bartowski

/

DeepSeek-R1-Distill-Qwen-14B-GGUF

like 174

VLLM GGUF support is in beta so not sure how well it runs in general but I can try this one locally later to confirm

Thanks a lot and also could you please make sure and double check if your gguf repos on 14b and 7b have the licenses required to be used in the AIMO-2 kaggle competition, your 7B models have been super useful and thanks a lot for the repo. But during final submission if any model does'nt have the licenses, our participation would be invalidated.
Do you need any help from my side for you to check the 14b 4bit models? I could send you any code? I tried on many versions of vLLM. The same code I used works for the 8bit 14b model

Yes they both contain the license file

running it locally it's fully coherent, can you give an example prompt?

"You are a helpful and harmless assistant. You are Qwen developed by Alibaba. You should think step-by-step. Return final answer within \boxed{}, after taking modulo 1000",

This prompt is used in this notebook, when I applied on the 14b it did'nt work.
https://www.kaggle.com/code/konstantinboyko/aimo-2-qwen-7b-gguf-q8-0-example

This maybe be a bug from vLLM, the 7B were only working on 0.7.1 vllm, I asked the author of this notebook and he didn't experiment much with the 14b. I'm still trying to figure out how to get these 14b up and working

You can also refer the last/latest comments in this discussion.
https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/565339#3151193

Could you specify which vLLM version you used? And also could you perform a small little test..uninstall and re-install the vLLM to a specific folder and uses import sys sys.path.append instead of making the installation in the default env path of python...