🦙 🚀 Alpaca-LoRA as a Chatbot Service

auto summarization

TODO (also contribution requests):

Save/load button to save conversation
Other model supports (baize, flan series)
Better context management: other than auto-summarization, retrieving relevant information from the past conversation history based on the current conversation window
transformers' Streamer: transformers starts supporting streaming generation. It will help us to remove the monkey patched StreamModel
WebGPT like feature: by referencing webgpt-cli

🔗 Demo link: Batch Mode with 30B and Streaming Mode with 30B (running on a single A6000 and 3xA6000 instances respectively at jarvislabs.ai), and Hugging Face Space which runs 13B on A10.

The easiest way to run this project is to use Colab. Just open up the alpaca_lora_in_colab notebook in Colab (there is a button open in colab), and run every cell sequentially. With the standard GPU instance(T4), you can run 7B and 13B models. With the premium GPU instance(A100 40GB), you can even run 30B model! Screenshot👇🏼 Just note that the connection could be somewhat unstable, so I recommend you to use Colab for development purpose.

This repository demonstrates Alpaca-LoRA as a Chatbot service with Alpaca-LoRA and Gradio. It comes with the following features:

Mode

1. Batch Generation Mode: batch generation mode aggregates requests up to batch_size, and pass the prompts in the requests to the model. It waits the current requests are fully handled. For instance, with batch_size=4, if a user sends a request, that is under processing. While it is under processing, if other users are connected, up to 4 requests from the users are aggregated and processed as soon as the current one is done.

2. Streaming Mode: streaming mode handles multiple requests in a interleaving way with threads. For instance, if there are two users (A and B) are connected, A's request is handled, and then B's request is handled, and then A's request is handled again.... This is because of the nature of streaming mode which generates and yield tokens in one by one manner.

Context management

Alpaca-LoRA as a Chatbot Service manages context in two ways. First of all, it remembers(stores) every history of the conversations by default as in the following code snippet. context_string is set as "Below is a history of instructions that describe tasks, paired with an input that provides further context. Write a response that appropriately completes the request by remembering the conversation history." by default, but it could be set manually via the Context field on top of the screen.
- additionally, there is a Summarize button in the middle (you need to expand the component labeled as "Helper Buttons"). If you click this button, it automatically input "summarize our conversations so far in three sentences." as a prompt, and the resulting generated text will be inserted into the Context field. THen all the conversation history up to this point will be ignored. That means the conversation fresh restarts with the below code snippet except context_string will be filled up with the model generated text.
- NOTE: the only last 2,000 characters are kept, and this number can be configured in constants.py

f"""{context_string}

### Input: {input} # Surrounding information to AI

### Instruction: {prompt1} # First instruction/prompt given by user

### Response {response1} # First response on the first prompt by AI

### Instruction: {prompt2} # Second instruction/prompt given by user

### Response: {response2} # Second response on the first prompt by AI
....
"""

misc.

There is a continue button in the middle of screen. What it does is to simply send "continue." prompt to the model. This is useful if you get incomplete previous response from the model. With the "continue.", the model tries to complete the response. Also, since this is a continuation of the response, the "continue." prompt will be hidden to make chatting history more natural.

Currently supported LoRA checkpoints

tloen/alpaca-lora-7b: the original 7B Alpaca-LoRA checkpoint by tloen
chansung/alpaca-lora-13b: the 13B Alpaca-LoRA checkpoint by myself(chansung) with the same script to tune the original 7B model
chansung/koalpaca-lora-13b: the 13B Alpaca-LoRA checkpoint by myself(chansung) with the Korean dataset created by KoAlpaca project by Beomi. It works for English(user) to Korean(AI) conversations.
chansung/alpaca-lora-30b: the 30B Alpaca-LoRA checkpoint by myself(chansung) with the same script to tune the original 7B model

Instructions

Prerequisites

Note that the code only works Python >= 3.9

$ conda create -n alpaca-serve python=3.9
$ conda activate alpaca-serve

Install dependencies

$ cd Alpaca-LoRA-Serve
$ pip install -r requirements.txt

Run Gradio application

$ BASE_URL=/path/to/trained/alpaca

$ python app.py --model_name_or_path $BASE_URL --peft --load_in_8bit

the following flags are supported

usage: app.py [-h] [--base_url BASE_URL] [--ft_ckpt_url FT_CKPT_URL] [--port PORT] [--batch_size BATCH_SIZE] [--api_open] [--share]
              [--gen_config_path GEN_CONFIG_PATH] [--multi_gpu]

Gradio Application for Alpaca-LoRA as a chatbot service

options:
  -h, --help            show this help message and exit
  --base_url BASE_URL   Hugging Face Hub URL
  --ft_ckpt_url FT_CKPT_URL
                        Hugging Face Hub URL
  --port PORT           PORT number where the app is served
  --batch_size BATCH_SIZE
                        Number of requests to handle at the same time
  --api_open            Open as API
  --share               Create and share temporary endpoint (useful in Colab env)
  --gen_config_path GEN_CONFIG_PATH
                        path to GenerationConfig file used in batch mode

Design figure

Acknowledgements

I am thankful to Jarvislabs.ai who generously provided free GPU resources to experiment with Alpaca-LoRA deployment and share it to communities to try out.

yukw777 / Alpaca-LoRA-Serve