/FastChat

An open platform for training, serving, and evaluating large languages. Release repo for Vicuna and FastChat-T5.

Primary LanguagePythonApache License 2.0Apache-2.0

FastChat

| Demo | Arena | Discord | Twitter |

FastChat is an open platform for training, serving, and evaluating large language model based chatbots. The core features include:

  • The weights, training code, and evaluation code for state-of-the-art models (e.g., Vicuna, FastChat-T5).
  • A distributed multi-model serving system with Web UI and OpenAI-compatible RESTful APIs.

News

  • [2023/05] 🔥 We introduced Chatbot Arena for battles among LLMs. Check out the blog post and demo.
  • [2023/04] We released FastChat-T5 compatible with commercial usage. Check out the weights and demo.
  • [2023/03] We released Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90% ChatGPT Quality. Check out the blog post and demo.

Contents

Install

Method 1: With pip

pip3 install fschat

Method 2: From source

  1. Clone this repository and navigate to the FastChat folder
git clone https://github.com/lm-sys/FastChat.git
cd FastChat

If you are running on Mac:

brew install rust cmake
  1. Install Package
pip3 install --upgrade pip  # enable PEP 660 support
pip3 install -e .

Model Weights

Vicuna Weights

We release Vicuna weights as delta weights to comply with the LLaMA model license. You can add our delta to the original LLaMA weights to obtain the Vicuna weights. Instructions:

  1. Get the original LLaMA weights in the Hugging Face format by following the instructions here.
  2. Use the following scripts to get Vicuna weights by applying our delta. They will automatically download delta weights from our Hugging Face account.

NOTE: Weights v1.1 are only compatible with transformers>=4.28.0 and fschat >= 0.2.0. Please update your local packages accordingly. If you follow the above commands to do a fresh install, then you should get all the correct versions.

Vicuna-7B

This conversion command needs around 30 GB of CPU RAM. See the "Low CPU Memory Conversion" section below if you do not have enough memory. Replace /path/to/* with the real paths.

python3 -m fastchat.model.apply_delta \
    --base-model-path /path/to/llama-7b \
    --target-model-path /path/to/output/vicuna-7b \
    --delta-path lmsys/vicuna-7b-delta-v1.1

Vicuna-13B

This conversion command needs around 60 GB of CPU RAM. See the "Low CPU Memory Conversion" section below if you do not have enough memory. Replace /path/to/* with the real paths.

python3 -m fastchat.model.apply_delta \
    --base-model-path /path/to/llama-13b \
    --target-model-path /path/to/output/vicuna-13b \
    --delta-path lmsys/vicuna-13b-delta-v1.1

Old weights

See docs/vicuna_weights_version.md for all versions of weights and their differences.

Low CPU Memory Conversion

You can try these methods to reduce the CPU RAM requirement of weight conversion.

  1. Append --low-cpu-mem to the commands above, which will split large weight files into smaller ones and use the disk as temporary storage. This can keep the peak memory at less than 16GB.
  2. Create a large swap file and rely on the operating system to automatically utilize the disk as virtual memory.

FastChat-T5

Simply run the line below to start chatting. It will automatically download the weights from a Hugging Face repo.

python3 -m fastchat.serve.cli --model-path lmsys/fastchat-t5-3b-v1.0

Inference with Command Line Interface

(Experimental Feature: You can specify --style rich to enable rich text output and better text streaming quality for some non-ASCII content. This may not work properly on certain terminals.)

Supported Models

The following models are tested:

Help us add more.

Single GPU

The command below requires around 28GB of GPU memory for Vicuna-13B and 14GB of GPU memory for Vicuna-7B. See the "No Enough Memory" section below if you do not have enough memory. Replace /path/to/model/weights with the a local folder or a Hugging repo id.

python3 -m fastchat.serve.cli --model-path /path/to/model/weights

Multiple GPUs

You can use model parallelism to aggregate GPU memory from multiple GPUs on the same machine.

python3 -m fastchat.serve.cli --model-path /path/to/model/weights --num-gpus 2

CPU Only

This runs on the CPU only and does not require GPU. It requires around 60GB of CPU memory for Vicuna-13B and around 30GB of CPU memory for Vicuna-7B.

python3 -m fastchat.serve.cli --model-path /path/to/model/weights --device cpu

Metal Backend (Mac Computers with Apple Silicon or AMD GPUs)

Use --device mps to enable GPU acceleration on Mac computers (requires torch >= 2.0). Use --load-8bit to turn on 8-bit compression.

python3 -m fastchat.serve.cli --model-path /path/to/model/weights --device mps --load-8bit

Vicuna-7B can run on a 32GB M1 Macbook with 1 - 2 words / second.

Intel XPU (Intel Data Center and Arc A-Series GPUs)

Install the Intel Extension for PyTorch.

Set the OneAPI environment variables:

source /opt/intel/oneapi/setvars.sh

Use --device xpu to enable XPU/GPU acceleration.

python3 -m fastchat.serve.cli --model-path /path/to/model/weights --device xpu

Vicuna-7B can run on an Intel Arc A770 16GB.

No Enough Memory

If you do not have enough memory, you can enable 8-bit compression by adding --load-8bit to commands above. This can reduce memory usage by around half with slightly degraded model quality. It is compatible with the CPU, GPU, and Metal backend. Vicuna-13B with 8-bit compression can run on a single NVIDIA 3090/4080/T4/V100(16GB) GPU.

python3 -m fastchat.serve.cli --model-path /path/to/model/weights --load-8bit

In addition to that, you can add --cpu-offloading to commands above to offload weights that don't fit on your GPU onto the CPU memory. This requires 8-bit compression to be enabled and the bitsandbytes package to be installed, which is only available on linux operating systems.

More Platforms

  • MLC LLM, backed by TVM Unity compiler, deploys Vicuna natively on phones, consumer-class GPUs and web browsers via Vulkan, Metal, CUDA and WebGPU.

GPTQ 4bit Support

FastChat provides fastest GPTQ 4bit inference with GPTQ-for-LLaMa. See docs/gptq.md.

Serving with Web GUI

To serve using the web UI, you need three main components: web servers that interface with users, model workers that host one or more models, and a controller to coordinate the webserver and model workers. You can learn more about the architecture here.

Here are the commands to follow in your terminal:

Launch the controller

python3 -m fastchat.serve.controller

This controller manages the distributed workers.

Launch the model worker(s)

python3 -m fastchat.serve.model_worker --model-path /path/to/model/weights

Wait until the process finishes loading the model and you see "Uvicorn running on ...". The model worker will register itself to the controller .

To ensure that your model worker is connected to your controller properly, send a test message using the following command:

python3 -m fastchat.serve.test_message --model-name vicuna-7b

You will see a short output.

Launch the Gradio web server

python3 -m fastchat.serve.gradio_web_server

This is the user interface that users will interact with.

By following these steps, you will be able to serve your models using the web UI. You can open your browser and chat with a model now. If the models do not show up, try to reboot the gradio web server.

(Optional): Advanced Features

  • You can register multiple model workers to a single controller, which can be used for serving a single model with higher throughput or serving multiple models at the same time. When doing so, please allocate different GPUs and ports for different model workers.
# worker 0
CUDA_VISIBLE_DEVICES=0 python3 -m fastchat.serve.model_worker --model-path lmsys/fastchat-t5-3b-v1.0 --controller http://localhost:21001 --port 31000 --worker http://localhost:31000
# worker 1
CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.model_worker --model-path ~/model_weights/vicuna-7b/ --controller http://localhost:21001 --port 31001 --worker http://localhost:31001
  • You can also launch a multi-tab gradio server, which includes the Chatbot Arena tabs.
python3 -m fastchat.serve.gradio_web_server_multi
  • You can protect your webserver with Gradio's Authentication with a password file. The password file should contain one or more "user:password" pairs in this format: u1:p1,u2:p2,u3:p3
python3 -m fastchat.serve.gradio_web_server --gradio-auth-path login.txt

API

OpenAI-Compatible RESTful APIs & SDK

FastChat provides OpenAI-compatible APIs for its supported models, so you can use FastChat as a local drop-in replacement for OpenAI APIs. The FastChat server is compatible with both openai-python library and cURL commands. See docs/openai_api.md.

Hugging Face Generation APIs

See fastchat/serve/huggingface_api.py.

LangChain Integration

See docs/langchain_integration.

Evaluation

Our AI-enhanced evaluation pipeline is based on GPT-4. This section provides a high-level summary of the pipeline. For detailed instructions, please refer to the evaluation documentation.

Pipeline Steps

  1. Generate answers from different models: Use qa_baseline_gpt35.py for ChatGPT, or specify the model checkpoint and run get_model_answer.py for Vicuna and other models.

  2. Generate reviews with GPT-4: Use GPT-4 to generate reviews automatically. This step can also be performed manually if the GPT-4 API is not available to you.

  3. Generate visualization data: Run generate_webpage_data_from_table.py to generate data for a static website, which allows you to visualize the evaluation data.

  4. Visualize the data: Serve a static website under the webpage directory. You can use python3 -m http.server to serve the website locally.

Data Format and Contribution

We use a data format encoded with JSON Lines for evaluation. The format includes information on models, prompts, reviewers, questions, answers, and reviews.

You can customize the evaluation process or contribute to our project by accessing the relevant data.

For detailed instructions, please refer to the evaluation documentation.

Fine-tuning

Data

Vicuna is created by fine-tuning a LLaMA base model using approximately 70K user-shared conversations gathered from ShareGPT.com with public APIs. To ensure data quality, we convert the HTML back to markdown and filter out some inappropriate or low-quality samples. Additionally, we divide lengthy conversations into smaller segments that fit the model's maximum context length. For detailed instructions to clean the ShareGPT data, check out here.

Due to some concerns, we may not release the ShareGPT dataset at the moment. If you would like to try the fine-tuning code, you can run it with some dummy conversations in dummy_conversation.json. You can follow the same format and plug in your own data.

Code and Hyperparameters

Our code is based on Stanford Alpaca with additional support for multi-turn conversations. We use similar hyperparameters as the Stanford Alpaca.

Hyperparameter Global Batch Size Learning rate Epochs Max length Weight decay
Vicuna-13B 128 2e-5 3 2048 0

Fine-tuning Vicuna-7B with Local GPUs

You can use the following command to train Vicuna-7B with 4 x A100 (40GB).

torchrun --nproc_per_node=4 --master_port=20001 fastchat/train/train_mem.py \
    --model_name_or_path ~/model_weights/llama-7b  \
    --data_path data/dummy_conversation.json \
    --bf16 True \
    --output_dir output_vicuna \
    --num_train_epochs 3 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1200 \
    --save_total_limit 10 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True

If you meet out-of-memory during model saving, see solutions here.

Fine-tuning FastChat-T5 with Local GPUs

You can use the following command to train FastChat-T5 with 4 x A100 (40GB).

torchrun --nproc_per_node=4 --master_port=9778 fastchat/train/train_flant5.py \
    --model_name_or_path google/flan-t5-xl \
    --data_path /data/dummy.json \
    --bf16 True \
    --output_dir ./checkpoints_flant5_3b \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 300 \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap T5Block \
    --tf32 True \
    --model_max_length 2048 \
    --preprocessed_path ./preprocessed_data/processed.json \
    --gradient_checkpointing True 

After training, please use our post-processing function to update the saved model weight. Additional discussions can be found here.

Fine-tuning on Any Cloud with SkyPilot

SkyPilot is a framework built by UC Berkeley for easily and cost effectively running ML workloads on any cloud (AWS, GCP, Azure, Lambda, etc.). To use SkyPilot, install it with the following command and setup the cloud credentials locally following the instructions here.

# Install skypilot from the master branch
pip install git+https://github.com/skypilot-org/skypilot.git

Vicuna

Vicuna can be trained on 8 A100 GPUs with 80GB memory. The following command will automatically launch a node satisfying the requirement, setup and run the training job on it.

sky launch -c vicuna -s scripts/train-vicuna.yaml --env WANDB_API_KEY

Other options are also valid:

# Launch it on managed spot to save 3x cost (train Vicuna-13B with around $300)
sky spot launch -n vicuna scripts/train-vicuna.yaml --env WANDB_API_KEY

# Train a 7B model
sky launch -c vicuna -s scripts/train-vicuna.yaml --env WANDB_API_KEY --env MODEL_SIZE=7

Note: Please make sure the WANDB_API_KEY has been setup on your local machine. You can find the API key on your wandb profile page. If you would like to train the model without using wandb, you can replace the --env WANDB_API_KEY flag with --env WANDB_MODE=offline.