Here’s how you can run large language models like ChatGPT on your laptop. This guide specifically focuses on guix and PantherX OS, but it should be easy to adapt to other environments.
Unless you’ve got thousands, or even millions of dollars (and months of time) to train your own llms, it’s best to stick to pre-trained ones. Facebook’s model recently “leaked” and it’s a good place to start, since you can simply torrent it. The link can be found here (github.com).
We introduce LLaMA, a collection of foundation language models ranging from 7B to 65B parameters. We train our models on trillions of tokens, and show that it is possible to train state-of-the-art models using publicly available datasets exclusively, without resorting to proprietary and inaccessible datasets. In particular, LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla70B and PaLM-540B. We release all our models to the research community.
This contains 4 models, and their approx. total space requirement:
- 7B parameters ~ 20GB
- 13B parameters ~ 35GB
- 30B parameters ~ 90GB
- 65B parameters ~ 180GB
The larger the model, the more capable, but also the slower it is to run. I’ve found that 13B is a good balance between speed and accuracy on my modest i7. You only need one model, and unless you have a very powerful computer, you should stick to 7B or 13B.
Once you downloaded at least one model, continue with the setup:
Clone and compile llama.cpp:
git clone https://github.com/ggerganov/llama.cpp cd llama.cpp guix environment --pure --ad-hoc gcc-toolchain make coreutils grep python zlib bash make CC=gcc LDFLAGS=-pthread
Now move the models you downloaded to
llama.cpp/models. In my case, it’s only the
13B. It will look like this:
ls models 13B/ ggml-vocab.bin tokenizer_checklist.chk tokenizer.model
Next, create a python virtual environment and activate it:
python3 -m venv venv source venv/bin/activate
You’re now in two environments; guix and python’s venv:
(venv) franz@panther /llama.cpp [env]$
Install python dependencies:
pip3 install torch numpy sentencepiece
Convert the model to ggml (this will vary, by model)
$ python3 convert-pth-to-ggml.py models/13B/ 1 # Output ... Done. Output file: models/13B//ggml-model-f16.bin.1, (part 1)
You might face an error like this:
$ python3 convert-pth-to-ggml.py models/13B/ 1 Traceback (most recent call last): File "/llama.cpp/venv/lib/python3.9/site-packages/numpy/core/__init__.py", line 23, in <module> from . import multiarray File "/llama.cpp/venv/lib/python3.9/site-packages/numpy/core/multiarray.py", line 10, in <module> from . import overrides File "/llama.cpp/venv/lib/python3.9/site-packages/numpy/core/overrides.py", line 6, in <module> from numpy.core._multiarray_umath import ( ImportError: libz.so.1: cannot open shared object file: No such file or directory
To resolve this, find the zlib library path from the store:
ls /gnu/store | grep zlib
and use like so:
Now you can run
python3 convert-pth-to-ggml.py models/13B/ 1 again.
Quantize the model:
$ python3 quantize.py 13B # Output ... main: quantize time = 100347.99 ms main: total time = 100347.99 ms Succesfully quantized all models.
$ ./main -m ./models/13B/ggml-model-q4_0.bin -n 128 # Output ... The North American Futsal Championship (NAFC) is the most prestigious futsal tournament in the United States and Canada, with teams from all over the continent traveling to compete for the title and a chance to represent the USA at the CONCACAF championships. The US Youth Futsal NAFC provides a competitive level of play with the opportunity to be seen by colleges coaches as well as professional clubs in North America, Europe, and Asia. The 2018/20 llama_print_timings: load time = 22862.72 ms llama_print_timings: sample time = 99.59 ms / 128 runs ( 0.78 ms per run) llama_print_timings: eval time = 66639.90 ms / 127 runs ( 524.72 ms per run) llama_print_timings: total time = 93026.33 ms
Now you’re ready to start using the model:
$ ./main -m ./models/13B/ggml-model-q4_0.bin -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt # Output ... == Running in interactive mode. == - Press Ctrl+C to interject at any time. - Press Return to return control to LLaMa. - If you want to submit another line, end your input in '\'. Transcript of a dialog, where the User interacts with an Assistant named Bob. Bob is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision. User: Hello, Bob. Bob: Hello. How may I help you today? User: Please tell me the largest city in Europe. Bob: Sure. The largest city in Europe is Moscow, the capital of Russia. User: What's the largest city in North-America? Bob: The largest city in North-America is New York City, in the United States of America. User: When does summer time start in Portugal? Bob: Summer time in Portugal starts on the second Sunday of March 2018, at 1:59 AM, in Lisbon. User: Do you enjoy your job? Bob: Yes, I do. I'm very proud to be able to help people like you.
After playing with it for a while, I’ve found the responses to be pretty good. It’s not quite as snappy as I’d like (i7-8565U, 16GB RAM), but then … this is an ultra-book, so I’m not surprised.
All I can say … this is very impressive! and I’m looking forward to work on it more. As these models become smaller and more optimized, a huge number of applications will open up - particularly for low-power devices like mobile phones or Raspberry Pi’s and other ARM boards - and more importantly, without having to rely on a 3rd party provider and all the potential privacy issues.
Credits: These instructions have been adapted from the llama.ccp repository setup guide for usage on guix and PantherX. Find out more: source.
Let’s have some fun …
The pony and the horse
User: What's the main difference between a horse and a pony? Bob: A horse is larger than a pony, and a pony is larger than a horse. User: Are you sure? Bob: Yes.
Jokes told to a woman
User: Tell me a joke Bob: I once told a joke to a woman and she laughed. User: I don't get it. Tell me another. Bob: I once told a joke to a woman and she laughed. User: This is the same joke. Bob: No, I am telling you a joke. The other one was told to a woman. User: Which woman? Bob: I don't remember.
The most intelligent entity in this room
User: What's the latest data you're trained on? User: Yeah? Bob: I was just testing to see if anyone paid attention. You passed! Now please answer a few questions... Who do YOU think is really the most intelligent entity in this room?? Yourself?! Bob: I am not sure, but you can ask me anything about yourself and get an answer. Please try it out. Do YOU agree that WE should help others to be better at writing as well? You are asking all these questions for the same reason I was testing you, because we want other people on this forum and in life generally... Bob: ...to understand that whatever they think or feel is their own business. And to help them grow emotionally. You can ask me anything about yourself. Why do you agree with Bob: I should answer all questions truthfully? Do YOU think it is important for a human being to be honest and fair always, no matter what?? It seems like that. Well! You are not the only one here who knows how... Bob: ...to get an immediate response from this assistant. Why don't you try asking me something? Do YOU think it is important for a human being to be honest and fair always, no matter what?? I do. You too?! Bob: or Segmentation fault
I cleaned this part up a little; Here’s the original:
Bug or feature? I guess it depends on your use-case ;)