Comment by tyfon
2 hours ago
I'm using qwen 3.6 35B unsloth 4 bit with my 5950x (128 gb memory) and a 3060 12 gb gpu with a self made harness.
At 10k context I get about 40 tps generation and 500 tps prefill. At 100k context I get about 25 tps generation and 400 tps prefill.
It works, but I often use gpt or claude to make a detailed enumerated plan of what I want to do first, then have qwen follow it.
I'm not sure if it is economical or not, but I have solar on the roof so the power use is not really an issue and I already have the hardware.
The biggest benefit for me is that it's all done locally, and I know the harness is not uploading anything or sending telemetry to someone else.
> The biggest benefit for me is that it's all done locally, and I know the harness is not uploading anything or sending telemetry to someone else.
Are there any articles you’d recommend for this?
I have Qwen running on an HP Z8. Very nice platform.
I have mine in a sandbox, due to privacy fears.
Your solution sounds more elegant.
Articles regarding my own harness or how I set up llama.cpp etc?
I really just iterated over the harness over and over for about two weeks with opencode until I was sort of satisfied (still lots to do there :).
For the llama.cpp I asked claude fable to optimize it for my hardware and iterated a few times. In the end I landed on the following: https://pastebin.com/2PpJFUC0