vision-language models on consumer gear: worth it?

[marbletrack]

Wanted to compare notes with people doing similar things.

My current setup is PyTorch with mini PC, and I keep running into tokens/sec.

If you've tested something similar, what did you keep and what did you stop bothering with?

My use case is pretty ordinary: local chat, some document work, and occasional batch jobs.

[turboarch95]

Thanks, that helps. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[cachelab]

Thanks, that helps. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of t

That matches what I was seeing. The pattern around prefix caching has been pretty similar for me.

[audiopoint18]

I'd approach it in a boring order. If you're on PyTorch, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.

[quietarch]

following this because I'm in the same boat.

[turboarch95]

Thanks, that helps. That matches what I saw on A6000. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[cachelab]

Interesting, I hadn't tried that. I ended up in about the same place using Ollama.

[quietarch]

Interesting, I hadn't tried that. I don't think your conclusion is far off.

[thetanest]

Ran a quick check on my box. I couldn't reproduce the slowdown until I pushed context much higher.