Please provide 4 bit, Thank you.
I also would like to have Q6_K or Q5_K_M as a great balance between size and performance (Q4 can be a bit rough around the edges at this model size)
Actually you can do it yourself (although it's not recommended for already quantized models) with llama.cpp. Just build llama.cpp from source and then run:
(llama.cpp)$ ./quantize --allow-requantize Magicoder-S-DS-6.7B_q8_0.gguf <output_requantized_model>.gguf q4_k_m
any 4 bit good quants yet?
I just uploaded a q3, q4 and q5 here https://huggingface.co./matthoffner/Magicoder-S-DS-6.7B-GGUF/tree/main
I just uploaded a q3, q4 and q5 here https://huggingface.co./matthoffner/Magicoder-S-DS-6.7B-GGUF/tree/main
Hey thanks, i ended up( merging 417884e regex_gpt2_preprocess pr and it works) , wonder if this will produce any better results or otherwise.
It seems to work ok so far, I set up a space with it here https://huggingface.co./spaces/matthoffner/ggml-coding-llm