Papamasher (@Papamasher) profile picture

Old man who plays big body grapplers badly

Joined July 2023
Photo by Papamasher (@Papamasher)Photo by Papamasher (@Papamasher)Photo by Papamasher (@Papamasher)Photo by Papamasher (@Papamasher)Photo by Papamasher (@Papamasher)Photo by Papamasher (@Papamasher)
Why in the age of AI do we have people worried about how they will heat their homes? Surely - distributing GPU's to everyone then using them to heat homes would solve two massive problems?
4
I'm starting to worry that Ablit models are going to be the reason we get legislation around running your own llms 😧
7
Qwen3.6 27Bは量子化によりどのくらい品質劣化するのか?(Quesmaより) 37時間・約390万tokenを費やし、 複数の量子化モデルを性能を比較したらしい。 ■元モデルとの出力差 Q8はBF16とほぼ同じ。 Q4までは差が小さく、Q3以下からズレが急に大きくなった。 ■SVG画像生成 ペリカンや歯車のSVGを生成して比較。 BF16〜Q4は、どれが高精度モデルか判別できないほど差が小さかった。 Q3は量子化方式によって差があり、Q2では明確に品質が落ちた。 ■数学問題 Q4以上はBF16とほぼ同等。 Q3は方式による差が大きく、 ・Q3_K_M:73.3% ・Q3_K_S:54.2% 同じ3bitでも、量子化方式によって性能が大きく変わる。 ■生成速度 Apple Siliconでは、量子化しても生成速度はあまり変わらなかった。 主なメリットは、モデルのサイズとメモリ使用量を減らせること。 ■まとめ Qwen 3.6 27Bは Q4_K_MかUD-Q4_K_XLまでなら 品質劣化が少ないかもしれない。 →16GB VRAM民  Q4サイズなかなかきついよ😓 #VRAM飢饉救済教
11
22
4
159
15,800
Is anyone having success running Laguna 2.1 FP8 on dual sparks? For me spec decoding seems to cause it to flatline if i go over 128k contexts, i've bumped memory and had a couple of models review, Sol/5.5 and Deepseek v4 flash (who got the closest to having it run well) - ideas?
22
Ok Local AI folks with the Dual spark setup. Laguna 2.1 twice, once on each node OR Dual sparks serving a single instance? I'm intrigued what people are doing atm? Dabbling with idea of Laguna on one and qwen 3.6 27b on the other - hermes MoA in the middle..
19
After Running Step 3.7 (NVFP4) on various dev tasks with deep contexts i can say - it's really damn stable and produces high quality work. Hallucinations are rare but when they do its bad - i had it crash out of one thought process claiming that i had insulted it. mid stream
13
Every time i try to get Spark 3.7 up and running across Dual Sparks i'm just getting gibberish responses - any of you smart people hit this problem and found the solution?
19
If this is Deepseek's preview I'm looking forward to the full version. With dual sparks it has been the strongest performing model for me so far.
DeepSeek V4 is coming. Mid-July. Yeah yeah you might think we had V4 for over 2 months already, but no, that was "preview of V4". They expect a lot of demand, and introduce a new mechanism: peak hour pricing. It doubles what we have now. The baseline is the same though.
22
Big Thank you to @MiaAI_lab the deepseek 1M context config works a treat. Hopefully I can get your Step3.7-NVFP4 working as well - every other instance (looking at u eugr) I've tried has yielded complete gibberish for responses. One question - why deepseek needed 20mins to boot?
12
I'm using Qwen3.6 27b NVF4P Text only for coding agents on DGX spark (ASUS-GX10). huggingface.co/sakamakismile… about 20 tps on each agent, seems pretty solid.
1
18
After further review and comparison i do not recommend this version of 27b - seems to have suffered badly under compression - results from prolonged attempts at work have not been good.
9
the Crossy road prompt is now going to be my defacto test of the impact of KV quantisation. The difference between the quants is wild. I'll never trust turbo2 with anything :D
1
7
Papamasher retweeted
I am confident with a month notice I could smoke almost everyone in the FGC on a 40yd dash
5
48
1
793
38,823
The time has come. Rollback netcode has been implemented in DRAGON BALL FighterZ! Thank you for your patience and support! #DBFZ More information: spr.ly/6011X1zDS
73
605
225
2,591
400,034
Papamasher retweeted
Replying to @unity
how can i trust that you'l be more transparent If you feel the need to hide replies that contain superficial things like jabs to your past as well as compliments? If you prioritize your brand image over transparecy, how can i trust that you wont hide more serious things from us?
2
3
184
16,561