Christopher Owen

terminally curious. ai-pilled.

[Christopher Owen @christopherowen]

@tenhkspark please do, or I welcome a collaboration?

https://t.co/EOx0AuCX10

we could rename the repo :)

[Christopher Owen @christopherowen]

@0xSero do you have a repo of your setup available for review?

[Christopher Owen @christopherowen]

@mfranz_on but when it starts inference and you can hear the thermal paste (or whatever) pop it's a bit distressing.

[Christopher Owen @christopherowen]

I wish that double clicking the title bar in macOS 27 would maximize the window like it did in 26.

[Christopher Owen @christopherowen]

@qubitium Are you experiencing context rot? What are you using for inference and what are you using for compaction?

[Christopher Owen @christopherowen]

There are many DeepSeek-on-Spark projects like it, but this one is mine.

3× DGX Spark. Switchless CX7 ring. Native DS4.1F weights. Upstream-aligned vLLM. Some unique patches.

92 tok/s aggregate at 8× concurrency.

Now it can be yours:
https://t.co/EOx0AuCX10

[Christopher Owen @christopherowen]

what would you do with a full local copy of Arxiv?

[Christopher Owen @christopherowen]

Thanks for posting! I still haven't put my 3x switchless spark ds41f repo up.

Every time I come across a repo like yours I take a quick scan for what I can do to improve mine.

Have you tried RoCEnante? I was able to get a couple percent improvement over NCCL for shorter cx lengths.

[Christopher Owen @christopherowen]

@jnardiello @antirez only think missing is a bit of quick storage. But I guess that depends on the Mac mini specs!

great setup.

[Christopher Owen @christopherowen]

@Hikari_07_jp I went with the Ugreen 480t and nvme drives. It's super quiet.

Be careful with the OS that these providers ship. For example uGreen made extensions to the operating system that meant mounting it elsewhere wasn't possible without recovery. Just install Ubuntu.

[Christopher Owen @christopherowen]

@burkov .5 days? What a timeline we live in :).

Back in the day we would spend weeks to setup something much less ;).

It took me about 1.5 days to get it working, btw.

[Christopher Owen @christopherowen]

@iam_zachi this has not been my experience.

[Christopher Owen @christopherowen]

@8my41 @iam_zachi which abliteration are you using? did you find that it was necessary compared to the normal weights? did it have any unintended consequences on intelligence?

[Christopher Owen @christopherowen]

@oldstackjournal ls -lasR

[Christopher Owen @christopherowen]

@trawasthi_ai Why the gatekeeping? It's great that more people are learning and trying.

The messages, when written by their AIs, do often sound a little over-confident though.

[Christopher Owen @christopherowen]

It’s just for me :). It is literally next to my tv in the living room.

Total throughput divided by concurrent threads is lower than the performance of a single session for sure, though. I think this is “the nature of the beast”.

I am using the full weights from DS. No extra quant.

[Christopher Owen @christopherowen]

@cozybearlog @Indy_triguy what's your setup like?

[Christopher Owen @christopherowen]

@mfranz_on 3 is enough with native quants, depending on context length.

[Christopher Owen @christopherowen]

@cozybearlog @Indy_triguy I'm getting about 90 over 4 concurrents and mid 40s on 1.

[Christopher Owen @christopherowen]

@tenhkspark nice!