Remix.run Logo
monster_truck a day ago

I really like Qwen, even the Q2KP quants of 3.6 27B have genuinely impressive local performance on a 24GB card. It has been good enough that I am happily giving them $60 right now to try this instead of waiting to try a slightly lesser version locally.

Was there ever an explanation for why we never got the weights of 3.7? I would like sourced quotes and not weird/cringe accusative speculation about distillation, or your take on The Big D.

jared0x90 a day ago | parent [-]

do you mind sharing your settings? i just picked up an r9700 to start playing with local qwen3.6 27b and your setup sounds promising and efficient on 24gb.

monster_truck 16 hours ago | parent | next [-]

Which settings exactly do you want?

As far as the model settings go I just follow what's on the card.

I'm using HauhauCS's models, they seem to do a slightly better job with their "P" quants. Especially wrt patching them to eliminate the "doom loops" that will time out the GPU (esp if you have not already given it the extra power budget, set fans to max, and lowered your max clock by about 8% to save yourself a crash/reboot).

Without getting into the weeds, unsloth covers a lot more ground so, when they're good they're great, but I've also had the most problems with them. Basically, don't be afraid to shop around and fuck with sliders.

These days it should almost always be enough to open up LMStudio, set context to max, K/V quant to 16 or 8, and off you go. I'm using a 7900XTX, with 128GB of memory for the cases where things don't fit. The default settings should be fine otherwise.

I don't know anything about the r9700, seems neat. Looks like it might play nicer with the vulkan backend than rocm, and you might have to `set GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1`. Perf should be at least as good as what I'm getting (180+ tk/s prompt, >50tk/s response), which imho is just fast enough that I can read it as it reasons and responds.

You can probably get more aggressive with KV quant (any of the 4's) and bump up the batch sizes (4096/1024 vs 2048/512, etc), it's quite situational. These days there generally should not be any serious loss of accuracy from the former. Keeping the rig cool will likely matter more, so turn your music up to hide the fans and set your rig on top of the AC exhaust lol. Oh and I guess make sure ReBAR is enabled and working, adrenaline should let you know if it's not enabled.

kadoban 19 hours ago | parent | prev [-]

Different quants, but I've been having success with: https://unsloth.ai/docs/models/qwen3.6 I do llama.cpp