1. Chatting with LLM and CLI Behavior¶

When no arguments are given, debgpt leads you into a general terminal chatting client with LLM backends. Use debgpt -h to see detailed options.

debgpt
    

During the interactive chatting mode, you may press / and see a list of available escaped commands that will not be seen as LLM prompt.

•

/save <path.txt>: save the last LLM response to the specified file.

•

/reset: clear the context. So you can start a new conversation without quiting.

•

/quit: quit the chatting mode. You can press Ctrl-D to quit as well.

The first user prompt can be provided through argument (--ask|-A|-a):

debgpt -A "Who are you? And what can LLM do?"
    

By specifying the --quit|-Q option, the program will quit after receiving the first response from LLM. For instance, we can let it mimic fortune with temperature 1.0 (--temperature|-T 1.0) for higher randomness:

debgpt -T 1.0 -QA 'Greet with me, and tell me a joke.'
    

After each session, the chatting history will be saved in ~/.debgpt as a json file in a unique name. The command debgpt replay can replay the last session if you forgot what LLM has replied to you.

The program can write the last LLM response to a file through -o <file>, and read question from stdin:

debgpt -Qa 'write a hello world in rakudo for me' -o hello.raku
debgpt -HQ stdin < question.txt | tee result.txt
    

After gettting familiarized with the fundamental usage and its CLI behavior, we can directly move on to the most important feature of this tool, namely the special prompt reader – MapReduce.

2. Context Readers for Additional Information¶

Context Reader is a function that reads the plain text contents from the specified resource, and wrap them as a part of a prompt for the LLM. The context readers can be arbitrarily combined together or specified multiple times through the unified argument --file|-f. It can read from a file, a PDF file, a directory, a URL, a Debian Policy section, a Debian Developer Reference section, a Debian BTS page, a Debian build status page (buildd), a Google search result, etc. The specification syntax of the unified reader --file|-f can be found using the command debgpt -f : or debgpt -f '?'.

Some examples are shown below:

# read a plain text file and ask a question
debgpt -Hf README.md -a 'very briefly teach me how to use this DebGPT.'
debgpt -Hf debgpt/policy.py -A 'explain this file'
debgpt -Hf my-resume.pdf -a 'Does this person have any foss-related experience?'
# read a directory
debgpt -Hf debian/ -a 'how is this package built? How many binary packages will be produced?'
# read a URL
debgpt -Hf 'https://www.debian.org/vote/2022/vote_003' -A 'Please explain the differences among the above choices.'
# reading debian bug tracking system
debgpt -Hf bts:src:pytorch -A 'Please summarize the above information. Make a table to organize it.'
debgpt -Hf bts:1056388 -A 'Please summarize the above information.'
# reading debian buildd status
debgpt -Hf buildd:glibc -A 'Please summarize the above information. Make a table to organize it.'
# piping from a command line
debgpt -Hf cmd:'apt list --upgradable' -A 'Briefly summarize the upgradable packages. You can categorize these packages.'
debgpt -Hf cmd:'git diff --staged' -A 'Briefly describe the change as a git commit message.'
# read manual page or tldr
debgpt -Hf man:debhelper-compat-upgrade-checklist -A "what's the change between compat 13 and compat 14?"
debgpt -H -f tldr:curl -f cmd:'curl -h' -A "download https://localhost/bigfile.iso to /tmp/workspace, in silent mode"
# read a Debian Policy section or Developer Reference section
debgpt -Hf policy:7.2 -A "what is the difference between Depends: and Pre-Depends: ?"
debgpt -Hf devref:5.5 -A 'Please summarize the above information.'
# when section is not specified, it will read the whole document. This may exceed the LLM context size limit.
debgpt -Hf policy: -A 'what is the latest changes in this policy?'
# more examples
debgpt -Hf pytorch/debian/control -f policy:7.4 -A "Explain what Conflicts+Replaces means in pytorch/debian/control based on the provided policy document"
debgpt -Hf pytorch/debian/rules -f policy:4.9.1 -A "Implement the support for the 'nocheck' tag based on the example provided in the policy document."
    

3. Inplace Editing of a File¶

The argument [–inplace|-i] is for in-place editing of a file. It is a read-write reader that does the same as --file|-f (read-only) does, but the inplace one will write the LLM response back to the file. We expect the user to use this feature for editing a file.

If specified, the edits (in UNIX diff format) will be printed to the screen. The --inplace|-i will mandate the --quit|-Q behavior, and will turn off markdown rendering.

The following example will ask LLM to edit the pyproject.toml file, adding pygments to its dependencies. This really works correctly.

debgpt -Hi pyproject.toml -a 'edit this file, adding pygments to its dependencies.'
    

If working in a Git repository, we can make things more automated: You may further append --inplace-git-add-commit to automatically add and commit the changes to the Git repository. If you want to review before commit, specify --inplace-git-p-add-commit|-I argument instead.

debgpt -Hi pyproject.toml -a 'edit this file, adding pygments to its dependencies.' --inplace-git-add-commit
    

The commit resulted by the above example can be seen at https://salsa.debian.org/deeplearning-team/debgpt/-/commit/968d7ab31cb3541f6733eb34bdf6cf13b6552b7d this link. Recent LLMs are strong enough to easily and correctly add type annotations and doc strings in DebGPT’s python codebase, see example https://salsa.debian.org/deeplearning-team/debgpt/-/commit/4735b38141eafd6aa9b0863fc73296aa41562aed here.

4. Vector Retriever for Most Relevant Information¶

5. MapReduce for Any Length Context¶

The “MapReduce” feature is the choice if you want the LLM to read bulk documentations.

Generally, LLMs have a limited context length. If you want to ask a question regarding a very long context, you can split the context into multiple parts, and extract the relevant information from each part. Then, you can ask the LLM to answer the question based on the extracted information.

The implementation of this is fairly simple: split the gathered information texts until the pre-defined maximum chunk size is satisfied, ask the LLM to extract relevant information from each chunk, and then repeatedly merge the extracted information through LLM summarization, untill there is only one chunk left. As a result, this functionality can be very quota-consuming if you are going to deal with long texts. Please keep an eye on your bill when you try this on a paied API service.

This functionality is implemented as the --mapreduce|-x argument. The user has to specify the --ask|-A|-a argument to tell LLM what kind of question we want to ask so it can extract the right information. It will summarize if the --ask|-A|-a argument is missing.

The key difference between the MapReduce and Vector Retriever is that MapReduce will really make the language model read all information you passed to it, while vector retriever will only make language model read the most relevant several pieces of information stored in the database.

Some usage examples of MapReduce are as follows:

•

Load a file and ask a question

debgpt -Hx resume.pdf -A 'Does this person know AI? To what extent?'
    

•

Load a directory and ask a question

debgpt -Hx . -a 'which file implemented mapreduce? how does it work?'
debgpt -Hx . -a 'teach me how to use this software. Is there any hidden functionality that is not written in its readme?'
debgpt -Hx ./debian -A 'how is this package built? how many binary packages will be produced?'
    

•

Load a URL and ask a question

debgpt -Hx 'https://www.debian.org/doc/debian-policy/policy.txt' -A 'what is the purpose of the archive?'
    

•

Load the whole Debian Policy document (plain text) and ask a question

debgpt -Hx policy:all -a "what is the latest changes in this policy?"
debgpt -Hx policy:all -A 'what package should enter contrib instead of main or non-free?'
    

•

Load the whole Debian Developer Reference document (plain text) and ask a question

debgpt -Hx devref:all -A 'How can I become a debian developer?'
debgpt -Hx devref:all -a 'how does general resolution work?'
    

•

If you don’t really bother to read policy: and devref:, or forgot which one is talking about the question in you mind, for instance:

debgpt -H -x policy:all -x devref:all -a 'which document (and which section) talk about Multi-Arch: ?'
    

•

Summarize the mailing list discussions within a month (MapReduce is more suitable than the retrieval (RAG) for this purpose):

debgpt -Hx ldo:debian-project/2024/10 -a 'write a news report based on the provided information. Cover as many topics as possible. You may expand a little bit on important matter. You must include links for every topic to the report.' --no-render
    

•

Load the latest sbuild log file and ask a question

debgpt -Hx sbuild: -A 'why does the build fail? do you have any suggestion?'
    

•

Google search: -x google: will use your prompt as the search query, and answer your question after reading the search results

debgpt -Hx google: -a 'how to start python programming?'
    

•

Google search: -x google:<search_query> gives more control over the search query. Here we let LLM answer the question provided by -a based on the search results of “debian packaging”.

debgpt -Hx google:'debian packaging' -a 'how to learn debian packaging?'
    

The -H argument will skip printing the first prompt generated by debgpt, because it is typically very lengthy, and only useful for debugging and development purpose. To further tweak the mapreduce behavior, you may want to check the --mapreduce_chunksize <int> and --mapreduce_parallelism <int> arguments.

6. Piping through Everywhere¶

Being able to pipe the inputs and outputs among different programs is one of the reasons why I love the UNIX philosophy.

The pipe mode is useful when you want to use debgpt in a shell script Try the follows on the Makefile in debgpt repo. Later we will introduce a in-place editing functionality which is more convenient than this one.

cat Makefile | debgpt -a 'delete the deprecated targets' pipe | tee tmp ; mv tmp Makefile; git diff
    

The pipe mode can be used for editing something in vim in-place.

# In vim debgpt/task.py, use 'V' mode to select the task_backend function, then 
:'<,'>!debgpt -a 'add type annotations and comments to this function' pipe
    

This looks interesting, right? debgpt has a git wrapper that automatically generates the git commit message for the staged contents and commit the message. Just try debgpt git commit --amend to see how it works. This will also be mentioned in the subcommands section.

7. DebGPT Subcommands¶

Git subcommand.

Let LLM automatically generate the git commit message, and call git to commit it:

debgpt git commit --amend
    

If you don’t even want to git commit --amend the commited message, just remove --amend from it.

8. Prompt Engineering¶

As you may have seen, the biggest variation in LLM usage happens in the context including how you provide the context readers, and how you ask the question through --ask|-A|-a. By adjusting the way you provide those information and ask the question, you can get significantly different results. To properly make LLM work for you, you may need to go through some basic prompt engineering methods.

The following are some references on this topic:

1.

OpenAI’s Guide https://platform.openai.com/docs/guides/prompt-engineering

Advanced usage of LLM such as Chain-of-Thought (CoT) will not be covered in this document. Please refer external resources for more information.

The usage of LLM is limited by our imaginations. I am glad to hear from you if you have more good ideas on how we can make LLMs useful for Debian development: https://salsa.debian.org/deeplearning-team/debgpt/-/issues

9. Debian Specific Usage Cases¶

•

Analysis of the sbuild buildlog

•

Analysis of the ratt buildlog directory

•

Evaluation on nm-templates. There is a reader spec design for this: -f nm:<question_id> for loading nm-template questions. It is just a convenient wrapper of multiple other readers. The following is a script for answering all questions from nm-templates automatically. The answers will be stored in plain text files.

# nm_assigned.txt
debgpt -f nm:nm_assigned -a 'pretend to be Mo Zhou <lumin@debian.org> and answer the question. Give concrete examples, and links as evidence supporting them are preferred.' -o nm-assigned-selfintro.txt
# nm_pp1.txt
for Q in PH0 PH1 PH2 PH3 PH4 PH5 PH6 PH7 PHa; do
debgpt -HQf nm:pp1.${Q} -a 'Be concise and answer in just several sentences.' -o nm-pp1-${Q}-brief.txt;
debgpt -HQf nm:pp1.${Q} -a 'Be precise and answer with details explained.' -o nm-pp1-${Q}-detail.txt;
done
# nm_pp1_extras.txt
for Q in PH0 PH8 PH9 PHb; do
debgpt -HQf nm:pp1e.${Q} -a 'Be concise and answer in just several sentences.' -o nm-pp1e-${Q}-brief.txt;
debgpt -HQf nm:pp1e.${Q} -a 'Be precise and answer with details explained.' -o nm-pp1e-${Q}-detail.txt;
done
# nm_pp2.txt
for Q in BT2 BT6 BT8; do
debgpt -HQf nm:pp2.${Q} -a 'Be concise and answer in just several sentences.' -o nm-pp2-${Q}-brief.txt;
debgpt -HQf nm:pp2.${Q} -a 'Be precise and answer with details explained.' -o nm-pp2-${Q}-detail.txt;
done
    

Ollama¶

Please refer to their official documentation for setting up. Note, you may need to adjust the num_ctx parameter in a Modelfile:

FROM qwen2.5
PARAMETER num_ctx 65536
    

Use ollama create <modelname> -f Modelfile to create a model with the new parameters. See upstream API documentation and Modelfile documentation for more details.

llamaFile¶

TODO: write this part of doc.

ZMQ (DebGPT Built-in)¶

This tool provides one backend implementation: zmq.

•

zmq: Only needed when you choose the ZMQ front end for self-hosted LLM inference server.

If you plan to use the openai or dryrun frontends, there is no specific hardware requirement. If you would like to self-host the LLM inference backend (ZMQ backend), powerful hardware is required.

LLM Selections

The concrete hardware requirement depends on the LLM you would like to use. A variety of open-access LLMs can be found here > https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard Generally, when trying to do prompt engineering only, the “instruction-tuned” LLMs and “RL-tuned” (RL is reinforcement learning) LLMs are recommended.

The pretrained (raw) LLMs are not quite useful in this case, as they have not yet gone through instruction tuning, nor reinforcement learning tuning procedure. These pretrained LLMs will more likely generate garbage and not follow your instructions, or simply repeat your instruction. We will only revisit the pretrained LLMs when we plan to start collecting data and fine-tune (e.g., LoRA) a model in the far future.

The following is a list of supported LLMs for self-hosting (this list will be updated when there are new state-of-the-art open-access LLMs available): • .RS 2

Mistral7B (Mistral-7B-Instruct-v0.2) (default)

This model requires roughly 15GB of disks space to download.

Mixtral8x7B (Mixtral-8x7B-Instruct-v0.1)

This model is larger yet more powerful than the default LLM. In exchange, it poses even higher hardware requirements. It takes roughly 60~100GB disk space (I forgot this number. Will check later).

Different LLMs will pose different hardware requirements. Please see the “Hardware Requirements” subsection below.

Hardware Requirements

By default, we recommend doing LLM inference in fp16 precision. If the VRAM (such as CUDA memory) is limited, you may also switch to even lower preicisions such as 8bit and 4bit. For pure CPU inference, we only support fp32 precision now.

Note, Multi-GPU inference is supported by the underlying transformers library. If you have multiple GPUs, this memory requirement is roughly divided by your number of GPUs.

Hardware requirements for the Mistral7B LLM:

•

Mistral7B + fp16 (cuda): 24GB+ VRAM preferred, but needs a 48GB GPU to run all the demos (some of them have a context as long as 8k). Example: Nvidia RTX A5000, Nvidia RTX 4090.

•

Mistral7B + 8bit (cuda): 12GB+ VRAM at minimum, but 24GB+ preferred so you can run all demos.

•

Mistral7B + 4bit (cuda): 6GB+ VRAM at minimum but 12GB+ preferred so you can run all demos. Example: Nvidia RTX 4070 (mobile) 8GB.

•

Mistral7B + fp32 (cpu): Requires 64GB+ of RAM, but a CPU is 100~400 times slower than a GPU for this workload and thus not recommended.

Hardware requirement for the Mixtral8x7B LLM:

•

Mixtral8x7B + fp16 (cuda): 90GB+ VRAM.

•

Mixtral8x7B + 8bit (cuda): 45GB+ VRAM.

•

Mixtral8x7B + 4bit (cuda): 23GB+ VRAM, but in order to make it work with long context such as 8k tokens, you still need 2x 48GB GPUs in 4bit precision.

See https://huggingface.co/blog/mixtral for more.

Usage of the ZMQ Backend

If you want to run the default LLM with different precisions:

debgpt backend --max_new_tokens=1024 --device cuda --precision fp16
debgpt backend --max_new_tokens=1024 --device cuda --precision bf16
debgpt backend --max_new_tokens=1024 --device cuda --precision 8bit
debgpt backend --max_new_tokens=1024 --device cuda --precision 4bit
    

The only supported precision on CPU is fp32 (for now). If you want to fall back to CPU computation (very slow):

debgpt backend --max_new_tokens=1024 --device cpu --precision fp32
    

If you want to run a different LLM, such as Mixtral8x7B than the default Mistral7B:

debgpt backend --max_new_tokens=1024 --device cuda --precision 4bit --llm Mixtral8x7B
    

The argument --max_new_tokens does not matter much and you can adjust it (it is the maximum length of each llm reply). You can adjust it as wish.