Unsloth as just released Unsloth Studio. It looks dead easy to use to train/fine-tune models on our GB10s.
You can install Studio on a separate Ubuntu system from the Spark system, and then perform the inference process on the Spark system. Thanks in advance.
i should check in with the boys .. i worked on studio for unsloth about half a year ago . roland tooked over most of the efforts .. maybe the spark needs a custom container. but this looks like a simple python dependency missing. this should be as easy as hopping into there discord . i can check the coming days but my week is packed rn
Thank you!
It’s good to see the great work done on this to make it easy for us!
I also posted this on the discord and Roland actually got pinged.
He told me to wait for the support for the spark because it needs some stuffs compiled for it (probably duo to arm architecture?). And aparently there is no spark for them to test on.
@jensenhuang please sponsor the team :))
i can talk to roland and give him access to mine if he needs it .. let me check with him
Hi. You can find torchcodec precompiled for DGX Spark (together with other wheels) here: GitHub - mamorett/Nvidia-DGX-prebuild-wheels: some precompiled wheels to be used with new Nvidia Spark DGX · GitHub .
Is the current 1-line installer working on DGX Spark? Any workaround if not?
I was able to get an earlier version installed, and also upgrade to the latest version, but I don’t think it’s ready for prime time in general: 1. The model picker is weird about selecting quants, it seems like you only get to pick one variant, and then it will only load that one in the future. 2. I don’t think downloads are using my HF token, larger model downloads would hang after a while.
For the Spark specifically, seems like something is wrong: Once I got a 27B model downloaded, and tried to load it, the progress bar in the log output was moving extremely slowly, it would have taken an hour or more to load.
Hello everyone,
Roland from Unsloth here.
I’ve recently acquired a dgx spark machine and i will be releasing a dgx spark compatible container (including studio) in the near future.
That will be awesome rolandtannous.
Thanks!
Hi Roland.
Working on LLMs, on vLLM inside Dockers, and making sure to have the proper stack to run like a race engine its exiting, but… most of the time one wants stability and functionality, and Unsloth bring this on another level. MUST ADMIT i have been only using unsloth for inference, training and tasting models flavors the last 4 days.
Thanks for the GB10 updates,!!
Hi Roland,
How’s the progress on DGX Spark support? I’m really excited and looking forward to its release!