Steven Martin

Manage IT for private equity company.
Pro-Choice Pro-Science Pro-Wrestling

[ÆON FORGE ✨ @SpaceTimeViking]

Built an Uncensored / Abliterated version of Gemma 4 12B that actually scored HIGHER on OpenAI Human Eval CODING over the base official 12B model.

Currently only in BF16 but NVFP4 Quantization coming in hot. Will evaluate the quantized model to see if we get similar results.

https://t.co/6TjStamWvu

[Steven Martin @StevenM_42]

@microcenter Thanks for the reply! I took them back the next day and swapped out for new ones that did in fact have the cover in them! Very easy and handled well by staff!

[Steven Martin @StevenM_42]

@nvidiacc Two of the four sparks I purchased were missing the magnetic covers. This seems like a QA issue at the factory. Online chat said to return them to @microcenter.

Can't you just mail me a couple $5 covers? I'm trying to expand this cluster!

[Steven Martin @StevenM_42]

@aijoey Im currently running alot of benchmarks using this llama.cpp forked for the spark:
https://t.co/yZD9T6nJaD
(i created two issues in the repo to get it working)

That plus "Nemotron 120B Q4_K_M" is looking real interesting atm...Not done yet, but here's a random chart!

[Steven Martin @StevenM_42]

Comparing Gemma4 and Qwen3.6 with vLLM bench serve using the ShareGPT V3 dataset on the DGX Spark.

[Steven Martin @StevenM_42]

@SpaceTimeViking I have two Sparks atm, may buy more if I can justify them. I'm working on putting together a benchmark suite cobbled together from various people's repos... I'll be sure to post some results once I have them.

[Steven Martin @StevenM_42]

@SpaceTimeViking
As a new Spark owner, I appreciate all your hard work!

You seem to be using NVFP4 on your github releases for Spark. Have you tried any other variants lately?

[Jonathan Yantis @mantis1]

@aijoey Have you tested fp4 and non-nvfp4 variants?

I’m running a highly optimized 122B model, and the most shocking thing I learned on the forums is the sm121 arch sucks at nvfp4. YMMV.

Check out the Qwen Intel autoround models, shocking performance increases.

[Steven Martin @StevenM_42]

@aijoey @KrystianLew Is it in your github? Id be interested in giving it a try. Been looking for a decent benchmark tool.

[Steven Martin @StevenM_42]

@brian_armstrong

[Steven Martin @StevenM_42]

@WildMediaOnly he played the hand 100% exactly correct, LOL

[Steven Martin @StevenM_42]

@Osinttechnical @vcdgf555 DRRRRRR..... how much does the missile cost?!?!???! DRRRRR

[Matt Harrison @__mharrison__]

For my friends who are still using UV and might be a little weary about recent compromises to PyPi packages, stick this in your pyproject.toml.

You can let all of those pip users find and report the compromises...

[Steven Martin @StevenM_42]

@rryssf

[Alabi @the_Lawrenz]

Bro 😳😳

[El Señor De Los Memes @UltimoLegend]

What a rollercoaster that was

[TacticalStNick @TacticalStNick]

Hahahaha. His Stone Cold impression is spot on. @steveaustinBSR

[Steven Martin @StevenM_42]

@tonysheng i asked Claude to adapt it for Windows... it works.

[Dan @robustus]

Turns out with claude code, my decades long strategy of NOT deeply learning:
- regexs
- sql
- nginx confs
- elaborate shell commands
- advanced shell scripting
- any javascript framework
- perf optimization
- webpack, cdns, bundlers
- 1000 other things

...was entirely correct.

[Steven Martin @StevenM_42]

@windscribecom Love Windscribe! Been using it for over five years!