How can I create a 128K GGUF of a non-instruct model?

#1
by nitinrathi - opened

I am using Qwen2.5 Coder 7B for FIM and code completion in my editor. I am using repo-level completion, but the 32K context is not enough. How can I create a 128K GGUF of the non-instruct model?

how are you doing FIM right now. can you share tcode

I wrote an Emacs extension. I used the following template:

<|repo_name|>{repo_name}
<|file_sep|>{file_path1}
{file_content1}
<|file_sep|>{file_path2}
{file_content2}
<|file_sep|>{file_path3}
<|fim_prefix|>{code_pre}<|fim_suffix|>{code_suf}<|fim_middle|>{code_fim}<|endoftext|>

For inference I am using llama.cpp and the /completion endpoint.

I will open source the Emacs extension once it matures.

Sign up or log in to comment