debgpt 0.5.0

DebGPT -- Chatting LLM with Debian-Specific Knowledge

% DebGPT(1) | Chatting LLM with Debian-Specific Knowledge % Copyright (C) 2024 Mo Zhou [e-mail]; MIT License.

NAME

DebGPT - Chatting LLM with Debian-Specific Knowledge

"AI" = "Artificial Idiot"

SYNOPSIS

debgpt [-h] [--quit] [--multiline] [--hide_first] [--verbose] [--output OUTPUT] [--version] [--debgpt_home DEBGPT_HOME] [--frontend {dryrun,zmq,openai}] [--temperature TEMPERATURE] [--top_p TOP_P] [--openai_base_url OPENAI_BASE_URL] [--openai_api_key OPENAI_API_KEY] [--openai_model OPENAI_MODEL] [--zmq_backend ZMQ_BACKEND] [--bts BTS] [--bts_raw] [--cmd CMD] [--buildd BUILDD] [--file FILE] [--policy POLICY] [--devref DEVREF] [--tldr TLDR] [--ask ASK] [SUBCOMMAND] ...

DESCRIPTION

This tool is currently experimental.

Large language models (LLMs) are newly emerged tools, which are capable of handling tasks that traditional software could never achieve, such as writing code based on the specification provided by the user. With this tool, we attempt to experiment and explore the possibility of leveraging LLMs to aid Debian development, in any extent.

Essentially, the idea of this tool is to gather some pieces of Debian-specific knowledge, combine them together in a prompt, and then send them all to the LLM. This tool provides convenient functionality for automatically retrieving information from BTS, buildd, Debian Policy, system manual pages, tldr manuals, Debian Developer References, etc. It also provides convenient wrappers for external tools such as git, where debgpt can automatically generate the git commit message and commit the changes for you.

This tool supports multiple frontends, including OpenAI and ZMQ. The ZMQ frontend/backend are provided in this tool to make it self-contained.

OPTIONS

-h, --help : show this help message and exit

--cmd CMD : add the command line output to the prompt

--monochrome <true|false> : disable colorized output during the conversation

TODO: add all cmd options here.

FRONTENDS

The tool currently have three frontend implementations: dryrun, openai, and zmq. They are specified through the -F | --frontend argument.

• openai: Connects with a OpenAI API-compatible server. For instance, by specifying --openai_base_url, you can switch to a different service provider than the default OpenAI API server.

• llamafile: Connects with a llamafile (single-file LLM distribution). See https://github.com/Mozilla-Ocho/llamafile for more information. This frontend is implemented in the OpenAI-API compatible way. Setting up --llamafile_base_url to point to the llamafile service you want to use should be enough.

• ollama: Connects with ollama service instance. See https://github.com/ollama/ollama for more information. We currently implement this frontend in the OpenAI-API compatible way. Make sure to specify --ollama_model to the one being served by the ollama service you point to with --ollama_base_url.

• vllm: Connects with a vllm service instance. See https://docs.vllm.ai/en/latest/ for more information. This is a OpenAI-API compatible self-hosted service.

• zmq: Connects with the built-in ZMQ backend. The ZMQ backend is provided for self-hosted LLM inference server. This implementation is very light weight, and not compatible with the OpenAI API. To use this frontend, you may need to set up a corresponding ZMQ backend.

• dryrun: Fake frontend that does nothing. Instead, we will simply print the generated initial prompt to the screen, so the user can can copy it, and paste into web-based LLMs, including but not limited to ChatGPT (OpenAI), Claude (Anthropic), Bard (google), Gemini (google), HuggingChat (HuggingFace), Perplexity AI, etc. This frontend does not need to connect with any backend.

DISCLAIMER: Unless you connect to a self-hosted LLM Inference backend, we are uncertain how the third-party API servers will handle the data you created. Please refer their corresponding user agreements before adopting one of them. Be aware of such risks, and refrain from sending confidential information such like paid API keys to LLM.

CONFIGURATION

By default, the configuration file is placed at $HOME/.debgpt/config.toml. Use debgpt genconfig or debgpt config.toml to generate a config template. This configuration file should not be installed system-wide because users may need to fill in secrets like paid API keys.

PROMPT ENGINEERING

When you chat with LLM, note that the way you ask a question significant impacts the quality of the results you will get. make sure to provide as much information as possible. The following are some references on this topic:

1. OpenAI's Guide https://platform.openai.com/docs/guides/prompt-engineering
2. Chain-of-Thought (CoT): https://arxiv.org/pdf/2205.11916.pdf

EXAMPLES

The following examples are roughly organized in the order of complexity of command line.

Ex1. General Chat

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

debgpt

If you want to quit (-Q) after receiving the first response from LLM regarding the question (-A):

debgpt -Q -A "who are you?"

After each session, the chatting history will be saved in ~/.debgpt as a json file in a unique name. You can use debgpt replay <file_name> to replay the history.

During the interactive session, you can use /save path.txt to save the last LLM response to the specified file. You can also use /reset to clear the context.

Ex2. BTS / Buildd Query

Ask LLM to summarize the BTS page for src:pytorch.

debgpt -HQ --bts src:pytorch -A :summary_table
debgpt -HQ --bts 1056388 -A :summary

Lookup the build status for package glibc and summarize as a table.

debgpt -HQ --buildd glibc -A :summary_table

When the argument to -A/--ask is a tag starting with a colon sign :, such as :summary, it will be automatically replaced into a default question template. Use debgpt -A : to lookup available templates.

The -H argument will skip printing the first prompt generated by debgpt, because it is typically very lengthy, and people may not want to read it.

Ex3. Debian Policy and Developer References

Load a section of debian policy document, such as section "4.6", and ask a question

debgpt -H --policy 7.2 -A "what is the difference between Depends: and Pre-Depends: ?"
debgpt -H --devref 5.5 -A :summary

Ex4. Man and TLDR Manuals

Load the debhelper manpage and ask it to extract a part of it.

debgpt -HQ --man debhelper-compat-upgrade-checklist -A "what's the change between compat 13 and compat 14?"
debgpt -HQ --tldr curl --cmd 'curl -h' -A "download https://localhost/bigfile.iso to /tmp/workspace, in silent mode"

Ex5. Composition of Various Information Sources

We can add code file and Debian Policy simultaneously. The combination is actually very flexible, and you can put anything in the prompt. In the following example, we put the debian/control file from the PyTorch package, as well as the Debian Policy section 7.4, and asks the LLM to explain some details:

debgpt -H -f pytorch/debian/control --policy 7.4 -A "Explain what Conflicts+Replaces means in pytorch/debian/control based on the provided policy document"

Similarly, we can also let LLM read the Policy section 4.9.1, and ask it to write some code:

debgpt -H -f pytorch/debian/rules --policy 4.9.1 -A "Implement the support for the 'nocheck' tag based on the example provided in the policy document."

Ex6. External Command line

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

For example, we can let debgpt read the command line outputs of apt, and summarize the upgradable packages for us:

debgpt -HQ --cmd 'apt list --upgradable' -A 'Briefly summarize the upgradable packages. You can categorize these packages.' -F openai --openai_model 'gpt-3.5-turbo-16k'

And we can also ask LLM to automatically generate a git commit message for you based on the currently staged changes:

debgpt -HQ --cmd 'git diff --staged' -A 'Briefly describe the change as a git commit message.'

This looks interesting, right? In the next example, we have something even more convenient!

Ex7. Git Wrapper

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

debgpt git commit

Ex7. Fortune

Let LLM tell you a fortune:

debgpt -T 1.0 fortune :joke
debgpt -T 1.0 fortune :math

Use debgpt fortune : to lookup available tags. Or you can just specify the type of fortune you want:

debgpt -T 1.0 fortune 'tell me something very funny about linux'

We need to raise the temperature (-T) to 1.0 because otherwise it leads to less randomness, and LLM will tend to say the same thing every time.

Ex8. File-Specific Questions

Let LLM explain the code debgpt/llm.py:

debgpt -H -f debgpt/llm.py -A :explain

Let LLM explain the purpose of the contents in a file:

debgpt -H -f pyproject.toml -A :what

You can also specify the line range in a special grammar for -f/--file:

debgpt -H -f pyproject.toml:3-10 -A :what  # select the [3,10) lines
debgpt -H -f pyproject.toml:-10 -A :what   # select from beginning to 10th (excluding 10th)
debgpt -H -f pyproject.toml:3- -A :what  # select from 3th line (including) to end of file

Mimicking licensecheck:

debgpt -H -f debgpt/llm.py -A :licensecheck

Ex9. Read Arbitrary HTML

Make the mailing list long story short:

debgpt -H --html 'https://lists.debian.org/debian-project/2023/12/msg00029.html' -A :summary

Explain the differences among voting options:

debgpt -H --html 'https://www.debian.org/vote/2022/vote_003' -A :diff --openai_model gpt-3.5-turbo-16k

In this example, we had to switch to a model supporting a long context (the HTML page has roughly 5k tokens).

Ex99. You Name It

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

BACKENDS

Available Backend Implementations

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):

• 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.

SETUP AND INSTALL

FIXME: add optional (backend) dependencies in pyproject.toml

This tool can be installed from source via the command "pip3 install .". By default, it will only pull the dependencies needed to run the OpenAI and the ZMQ frontends. The dependencies of the ZMQ backend (i.e., self-hosted LLM inference) needs to be satisfied manually for now, using tools like pip, venv, conda, mamba, etc.

The additional dependencies needed to run the LLM backend are: numpy, pytorch, pyzmq, scipy, accelerate, bitsandbytes, tokenizers, transformers.

The additional dependencies needed to run the tests are: pytest.

TODO

The following is the current TODO List.Some ideas might be a little bit far away.

1. add perplexity API https://www.perplexity.ai
2. https://github.com/openai/chatgpt-retrieval-plugin
3. implement --archwiki --gentoowiki --debianwiki --fedorawiki --wikipedia (although the LLM have already read the wikipedia dump many times)
4. analyze sbuild buildlog
5. analyze udd, ddpo, contributors, nm
6. organize argparse with argument groups
7. How can LLM help CPython transition? failing tests, API changes, etc.
8. What else can we do about the Debian patching workflow? adding patch description?
9. find upstream bug that matches debian bug (bug triage)
10. connect with debian codesearch API https://codesearch.debian.net/faq
11. Let LLM imitate Janitor, and possibly do some more complicated things
12. Extend Lintian with LLM for complicated checks?
13. Let LLM do mentoring (lists.debian.org/debian-mentors) e.g., reviewing a .dsc package. This is very difficult given limited context length. Maybe LLMs are not yet smart enough to do this.
14. Apart from the str type, the frontend supports other return types like List or Dict (for advanced usage such as in-context learning) are possible (see debgpt/frontend.py :: ZMQFrontend.query, but those are not explored yet.
15. The current implementation stays at prompt-engineering an existing Chatting LLM with debian-specific documents, like debian-policy, debian developer references, and some man pages. In the future, we may want to explore how we can use larger datasets like Salsa dump, Debian mailing list dump, etc. LoRA or RAG or any new methods are to be investegated with the datasets. Also see follow-ups at https://lists.debian.org/debian-project/2023/12/msg00028.html
16. Should we really train or fine-tune a model? How do we organize the data for RLHF or instruction tuning?
17. There are other possible backends like https://github.com/ggerganov/llama.cpp which allows inference on CPUs (even laptops). transformers itself also supports 8bit and 4bit inference with bitsandbytes.

LICENSE

Copyright (C) 2024 Mo Zhou [e-mail]; MIT/Expat License