Uwe Schuster

[Uwe Schuster @UScLE]

@TeksEdge Trying with a rented RTX Pro 4500 is on my TODO list to see if it is faster than the Unsloth Q4 K XL MTP.

[Uwe Schuster @UScLE]

@Powermikey2206 Kein Schlangen und ja, richtig der Preis heute Abend war bereits über 20 Cent über dem Preis von heute früh.

[Uwe Schuster @UScLE]

@CWRoehl Sie wären damit nicht alleine.

[Uwe Schuster @UScLE]

@AlexFinn Especially with the additional RTX Pro 6000 this sounds like a dream machine from my POV. Which version of the 6000? Workstation or Max-Q? Any plans regarding the cooling? Seen sometime ago someone changing its dual 5090 setup to watercooling, because one of the cards got too hot

[Uwe Schuster @UScLE]

@ItakGol LOL

[Uwe Schuster @UScLE]

@UnslothAI Qwen 3.6 27B

[Uwe Schuster @UScLE]

@UlisesDavid__ Liegt wohl eher am Kameramann der die "Statistik" verfälsch :-)

[Uwe Schuster @UScLE]

@airoid1111 Do not know the context from the text, but did not know that japanese woman are that curvy.

[Uwe Schuster @UScLE]

@Lars_Kruse_1973 @olivergorus Sobald es schwierig wird und die KI sich im Kreis dreht sind alle raus die den Code nicht verstehen. Da gilt dann "a fool with a tool is still fool".

[Uwe Schuster @UScLE]

@aktien_max Nein, Konkurrenz belebt bekanntlich das Geschäft und besser nichts veröffentlichen (Roadster 2.0) als was schlechtes.

[Uwe Schuster @UScLE]

@RonnyGolisch Deutschland wartet, bis die KI gelernt hat, ein Fax zu senden. (Kudos an Gemini für alternative Vorschläge zu "Bürokratie != FSD")

[Uwe Schuster @UScLE]

@TeksEdge Looking forward to it

[Uwe Schuster @UScLE]

@rumgewieselt @UnslothAI Nothing changed anything...till I have got Nvidias nccl-tests compiling + running with arch 60 and pushed harder. The missing piece was the llama.cpp param "--split-mode row". Hear coil whine, almost 20 t/s with two P100 (x8 and x4) and almost 22 t/s with the GTX 1080 (x8) added.

[Uwe Schuster @UScLE]

@rumgewieselt @UnslothAI Thank you for the hints, I will try them and especially nccl sounds promising / sounds like the missing piece.

[Uwe Schuster @UScLE]

@rumgewieselt @UnslothAI I see 17 t/s using Qwen3.6-27B-Q4_K_M.gguf from unsloth/Qwen3.6-27B-MTP-GGUF on Ubuntu 22 with two P100, one at 8x and one at 4x.
No idea if the code of the PR changed the --spec-type param value is draft-mtp. I have built llama.cpp againts cuda toolkit 12.6.3+architectures 60,61

[Uwe Schuster @UScLE]

@RRunner144 Can you expand on that? Your image and 3090s does not match.

[Uwe Schuster @UScLE]

@joaosump @AiFreak_ Keep us posted about your 384 GB VRAM monster

[Uwe Schuster @UScLE]

@SawyerMerritt I have questions! Does this mean @Tesla_AI is not longer compute constraint and renting Collosus 1 would not speed up FSD training?

[Uwe Schuster @UScLE]

@dholzric @bijanbowen BTW, I would not have said anything with two 5.25" drives. :-) I have had an internship back in school days as data typist on such a machine. https://t.co/oQWYtp8ogY