The ChatAPI server module for providing API service and running instance of GPT-J-6B.

Files:

chatapi-gpt-j-6b.py

python -m pip install torch transformers

In the first running, the program will download the GPT-J-6B model from the Internet (about 25.5 GB). The model will be cached under ./.cache/.

And running with:

python3 chatapi-gpt-j-6b.py

When you see:

[Info <time>]: Loaded.
[Info <time>]: Running...

That represents the prepare work has been done.

Now, you can send a request by save a file into the requests folder (will be craete at python work dictory):

echo 'I like oranges. How about you?' > requests/1.request #Filename must ends with ".request".

After the program finished work, the request file will be delete and a new file with extname ".back" which including the result of your request will be create:

cat requests/1.back

You can send Ctrl+C key as normal.

Or you can create a file named quit.flag into flags folder:

touch flags/quit.flag

If the program detects the flag file (the detect work is after generate text), the program will quit.

Or you can create a file named status.flag into flags folder:

touch flags/status.flag

And you will get the status info:

cat flags/status.back

In fact, the Status value will always be running ( because the detect work is after generate text**).**

@misc{gpt-j,
  author = {Wang, Ben and Komatsuzaki, Aran},
  title = {{GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model}},
  howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
  year = 2021,
  month = May
}

View this repository on GitHub