Selfhosted

hideshow 9 child comments

parent

[–] atzanteol@sh.itjust.works 1 point 4 months ago (8 children)

I'll check that out - speed isn't my biggest issue so much as coding performance... The qwen 3.5 model I was using can write code, but it's... Meh? Like sometimes it doesn't even compile.

I did try tweaking llama.cpp to do some cpu offloading and it does seem to allow for much larger contexts at a modest performance loss. I'll check out larger models.

hideshow 8 child comments

parent

[–] brucethemoose@lemmy.world 2 points 4 months ago* (3 children)

CPU offloading is too slow unless you use a hybrid MoE model, with the --n-cpu-moe parameter, specifically.

This only offloads "sparse" parts of the model to the CPU, which take up a lot of RAM but are very compute-lite to run. In practice, thats most of the size of modern MoE LLMs.

hideshow 3 child comments

parent

[–] robber@lemmy.ml 1 point 4 months ago (2 children)

Since implementation of the --fit parameter and its relatives, and --fit on becoming the default, llama.cpp intelligently decides what to offload. For me, it made --n-cpu-moe obsolete.

hideshow 2 child comments

parent

[–] brucethemoose@lemmy.world 2 points 4 months ago* (1 child)

Mostly, yeah.

Sometimes it’s better to “cut it close,” with (for instance) a 27B model that’s nearly OOMing your VRAM fully offloaded, but you know will be fine in regular use without too many programs open.

In my case, with MiMo 2.5, it fills both my CPU and GPU RAM rather completely, so it’s best to set a static value so I don’t swap CPU RAM, and don’t OOM on the GPU either.

parent

hideshow 1 child comment

[–] robber@lemmy.ml 1 point 4 months ago

You can control how much context should be fitted with --fit-ctx and how much space the algorithm should leave unallocated (even on a per-GPU basis) with --fit-target.

parent

[–] theterrasque@infosec.pub 1 point 4 months ago (3 children)

Try qwen3.6-35b-a3b with a lightweight harness like pi.dev

Having it be able to run commands and try to compile or run the code and see the output helps especially on the "doesn't compile" part of things

hideshow 3 child comments

parent

[–] atzanteol@sh.itjust.works 1 point 4 months ago* (last edited 4 months ago) (2 children)

Yeah - I've been playing around more with the Qwen3-Coder-30B-A3B-Instruct MoE model and it's still quite... Meh. I've been using llama.cpp and I've tried a bunch of tuning. It works and performs well enough (15t/s) but the output is just garbage. I can do some simple coding but I'm finding I'm fighting with it more than if I just wrote the code myself. Maybe I just have standards that are too high. Claude Opus 3.7 is just in an entirely different league...

hideshow 2 child comments

parent

[–] theterrasque@infosec.pub 1 point 4 months ago (1 child)

When you run it, do you use unsloth's recommended settings for coding?

https://unsloth.ai/docs/models/qwen3.6

Also have preserve thinking on, it helps it stay consistent in multi turn work.

Which model version you're using can also affect results, usually unsloth's ones are good.

With all that said, it's of course a small model so it's not a super coder. The 27b is better (I'd guess 25-35% better), but of course still a small model so..

So it'll maybe not be good enough still, but should give it the chance to let it do the best it can :)

parent

hideshow 1 child comment

[–] atzanteol@sh.itjust.works 1 point 4 months ago

So - I setup that model according to the docs and gave it this prompt:

Write me a highly optimized n-queens solver in go. It should take advantage of parallelism (what little there is) and output only the solution and how long it took.

After 10 minutes it gave me code that didn't compile.

It took another 3 mins to fix the compile error and the output is not correct.

As I said - LLMs on 8Gig VRAM just aren't worth it.