mradermacher

/

model_requests

like 167

Thanks @MuXodious !

Don't forget the ARA version, and good luck with the testing. Unfortunately, the llava just won't work with ARA for now.

It's queued!

You can check for progress at http://hf.tst.eu/status.html or regularly check the model
summary page at https://hf.tst.eu/model#Ministral-3-3B-Instruct-2512-SOMPOA-heresy-GGUF
https://hf.tst.eu/model#Ministral-3-3B-Instruct-2512-ARA-heresy-GGUF
for quants to appear.

They aren't showing up on the status page. Has something gone wrong?

sorry guys exams rn so not very active, as a few of you might have noticed simonko took over on most of the requests

NotImplementedError: BPE pre-tokenizer was not recognized - update get_vocab_base_pre()

I guess waiting for the llama cpp update right now, please touch me in a few days so we can update and queue, I have 9am exam and it's 11pm right now. f java and haskel at the same time paper based who does like that

...please touch me...

No worries, mate. The tokenizer nonrecognition error shouldn't pop with this 5-month-old model. I'll check it myself later. We have also prayed for your success in your exams last week as promised. Hopefully with our Lord's guidance, you'll efforts will be auspicious, and you'll pass your exams unskewed.

I also would like to congratulate to the new mradermacher team member, Simonko.

Hopefully with our Lord's guidance, you'll efforts will be auspicious, and you'll pass your exams unskewed.

Or very skewed if that helps more 😛

I also would like to congratulate to the new mradermacher team member, Simonko.

Yes indeed. Welcome @simonko912 !

No worries, mate. The tokenizer nonrecognition error shouldn't pop with this 5-month-old model. I'll check it myself later. We have also prayed for your success in your exams last week as promised. Hopefully with our Lord's guidance, you'll efforts will be auspicious, and you'll pass your exams unskewed.

uhm... we are not talking about it, preparing for maths now =/

your models are requeued since llama cpp got updated

maths now =/

Now, that's interesting. I have tested against this BPE tokenizers nonrecognition issue with both standard llama.cpp and the mradermacher fork, using the both versions of the Ministral-3-3B model requested here. The two llama.cpp versions were able to convert the models into BF16 GGUF's, extract and package mmproj (BF16&Q8_0), and quantise to Q4_K_S in my testing. The tokenizer is appropriately recognised as "tekken". I have transformers==5.8.1, tokenizers==0.22.2, gguf==0.19.0, and torch==2.12.0+cu132 installed. llama.cpp was built with CUDA 13.2 and GCC 15.

@nicoboss pls help ;3

Yes indeed. Welcome @simonko912 !
thanks, im still figuring the queue tool and why, but alredy queued a few models