn't end soon, waiting might mean nobody can afford anything at all. 🧵👇

[Gustav RLo @gustavrlo]

@needmorevram OMG HOW?

[Gustav RLo @gustavrlo]

@shuklareal_ @grok They are likely deodorant marks. Try to rub some jeans on them. If they disappear it was deodorant.

[Gustav RLo @gustavrlo]

wow I wonder if we going to ever get numbers like these on TP=2

[Matt Mastracci @mmastrac]

There is a _lot_ of optimization potential left in GLM 5.3 Flash on DGX Sparks. I'm excited to share my MTP approach - I'll be upstreaming it to vLLM once it's fully tested.

Already prefilling almost double what I was before the experiment (!).

TP=4 for now, but stay tuned.

[Gustav RLo @gustavrlo]

@BroWire_ that prank can blind a kid, or worse.

[Gustav RLo @gustavrlo]

@ThePeterMick the premium ++, but hey, no subscription.

[Gustav RLo @gustavrlo]

@MegaMarian12350 wow. so sad

[Gustav RLo @gustavrlo]

@SlickmediaX De mulher pra mulher Marisa.

[Gustav RLo @gustavrlo]

@Shahlin_ibrahim @ThePeterMick LOL exactly what I was thinking.

[mr_r0b0t @mr_r0b0t]

r0b0tlab/GLM-5.3-Flash-EXL3-2.25bpw-sm121 + 3.00bpw DFlash2 drafter + native ExLlamaV3 runtime!

Tested on a single @NVIDIAAI
GB10 ♥️

At 8K C1, code: 22.61→49.96 tok/s with K=5
Exact single-key retrieval at 259,993 prompt tokens
Extended eval suite in progress 🤓

[Gustav RLo @gustavrlo]

@MiaAI_lab @christomitov @Zai_org 🤞

[Gustav RLo @gustavrlo]

People are addicted to "Us against them" mentality, that is not productive.

I would not have had the money to have used the 6 Billion tokens from the past 2 months if I only had API to use.

At least 1-3 B were from Mia's recipe for GLM 5.3 Flash.

1-2 were smaller models for smaller tasks.

around 1.2B was a combination of Claude, Astra and Deep Seek V4 and V4.1 Flash (API).

I got over 3 times the volume of work that I would have done for less.

Not seeing things from this angle is just wasteful.

[Gustav RLo @gustavrlo]

@WescheNex1q If this could do GLM 5.3 Flash with these speed gains, would probably become 2x cluster DGX Spark main for most everyone.