FastChat Documentation

Technical Documentation: lm-sys/FastChat

ℹ️ Provenance: Multi-Tier Ingestion · CodeWiki Reference

Overview (`README.md`) Copy Section

FastChat

| Demo | Discord | X |

FastChat is an open platform for training, serving, and evaluating large language model based chatbots.

• FastChat powers Chatbot Arena (lmarena.ai), serving over 10 million chat requests for 70+ LLMs.
• Chatbot Arena has collected over 1.5M human votes from side-by-side LLM battles to compile an online LLM Elo leaderboard.

FastChat's core features include:

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

News Copy Chapter

• [2024/03] 🔥 We released Chatbot Arena technical report.
• [2023/09] We released LMSYS-Chat-1M, a large-scale real-world LLM conversation dataset. Read the report.
• [2023/08] We released Vicuna v1.5 based on Llama 2 with 4K and 16K context lengths. Download weights.
• [2023/07] We released Chatbot Arena Conversations, a dataset containing 33k conversations with human preferences. Download it here.

More

• [2023/08] We released LongChat v1.5 based on Llama 2 with 32K context lengths. Download weights.
• [2023/06] We introduced MT-bench, a challenging multi-turn question set for evaluating chatbots. Check out the blog post.
• [2023/06] We introduced LongChat, our long-context chatbots and evaluation tools. Check out the blog post.
• [2023/05] We introduced Chatbot Arena for battles among LLMs. Check out the blog post.
• [2023/03] We released Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90% ChatGPT Quality. Check out the blog post.

Contents Copy Chapter

• Install
• Model Weights
• Inference with Command Line Interface
• Serving with Web GUI
• API
• Evaluation
• Fine-tuning
• Citation

Install Copy Chapter

Method 1: With pip Copy Section

pip3 install "fschat[model_worker,webui]"

Method 2: From source Copy Section

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

2. Install Package

pip3 install --upgrade pip  # enable PEP 660 support
pip3 install -e ".[model_worker,webui]"

Model Weights Copy Chapter

Vicuna Weights Copy Section

Vicuna is based on Llama 2 and should be used under Llama's model license.

You can use the commands below to start chatting. It will automatically download the weights from Hugging Face repos.
Downloaded weights are stored in a .cache folder in the user's home folder (e.g., ~/.cache/huggingface/hub/<model_name>).

See more command options and how to handle out-of-memory in the "Inference with Command Line Interface" section below.

NOTE: transformers>=4.31 is required for 16K versions.

Old weights: see docs/vicuna_weights_version.md for all versions of weights and their differences.

Other Models Copy Section

Besides Vicuna, we also released two additional models: LongChat and FastChat-T5.
You can use the commands below to chat with them. They will automatically download the weights from Hugging Face repos.

Inference with Command Line Interface Copy Chapter

(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

FastChat supports a wide range of models, including
LLama 2, Vicuna, Alpaca, Baize, ChatGLM, Dolly, Falcon, FastChat-T5, GPT4ALL, Guanaco, MTP, OpenAssistant, OpenChat, RedPajama, StableLM, WizardLM, xDAN-AI and more.

See a complete list of supported models and instructions to add a new model here.

Single GPU

The command below requires around 14GB of GPU memory for Vicuna-7B and 28GB of GPU memory for Vicuna-13B.
See the "Not Enough Memory" section below if you do not have enough memory.
--model-path can be a local folder or a Hugging Face repo name.

python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5

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 lmsys/vicuna-7b-v1.5 --num-gpus 2

Tips:
Sometimes the "auto" device mapping strategy in huggingface/transformers does not perfectly balance the memory allocation across multiple GPUs.
You can use --max-gpu-memory to specify the maximum memory per GPU for storing model weights.
This allows it to allocate more memory for activations, so you can use longer context lengths or larger batch sizes. For example,

python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5 --num-gpus 2 --max-gpu-memory 8GiB

CPU Only

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

python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5 --device cpu

Use Intel AI Accelerator AVX512_BF16/AMX to accelerate CPU inference.

CPU_ISA=amx python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5 --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 lmsys/vicuna-7b-v1.5 --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 lmsys/vicuna-7b-v1.5 --device xpu

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

Ascend NPU

Install the Ascend PyTorch Adapter. Set the CANN environment variables:

source /usr/local/Ascend/ascend-toolkit/set_env.sh

Use --device npu to enable NPU acceleration.

python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5 --device npu

Vicuna-7B/13B can run on an Ascend NPU.

Not 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 GPU with 16 GB of VRAM, like an Nvidia RTX 3090, RTX 4080, T4, V100 (16GB), or an AMD RX 6800 XT.

python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5 --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 and Quantization

• For AMD GPU users, please install ROCm and the ROCm version of PyTorch before you install FastChat. See also this post.
• FastChat supports ExLlama V2. See docs/exllama_v2.md.
• FastChat supports GPTQ 4bit inference with GPTQ-for-LLaMa. See docs/gptq.md.
• FastChat supports AWQ 4bit inference with mit-han-lab/llm-awq. See docs/awq.md.
• MLC LLM, backed by TVM Unity compiler, deploys Vicuna natively on phones, consumer-class GPUs and web browsers via Vulkan, Metal, CUDA and WebGPU.

Use models from modelscope

For Chinese users, you can use models from www.modelscope.cn via specify the following environment variables.

export FASTCHAT_USE_MODELSCOPE=True

Serving with Web GUI Copy Chapter

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 lmsys/vicuna-7b-v1.5

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-v1.5

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.

Launch Chatbot Arena (side-by-side battle UI) Copy Chapter

Currently, Chatbot Arena is powered by FastChat. Here is how you can launch an instance of Chatbot Arena locally.

FastChat supports popular API-based models such as OpenAI, Anthropic, Gemini, Mistral and more. To add a custom API, please refer to the model support doc. Below we take OpenAI models as an example.

Create a JSON configuration file api_endpoint.json with the api endpoints of the models you want to serve, for example:

{
    "gpt-4o-2024-05-13": {
        "model_name": "gpt-4o-2024-05-13",
        "api_base": "https://api.openai.com/v1",
        "api_type": "openai",
        "api_key": [Insert API Key],
        "anony_only": false
    }
}

For Anthropic models, specify "api_type": "anthropic_message" with your Anthropic key. Similarly, for gemini model, specify "api_type": "gemini". More details can be found in api_provider.py.

To serve your own model using local gpus, follow the instructions in Serving with Web GUI.

Now you're ready to launch the server:

python3 -m fastchat.serve.gradio_web_server_multi --register-api-endpoint-file api_endpoint.json

(Optional): Advanced Features, Scalability, Third Party UI

• 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/vicuna-7b-v1.5 --controller http://localhost:21001 --port 31000 --worker http://localhost:31000
# worker 1
CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.model_worker --model-path lmsys/fastchat-t5-3b-v1.0 --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

• The default model worker based on huggingface/transformers has great compatibility but can be slow. If you want high-throughput batched serving, you can try vLLM integration.
• If you want to host it on your own UI or third party UI, see Third Party UI.

API Copy Chapter

OpenAI-Compatible RESTful APIs & SDK Copy Section

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.
The REST API is capable of being executed from Google Colab free tier, as demonstrated in the FastChat_API_GoogleColab.ipynb notebook, available in our repository.
See docs/openai_api.md.

Hugging Face Generation APIs Copy Section

See fastchat/serve/huggingface_api.py.

LangChain Integration Copy Section

See docs/langchain_integration.

Evaluation Copy Chapter

We use MT-bench, a set of challenging multi-turn open-ended questions to evaluate models.
To automate the evaluation process, we prompt strong LLMs like GPT-4 to act as judges and assess the quality of the models' responses.
See instructions for running MT-bench at fastchat/llm_judge.

MT-bench is the new recommended way to benchmark your models. If you are still looking for the old 80 questions used in the vicuna blog post, please go to vicuna-blog-eval.

Fine-tuning Copy Chapter

Data Copy Section

Vicuna is created by fine-tuning a Llama base model using approximately 125K 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.

We will not release the ShareGPT dataset. 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 Copy Section

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

Fine-tuning Vicuna-7B with Local GPUs Copy Section

• Install dependency

pip3 install -e ".[train]"

• You can use the following command to train Vicuna-7B with 4 x A100 (40GB). Update --model_name_or_path with the actual path to Llama weights and --data_path with the actual path to data.

torchrun --nproc_per_node=4 --master_port=20001 fastchat/train/train_mem.py \
    --model_name_or_path meta-llama/Llama-2-7b-hf \
    --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

Tips:

• If you are using V100 which is not supported by FlashAttention, you can use the memory-efficient attention implemented in xFormers. Install xformers and replace fastchat/train/train_mem.py above with fastchat/train/train_xformers.py.
• If you meet out-of-memory due to "FSDP Warning: When using FSDP, it is efficient and recommended... ", see solutions here.
• If you meet out-of-memory during model saving, see solutions here.
• To turn on logging to popular experiment tracking tools such as Tensorboard, MLFlow or Weights & Biases, use the report_to argument, e.g. pass --report_to wandb to turn on logging to Weights & Biases.

Other models, platforms and LoRA support Copy Section

More instructions to train other models (e.g., FastChat-T5) and use LoRA are in docs/training.md.

Fine-tuning on Any Cloud with SkyPilot Copy Section

SkyPilot is a framework built by UC Berkeley for easily and cost effectively running ML workloads on any cloud (AWS, GCP, Azure, Lambda, etc.).
Find SkyPilot documentation here on using managed spot instances to train Vicuna and save on your cloud costs.

Citation Copy Chapter

The code (training, serving, and evaluation) in this repository is mostly developed for or derived from the paper below.
Please cite it if you find the repository helpful.

@misc{zheng2023judging,
      title={Judging LLM-as-a-judge with MT-Bench and Chatbot Arena},
      author={Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric. P Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica},
      year={2023},
      eprint={2306.05685},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

We are also planning to add more of our research to this repository.

Overview (`tests/README.md`) Copy Section

Unit tests for FastChat Copy Chapter

Test CLI Inference Copy Section

python3 test_cli.py

Test OpenAI API Server Copy Section

python3 launch_openai_api_test_server.py

python3 test_openai_api.py

Test GUI Serving Copy Section

python3 -m fastchat.serve.controller

/* Detailed source-code truncated for AI context efficiency. */

python3 -m fastchat.serve.gradio_web_server_multi

Test Peft Serving Copy Section

python3 -m fastchat.serve.controller

PEFT_SHARE_BASE_WEIGHTS=true python3 -m fastchat.serve.multi_model_worker \
  --model-path SurfaceData/dummy_pythia160m_lora16_peft_chat \
  --model-path SurfaceData/dummy_pythia160m_lora8_peft_chat

Overview (`fastchat/llm_judge/README.md`) Copy Section

LLM Judge

| Paper | Leaderboard |

In this package, you can use MT-bench questions and prompts to evaluate your models with LLM-as-a-judge.
MT-bench is a set of challenging multi-turn open-ended questions for evaluating chat assistants.
To automate the evaluation process, we prompt strong LLMs like GPT-4 to act as judges and assess the quality of the models' responses.

Contents Copy Chapter

• Install
• Review Pre-Generated Model Answers and Judgments
• MT-Bench
• Agreement Computation
• Datasets
• Citation

Install Copy Chapter

git clone https://github.com/lm-sys/FastChat.git
cd FastChat
pip install -e ".[model_worker,llm_judge]"

Review Pre-Generated Model Answers and Judgments Copy Chapter

We provide pre-generated model answers and judgments for some models.
You can view them at this demo.

To download the pre-generated data, use

python3 download_mt_bench_pregenerated.py

After downloading the data, you can view them locally by

python3 qa_browser.py --share

You can use this QA browser to view the answers generated by you later.

MT-Bench Copy Chapter

Evaluate a model on MT-bench Copy Section

Step 1. Generate model answers to MT-bench questions

python gen_model_answer.py --model-path [MODEL-PATH] --model-id [MODEL-ID]

Arguments:

• [MODEL-PATH] is the path to the weights, which can be a local folder or a Hugging Face repo ID.
• [MODEL-ID] is a name you give to the model.

e.g.,

python gen_model_answer.py --model-path lmsys/vicuna-7b-v1.5 --model-id vicuna-7b-v1.5

The answers will be saved to data/mt_bench/model_answer/[MODEL-ID].jsonl.

To make sure FastChat loads the correct prompt template, see the supported models and how to add a new model here.

You can also specify --num-gpus-per-model for model parallelism (needed for large 65B models) and --num-gpus-total to parallelize answer generation with multiple GPUs.

Note: if you experience slow answer generation, please refer to Other Backends section to use inference engine to speed up by 20x.

Step 2. Generate GPT-4 judgments

There are several options to use GPT-4 as a judge, such as pairwise winrate and single-answer grading.
In MT-bench, we recommend single-answer grading as the default mode.
This mode asks GPT-4 to grade and give a score to model's answer directly without pairwise comparison.
For each turn, GPT-4 will give a score on a scale of 10. We then compute the average score on all turns.

export OPENAI_API_KEY=XXXXXX  # set the OpenAI API key
python gen_judgment.py --model-list [LIST-OF-MODEL-ID] --parallel [num-concurrent-api-call]

e.g.,

python gen_judgment.py --model-list vicuna-13b-v1.3 alpaca-13b llama-13b claude-v1 gpt-3.5-turbo gpt-4 --parallel 2

The judgments will be saved to data/mt_bench/model_judgment/gpt-4_single.jsonl

Step 3. Show MT-bench scores

• Show the scores for selected models
  text Copy

  python show_result.py --model-list vicuna-13b-v1.3 alpaca-13b llama-13b claude-v1 gpt-3.5-turbo gpt-4

• Show all scores
  text Copy

  python show_result.py

Other grading options Copy Section

Besides score-based single-answer grading, we also support two additional grading options based on win rates:

• pariwise-baseline: run pairwise comparison against a baseline model.
• pairwise-all: run pairwise comparison between all model pairs on all questions.

Option 2: pairwise comparison against a baseline (default: gpt-3.5-turbo)

• Generate GPT-4 judgments

python gen_judgment.py --mode pairwise-baseline --model-list vicuna-13b-v1.3 alpaca-13b llama-13b --parallel 2

The judgments will be saved to data/mt_bench/model_judgment/gpt-4_pair.jsonl

• Show results

python show_result.py --mode pairwise-baseline

Option 3: Run GPT-4 judge with all pair comparisons

Another option is to run pairwise comparisons on all possible pairs.
This could be more expensive when #models increases, but it gives you a more comprehensive information.

python gen_judgment.py --mode pairwise-all --model-list [LIST-OF-MODEL-ID] --parallel [num-concurrent-api-call]

python show_result.py --mode pairwise-all

How to get GPT-3.5/GPT-4/Claude's answer? Copy Section

• python gen_api_answer.py --model [MODEL-NAME] to generate GPT-3.5/4 and Claude's answers.

How to plot the radar figure? Copy Section

You can use this colab notebook to plot the radar figure for MT-bench.

Other backends Copy Section

We can also use vLLM for answer generation, which can be faster for the models supported by vLLM.

1. Launch a vLLM worker

vllm serve [MODEL-PATH] --dtype auto

• Arguments:
  • [MODEL-PATH] is the path to the weights, which can be a local folder or a Hugging Face repo ID.

2. Generate the answers

python gen_api_answer.py --model [MODEL-NAME] --openai-api-base http://localhost:8000/v1 --parallel 50

• Arguments:
  • [MODEL-NAME] is the name of the model from Step 1.
  • --parallel is the number of concurrent API calls to the vLLM worker.

Agreement Computation Copy Chapter

We released 3.3K human annotations for model responses generated by 6 models in response to 80 MT-bench questions. The dataset is available at lmsys/mt_bench_human_judgments.

This Colab notebook shows how to compute the agreement between humans and GPT-4 judge with the dataset. Our results show that humans and GPT-4 judge achieve over 80% agreement, the same level of agreement between humans.

Datasets Copy Chapter

• Chatbot Arena Conversation Dataset
• MT-bench Human Annotation Dataset

Citation Copy Chapter

Please cite the following paper if you find the code or datasets helpful.

@misc{zheng2023judging,
      title={Judging LLM-as-a-judge with MT-Bench and Chatbot Arena}, 
      author={Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric. P Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica},
      year={2023},
      eprint={2306.05685},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Overview (`playground/test_embedding/README.md`) Copy Section

Machine Learning with Embeddings Copy Chapter

You can use embeddings to

• Evaluate text similarity, see test_sentence_similarity.py
• Build your own classifier, see test_classification.py
• Search relative texts, see test_semantic_search.py

To these tests, you need to download the data here. You also need an OpenAI API key for comparison.

Run with:

cd playground/test_embedding
python3 test_classification.py

The script will train classifiers based on vicuna-7b, text-similarity-ada-001 and text-embedding-ada-002 and report the accuracy of each classifier.

Overview (`fastchat/serve/gateway/README.md`) Copy Section

fastchat Nginx Gateway

Purpose of the Gateway Copy Chapter

The Nginx gateway serves the following purposes:

1. Protects Gradio servers by acting as a firewall.
2. Facilitates dynamic mounting and unmounting of Gradio servers.
3. Provides load balancing for Gradio servers.
4. Offers additional security features, such as total connection limit.
5. Reduces attack surface by requiring only a single public port to be exposed for serving.

Deployment and Updating of the Gateway Copy Chapter

Installing Nginx Copy Section

On Debian-based distributions (e.g., Ubuntu):

sudo apt update
sudo apt install nginx

On Red Hat-based distributions (e.g., CentOS, Fedora):

sudo yum install epel-release
sudo yum install nginx

Deployment Copy Section

Copy nginx.conf to /etc/nginx/nginx.conf (need sudo permission).

Replace the port number 7860 in server localhost:7860 with the port where you deploy the Gradio web server.

Modify upstream websocket to configure Gradio servers behind the gateway.

Lastly, update Nginx.

HTTPS Deployment with a Public Domain URL Copy Section

Make sure you obtain the HTTPS certificate and the private key used to generate the certificate.

Fill the addresses to your certificate and private key in the [PATH_TO_SSL_CERT] and [PATH_TO_PRIVATE_KEY] fields.

If you have your own domain url to serve the chatbot, replace the chat.lmsys.org url with your own domain url.

Updating Copy Section

Every time when /etc/nginx/nginx.conf is modified, you need to update the Nginx service:

sudo nginx -t  # check `/etc/nginx/nginx.conf`
sudo systemctl reload nginx  # restart Nginx service to load the new config
sudo systemctl status nginx  # check the status of the Nginx service. It should be active (running).

Overview (`fastchat/serve/monitor/classify/README.md`) Copy Section

Download dataset Copy Chapter

We have pre-generated several category classifier benchmarks and ground truths. You can download them (with git-lfs installed) to the directory classify/ by running

> git clone https://huggingface.co/datasets/lmarena-ai/categories-benchmark-eval
// cd into classify/ and then copy the label_bench directory to the current directory
> cp -r categories-benchmark-eval/label_bench .

Your label_bench directory should follow the structure:

├── label_bench/
│   ├── creative_writing_bench/
│   │   ├── data/
│   │   │    └── llama-v3p1-70b-instruct.json
│   │   └── test.json
│   ├── ...
│   ├── your_bench_name/
│   │   ├── data/
│   │   │    ├── your_classifier_data_1.json
│   │   │    ├── your_classifier_data_2.json
│   │   │    └── ...
│   │   └── test.json (your ground truth)
└── ...

How to evaluate your category classifier? Copy Chapter

To test your new classifier for a new category, you would have to make sure you created the category child class in category.py. Then, to generate classification labels, make the necessary edits in config.yaml and run

python label.py --config config.yaml --testing

If you are labeling a vision category, add the --vision flag to the command. This will add a new column to the input data called image_path that contains the path to the image corresponding to each conversation. Ensure that you update your config with the correct image_dir where the images are stored.

Then, add your new category bench to tag_names in display_score.py. After making sure that you also have a correctly formatted ground truth json file, you can report the performance of your classifier by running

python display_score.py --bench <your_bench>

If you want to check out conflicts between your classifier and ground truth, use

python display_score.py --bench <your_bench> --display-conflict

Example output:

> python display_score.py --bench if_bench --display-conflict
Model: gpt-4o-mini-2024-07-18
Accuracy: 0.967
Precision: 0.684
Recall: 0.918

###### CONFLICT ######

Ground Truth = True; Pred = False
\####################
...

Ground Truth = False; Pred = True
\####################
...

Overview (`fastchat/serve/monitor/vote_time_stats/README.md`) Copy Section

Instructions

First run analyze_data.py to collect metadata of all votes.

Then run plot.py to get the plot. You need to edit these files for proper input or output filename

Arena (`docs/arena.md`) Copy Section

Chatbot Arena

Chatbot Arena is an LLM benchmark platform featuring anonymous, randomized battles, available at https://lmarena.ai.
We invite the entire community to join this benchmarking effort by contributing your votes and models.

How to add a new model Copy Chapter

If you want to see a specific model in the arena, you can follow the methods below.

Method 1: Hosted by 3rd party API providers or yourself Copy Section

If you have a model hosted by a 3rd party API provider or yourself, please give us the access to an API endpoint.

• We prefer OpenAI-compatible APIs, so we can reuse our code for calling OpenAI models.
• If you have your own API protocol, please follow the instructions to add them. Contribute your code by sending a pull request.

Method 2: Hosted by LMSYS Copy Section

1. Contribute the code to support this model in FastChat by submitting a pull request. See instructions.
2. After the model is supported, we will try to schedule some compute resources to host the model in the arena. However, due to the limited resources we have, we may not be able to serve every model. We will select the models based on popularity, quality, diversity, and other factors.

How to launch vision arena Copy Chapter

1. Run python3 -m fastchat.serve.controller to start the controller and begin registering local model workers and API-provided workers.
2. Run python3 -m fastchat.serve.sglang_worker --model-path <model-path> --tokenizer-path <tokenizer-path> to run local vision-language models. Currently supported models include the LLaVA and Yi-VL series.
3. If you are using a 3rd party model with an API provider (e.g. GPT-4-V, Gemini 1.5), please follow the instructions model_support.md to add a json file api_endpoints.json.
4. Run the gradio server with the --vision-arena flag on.
5. To run and store images into a remote directory, add the flag: --use-remote-storage
6. To run and allow samples of random questions, add --random_questions metadata_sampled.json. Check sections below for how to generate this.

Example command:

python3 -m fastchat.serve.gradio_web_server_multi --share --register-api-endpoint-file api_endpoints.json --vision-arena --use-remote-storage --random-questions metadata_sampled.json

NSFW and CSAM Detection Copy Section

1. Adding NSFW Endpoint and API key: Please add the following environment variables to run the NSFW moderation filter for images:

• AZURE_IMG_MODERATION_ENDPOINT: This is the endpoint that the NSFW moderator is hosted (e.g. https://{endpoint}/contentmoderator/moderate/v1.0/ProcessImage/Evaluate). Change the endpoint to your own.
• AZURE_IMG_MODERATION_API_KEY: Your API key to run this endpoint.

2. Adding CSAM API key:

• PHOTODNA_API_KEY: The API key that runs the CSAM detector endpoint.

Example in ~/.bashrc:

export AZURE_IMG_MODERATION_ENDPOINT=https://<endpoint>/contentmoderator/moderate/v1.0/ProcessImage/Evaluate
export AZURE_IMG_MODERATION_API_KEY=<api-key>
export PHOTODNA_API_KEY=<api-key>

Adding Random Samples for VQA Copy Section

We provide random samples of example images for users to interact with coming from various datasets including DocVQA, RealWorldQA, ChartQA and VizWiz-VQA.

1. Download the images and generate random questions file by running python fastchat/serve/vision/create_vqa_examples_dir.py

Awq (`docs/awq.md`) Copy Section

AWQ 4bit Inference

We integrated AWQ into FastChat to provide efficient and accurate 4bit LLM inference.

Install AWQ Copy Chapter

Setup environment (please refer to this link for more details):

conda create -n fastchat-awq python=3.10 -y
conda activate fastchat-awq
# cd /path/to/FastChat
pip install --upgrade pip    # enable PEP 660 support
pip install -e .             # install fastchat

git clone https://github.com/mit-han-lab/llm-awq repositories/llm-awq
cd repositories/llm-awq
pip install -e .             # install awq package

cd awq/kernels				
python setup.py install	     # install awq CUDA kernels

Chat with the CLI Copy Chapter

# Download quantized model from huggingface
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/mit-han-lab/vicuna-7b-v1.3-4bit-g128-awq

# You can specify which quantized model to use by setting --awq-ckpt
python3 -m fastchat.serve.cli \
    --model-path models/vicuna-7b-v1.3-4bit-g128-awq \
    --awq-wbits 4 \
    --awq-groupsize 128

Benchmark Copy Chapter

• Through 4-bit weight quantization, AWQ helps to run larger language models within the device memory restriction and prominently accelerates token generation. All benchmarks are done with group_size 128.

• Benchmark on NVIDIA RTX A6000:

  Model	Bits	Max Memory (MiB)	Speed (ms/token)	AWQ Speedup
  vicuna-7b	16	13543	26.06	/
  vicuna-7b	4	5547	12.43	2.1x
  llama2-7b-chat	16	13543	27.14	/
  llama2-7b-chat	4	5547	12.44	2.2x
  vicuna-13b	16	25647	44.91	/
  vicuna-13b	4	9355	17.30	2.6x
  llama2-13b-chat	16	25647	47.28	/
  llama2-13b-chat	4	9355	20.28	2.3x

• NVIDIA RTX 4090:

  Model	AWQ 4bit Speed (ms/token)	FP16 Speed (ms/token)	AWQ Speedup
  vicuna-7b	8.61	19.09	2.2x
  llama2-7b-chat	8.66	19.97	2.3x
  vicuna-13b	12.17	OOM	/
  llama2-13b-chat	13.54	OOM	/

• NVIDIA Jetson Orin:

  Model	AWQ 4bit Speed (ms/token)	FP16 Speed (ms/token)	AWQ Speedup
  vicuna-7b	65.34	93.12	1.4x
  llama2-7b-chat	75.11	104.71	1.4x
  vicuna-13b	115.40	OOM	/
  llama2-13b-chat	136.81	OOM	/

Dashinfer Integration (`docs/dashinfer_integration.md`) Copy Section

dash-infer Integration

DashInfer is a high-performance inference engine specifically optimized for CPU environments, delivering exceptional performance boosts for LLM inference tasks. It supports acceleration for a variety of models including Llama, Qwen, and ChatGLM, making it a versatile choice as a performant worker in FastChat. Notably, DashInfer exhibits significant performance enhancements on both Intel x64 and ARMv9 processors, catering to a wide spectrum of hardware platforms. Its efficient design and optimization techniques ensure rapid and accurate inference capabilities, making it an ideal solution for deploying large language models in resource-constrained environments or scenarios where CPU utilization is preferred over GPU acceleration.

Instructions Copy Chapter

1. Install dash-infer.

   text Copy

   pip install dashinfer

2. When you launch a model worker, replace the normal worker (fastchat.serve.model_worker) with the dash-infer worker (fastchat.serve.dashinfer_worker). All other commands such as controller, gradio web server, and OpenAI API server are kept the same.

   text Copy

   python3 -m fastchat.serve.dashinfer_worker --model-path qwen/Qwen-7B-Chat --revision=master /path/to/dashinfer-model-generation-config.json

Here is an example:

python3 -m fastchat.serve.dashinfer_worker --model-path qwen/Qwen-7B-Chat --revision=master dash-infer/examples/python/model_config/config_qwen_v10_7b.json

If you use an already downloaded model, try to replace model-path with a local one and choose a conversation template via --conv-template option
'''
python3 -m fastchat.serve.dashinfer_worker --model-path ~/.cache/modelscope/hub/qwen/Qwen-7B-Chat --conv-template qwen-7b-chat /path/to/dashinfer-model-generation-config.json
'''
All avaliable conversation chat templates are listed at fastchat/conversation.py

Dataset Release (`docs/dataset_release.md`) Copy Section

Datasets Copy Chapter

We release the following datasets based on our projects and websites.

• LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset
• LMSYS-Human-Preference-55k
• Chatbot Arena Conversation Dataset
• MT-bench Human Annotation Dataset

Exllama V2 (`docs/exllama_v2.md`) Copy Section

ExllamaV2 GPTQ Inference Framework

Integrated ExllamaV2 customized kernel into Fastchat to provide Faster GPTQ inference speed.

Note: Exllama not yet support embedding REST API.

Install ExllamaV2 Copy Chapter

Setup environment (please refer to this link for more details):

git clone https://github.com/turboderp/exllamav2
cd exllamav2
pip install -e .

Chat with the CLI:

python3 -m fastchat.serve.cli \
    --model-path models/vicuna-7B-1.1-GPTQ-4bit-128g \
    --enable-exllama

Start model worker:

# Download quantized model from huggingface
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/TheBloke/vicuna-7B-1.1-GPTQ-4bit-128g models/vicuna-7B-1.1-GPTQ-4bit-128g

# Load model with default configuration (max sequence length 4096, no GPU split setting).
python3 -m fastchat.serve.model_worker \
    --model-path models/vicuna-7B-1.1-GPTQ-4bit-128g \
    --enable-exllama

#Load model with max sequence length 2048, allocate 18 GB to CUDA:0 and 24 GB to CUDA:1.
python3 -m fastchat.serve.model_worker \
    --model-path models/vicuna-7B-1.1-GPTQ-4bit-128g \
    --enable-exllama \
    --exllama-max-seq-len 2048 \
    --exllama-gpu-split 18,24

--exllama-cache-8bit can be used to enable 8-bit caching with exllama and save some VRAM.

Performance Copy Chapter

Reference: https://github.com/turboderp/exllamav2#performance

Gptq (`docs/gptq.md`) Copy Section

GPTQ 4bit Inference

Support GPTQ 4bit inference with GPTQ-for-LLaMa.

1. Window user: use the old-cuda branch.
2. Linux user: recommend the fastest-inference-4bit branch.

Install Copy Chapter

Setup environment:

# cd /path/to/FastChat
git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git repositories/GPTQ-for-LLaMa
cd repositories/GPTQ-for-LLaMa
# Window's user should use the `old-cuda` branch
git switch fastest-inference-4bit
# Install `quant-cuda` package in FastChat's virtualenv
python3 setup_cuda.py install
pip3 install texttable

Chat with the CLI:

python3 -m fastchat.serve.cli \
    --model-path models/vicuna-7B-1.1-GPTQ-4bit-128g \
    --gptq-wbits 4 \
    --gptq-groupsize 128

Start model worker:

# Download quantized model from huggingface
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/TheBloke/vicuna-7B-1.1-GPTQ-4bit-128g models/vicuna-7B-1.1-GPTQ-4bit-128g

python3 -m fastchat.serve.model_worker \
    --model-path models/vicuna-7B-1.1-GPTQ-4bit-128g \
    --gptq-wbits 4 \
    --gptq-groupsize 128

# You can specify which quantized model to use
python3 -m fastchat.serve.model_worker \
    --model-path models/vicuna-7B-1.1-GPTQ-4bit-128g \
    --gptq-ckpt models/vicuna-7B-1.1-GPTQ-4bit-128g/vicuna-7B-1.1-GPTQ-4bit-128g.safetensors \
    --gptq-wbits 4 \
    --gptq-groupsize 128 \
    --gptq-act-order

Benchmark Copy Chapter

Langchain Integration (`docs/langchain_integration.md`) Copy Section

Local LangChain with FastChat

LangChain is a library that facilitates the development of applications by leveraging large language models (LLMs) and enabling their composition with other sources of computation or knowledge.
FastChat's OpenAI-compatible API server enables using LangChain with open models seamlessly.

Launch RESTful API Server Copy Chapter

Here are the steps to launch a local OpenAI API server for LangChain.

First, launch the controller

python3 -m fastchat.serve.controller

LangChain uses OpenAI model names by default, so we need to assign some faux OpenAI model names to our local model.
Here, we use Vicuna as an example and use it for three endpoints: chat completion, completion, and embedding.
--model-path can be a local folder or a Hugging Face repo name.
See a full list of supported models here.

python3 -m fastchat.serve.model_worker --model-names "gpt-3.5-turbo,text-davinci-003,text-embedding-ada-002" --model-path lmsys/vicuna-7b-v1.5

Finally, launch the RESTful API server

python3 -m fastchat.serve.openai_api_server --host localhost --port 8000

Set OpenAI Environment Copy Chapter

You can set your environment with the following commands.

Set OpenAI base url

export OPENAI_API_BASE=http://localhost:8000/v1

Set OpenAI API key

export OPENAI_API_KEY=EMPTY

If you meet the following OOM error while creating embeddings, please set a smaller batch size by using environment variables.

openai.error.APIError: Invalid response object from API: '{"object":"error","message":"**NETWORK ERROR DUE TO HIGH TRAFFIC. PLEASE REGENERATE OR REFRESH THIS PAGE.**\\n\\n(CUDA out of memory. Tried to allocate xxx MiB (GPU 0; xxx GiB total capacity; xxx GiB already allocated; xxx MiB free; xxx GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF)","code":50002}' (HTTP response code was 400)

You can try export FASTCHAT_WORKER_API_EMBEDDING_BATCH_SIZE=1.

Try local LangChain Copy Chapter

Here is a question answerting example.

Download a text file.

wget https://raw.githubusercontent.com/hwchase17/langchain/v0.0.200/docs/modules/state_of_the_union.txt

Run LangChain.

from langchain.chat_models import ChatOpenAI
from langchain.document_loaders import TextLoader
from langchain.embeddings import OpenAIEmbeddings
from langchain.indexes import VectorstoreIndexCreator

embedding = OpenAIEmbeddings(model="text-embedding-ada-002")
loader = TextLoader("state_of_the_union.txt")
index = VectorstoreIndexCreator(embedding=embedding).from_loaders([loader])
llm = ChatOpenAI(model="gpt-3.5-turbo")

questions = [
    "Who is the speaker",
    "What did the president say about Ketanji Brown Jackson",
    "What are the threats to America",
    "Who are mentioned in the speech",
    "Who is the vice president",
    "How many projects were announced",
]

for query in questions:
    print("Query:", query)
    print("Answer:", index.query(query, llm=llm))

Lightllm Integration (`docs/lightllm_integration.md`) Copy Section

LightLLM Integration

You can use LightLLM as an optimized worker implementation in FastChat.
It offers advanced continuous batching and a much higher (~10x) throughput.
See the supported models here.

Instructions Copy Chapter

1. Please refer to the Get started to install LightLLM. Or use Pre-built image

2. When you launch a model worker, replace the normal worker (fastchat.serve.model_worker) with the LightLLM worker (fastchat.serve.lightllm_worker). All other commands such as controller, gradio web server, and OpenAI API server are kept the same. Refer to --max_total_token_num to understand how to calculate the --max_total_token_num argument.

   text Copy

   python3 -m fastchat.serve.lightllm_worker --model-path lmsys/vicuna-7b-v1.5 --tokenizer_mode "auto" --max_total_token_num 154000

   If you what to use quantized weight and kv cache for inference, try

   text Copy

   python3 -m fastchat.serve.lightllm_worker --model-path lmsys/vicuna-7b-v1.5 --tokenizer_mode "auto" --max_total_token_num 154000 --mode triton_int8weight triton_int8kv

Mlx Integration (`docs/mlx_integration.md`) Copy Section

Apple MLX Integration

You can use Apple MLX as an optimized worker implementation in FastChat.

It runs models efficiently on Apple Silicon

See the supported models here.

Note that for Apple Silicon Macs with less memory, smaller models (or quantized models) are recommended.

Instructions Copy Chapter

1. Install MLX.

   text Copy

   pip install "mlx-lm>=0.0.6"

2. When you launch a model worker, replace the normal worker (fastchat.serve.model_worker) with the MLX worker (fastchat.serve.mlx_worker). Remember to launch a model worker after you have launched the controller (instructions)

   text Copy

   python3 -m fastchat.serve.mlx_worker --model-path TinyLlama/TinyLlama-1.1B-Chat-v1.0

Model Support (`docs/model_support.md`) Copy Section

Model Support

This document describes how to support a new model in FastChat.

Content Copy Chapter

• Local Models
• API-Based Models

Local Models Copy Chapter

To support a new local model in FastChat, you need to correctly handle its prompt template and model loading.
The goal is to make the following command run with the correct prompts.

python3 -m fastchat.serve.cli --model [YOUR_MODEL_PATH]

You can run this example command to learn the code logic.

python3 -m fastchat.serve.cli --model lmsys/vicuna-7b-v1.5

You can add --debug to see the actual prompt sent to the model.

Steps Copy Section

FastChat uses the Conversation class to handle prompt templates and BaseModelAdapter class to handle model loading.

1. Implement a conversation template for the new model at fastchat/conversation.py. You can follow existing examples and use register_conv_template to add a new one. Please also add a link to the official reference code if possible.
2. Implement a model adapter for the new model at fastchat/model/model_adapter.py. You can follow existing examples and use register_model_adapter to add a new one.
3. (Optional) add the model name to the "Supported models" section above and add more information in fastchat/model/model_registry.py.

After these steps, the new model should be compatible with most FastChat features, such as CLI, web UI, model worker, and OpenAI-compatible API server. Please do some testing with these features as well.

Supported models Copy Section

• meta-llama/Llama-2-7b-chat-hf
  • example: python3 -m fastchat.serve.cli --model-path meta-llama/Llama-2-7b-chat-hf
• Vicuna, Alpaca, LLaMA, Koala
  • example: python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5
• allenai/tulu-2-dpo-7b
• BAAI/AquilaChat-7B
• BAAI/AquilaChat2-7B
• BAAI/AquilaChat2-34B
• BAAI/bge-large-en
• argilla/notus-7b-v1
• baichuan-inc/baichuan-7B
• BlinkDL/RWKV-4-Raven
  • example: python3 -m fastchat.serve.cli --model-path ~/model_weights/RWKV-4-Raven-7B-v11x-Eng99%-Other1%-20230429-ctx8192.pth
• bofenghuang/vigogne-2-7b-instruct
• bofenghuang/vigogne-2-7b-chat
• camel-ai/CAMEL-13B-Combined-Data
• codellama/CodeLlama-7b-Instruct-hf
• databricks/dolly-v2-12b
• deepseek-ai/deepseek-llm-67b-chat
• deepseek-ai/deepseek-coder-33b-instruct
• FlagAlpha/Llama2-Chinese-13b-Chat
• FreedomIntelligence/phoenix-inst-chat-7b
• FreedomIntelligence/ReaLM-7b-v1
• h2oai/h2ogpt-gm-oasst1-en-2048-open-llama-7b
• HuggingFaceH4/starchat-beta
• HuggingFaceH4/zephyr-7b-alpha
• internlm/internlm-chat-7b
• cllm/consistency-llm-7b-codesearchnet/consistency-llm-7b-gsm8k/consistency-llm-7b-sharegpt48k/consistency-llm-7b-spider
• IEITYuan/Yuan2-2B/51B/102B-hf
• lcw99/polyglot-ko-12.8b-chang-instruct-chat
• lmsys/fastchat-t5-3b-v1.0
• meta-math/MetaMath-7B-V1.0
• Microsoft/Orca-2-7b
• mosaicml/mpt-7b-chat
  • example: python3 -m fastchat.serve.cli --model-path mosaicml/mpt-7b-chat
• Neutralzz/BiLLa-7B-SFT
• nomic-ai/gpt4all-13b-snoozy
• NousResearch/Nous-Hermes-13b
• openaccess-ai-collective/manticore-13b-chat-pyg
• OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5
• openchat/openchat_3.5
• Open-Orca/Mistral-7B-OpenOrca
• OpenLemur/lemur-70b-chat-v1
• Phind/Phind-CodeLlama-34B-v2
• project-baize/baize-v2-7b
• Qwen/Qwen-7B-Chat
• rishiraj/CatPPT
• Salesforce/codet5p-6b
• StabilityAI/stablelm-tuned-alpha-7b
• tenyx/TenyxChat-7B-v1
• TinyLlama/TinyLlama-1.1B-Chat-v1.0
• THUDM/chatglm-6b
• THUDM/chatglm2-6b
• tiiuae/falcon-40b
• tiiuae/falcon-180B-chat
• timdettmers/guanaco-33b-merged
• togethercomputer/RedPajama-INCITE-7B-Chat
• VMware/open-llama-7b-v2-open-instruct
• WizardLM/WizardLM-13B-V1.0
• WizardLM/WizardCoder-15B-V1.0
• Xwin-LM/Xwin-LM-7B-V0.1
• Any EleutherAI pythia model such as pythia-6.9b
• Any Peft adapter trained on top of a
  model above. To activate, must have peft in the model path. Note: If
  loading multiple peft models, you can have them share the base model weights by
  setting the environment variable PEFT_SHARE_BASE_WEIGHTS=true in any model
  worker.

API-Based Models Copy Chapter

To support an API-based model, consider learning from the existing OpenAI example.
If the model is compatible with OpenAI APIs, then a configuration file is all that's needed without any additional code.
For custom protocols, implementation of a streaming generator in fastchat/serve/api_provider.py is required, following the provided examples. Currently, FastChat is compatible with OpenAI, Anthropic, Google Vertex AI, Mistral, Nvidia NGC, YandexGPT and Reka.

Steps to Launch a WebUI with an API Model Copy Section

1. Specify the endpoint information in a JSON configuration file. For instance, create a file named api_endpoints.json:

{
  "gpt-3.5-turbo": {
    "model_name": "gpt-3.5-turbo",
    "api_type": "openai",
    "api_base": "https://api.openai.com/v1",
    "api_key": "sk-******",
    "anony_only": false,
    "recommended_config": {
      "temperature": 0.7,
      "top_p": 1.0
    },
    "text-arena": true,
    "vision-arena": false,
  }
}

• "api_type" can be one of the following: openai, anthropic, gemini, mistral, yandexgpt or reka. For custom APIs, add a new type and implement it accordingly.
• "anony_only" indicates whether to display this model in anonymous mode only.
• "recommended_config" indicates the recommended generation parameters for temperature and top_p.
• "text-arena" indicates whether the model should be displayed in the Text Arena.
• "vision-arena" indicates whether the model should be displayed in the Vision Arena.

2. Launch the Gradio web server with the argument --register api_endpoints.json:

python3 -m fastchat.serve.gradio_web_server --controller "" --share --register api_endpoints.json

Now, you can open a browser and interact with the model.

Openai Api (`docs/openai_api.md`) Copy Section

OpenAI-Compatible RESTful APIs

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.

The following OpenAI APIs are supported:

• Chat Completions. (Reference: https://platform.openai.com/docs/api-reference/chat)
• Completions. (Reference: https://platform.openai.com/docs/api-reference/completions)
• Embeddings. (Reference: https://platform.openai.com/docs/api-reference/embeddings)

The REST API can be seamlessly operated from Google Colab, as demonstrated in the FastChat_API_GoogleColab.ipynb notebook, available in our repository. This notebook provides a practical example of how to utilize the API effectively within the Google Colab environment.

RESTful API Server Copy Chapter

First, launch the controller

python3 -m fastchat.serve.controller

Then, launch the model worker(s)

python3 -m fastchat.serve.model_worker --model-path lmsys/vicuna-7b-v1.5

Finally, launch the RESTful API server

python3 -m fastchat.serve.openai_api_server --host localhost --port 8000

Now, let us test the API server.

OpenAI Official SDK Copy Section

The goal of openai_api_server.py is to implement a fully OpenAI-compatible API server, so the models can be used directly with openai-python library.

First, install OpenAI python package >= 1.0:

pip install --upgrade openai

Then, interact with the Vicuna model:

import openai

openai.api_key = "EMPTY"
openai.base_url = "http://localhost:8000/v1/"

model = "vicuna-7b-v1.5"
prompt = "Once upon a time"

# create a completion
completion = openai.completions.create(model=model, prompt=prompt, max_tokens=64)
# print the completion
print(prompt + completion.choices[0].text)

# create a chat completion
completion = openai.chat.completions.create(
  model=model,
  messages=[{"role": "user", "content": "Hello! What is your name?"}]
)
# print the completion
print(completion.choices[0].message.content)

Streaming is also supported. See test_openai_api.py. If your api server is behind a proxy you'll need to turn off buffering, you can do so in Nginx by setting proxy_buffering off; in the location block for the proxy.

cURL Copy Section

cURL is another good tool for observing the output of the api.

List Models:

curl http://localhost:8000/v1/models

Chat Completions:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vicuna-7b-v1.5",
    "messages": [{"role": "user", "content": "Hello! What is your name?"}]
  }'

Text Completions:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vicuna-7b-v1.5",
    "prompt": "Once upon a time",
    "max_tokens": 41,
    "temperature": 0.5
  }'

Embeddings:

curl http://localhost:8000/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vicuna-7b-v1.5",
    "input": "Hello world!"
  }'

Running multiple Copy Section

If you want to run multiple models on the same machine and in the same process,
you can replace the model_worker step above with a multi model variant:

python3 -m fastchat.serve.multi_model_worker \
    --model-path lmsys/vicuna-7b-v1.5 \
    --model-names vicuna-7b-v1.5 \
    --model-path lmsys/longchat-7b-16k \
    --model-names longchat-7b-16k

This loads both models into the same accelerator and in the same process. This
works best when using a Peft model that triggers the PeftModelAdapter.

TODO: Base model weight optimization will be fixed once this
Peft issue is resolved.

LangChain Support Copy Chapter

This OpenAI-compatible API server supports LangChain. See LangChain Integration for details.

Adjusting Environment Variables Copy Chapter

Timeout Copy Section

By default, a timeout error will occur if a model worker does not response within 100 seconds. If your model/hardware is slower, you can change this timeout through an environment variable:

export FASTCHAT_WORKER_API_TIMEOUT=<larger timeout in seconds>

Batch size Copy Section

If you meet the following OOM error while creating embeddings. You can use a smaller batch size by setting

export FASTCHAT_WORKER_API_EMBEDDING_BATCH_SIZE=1

Todos Copy Chapter

Some features to be implemented:

• Support more parameters like logprobs, logit_bias, user, presence_penalty and frequency_penalty
• Model details (permissions, owner and create time)
• Edits API
• Rate Limitation Settings

Third Party Ui (`docs/third_party_ui.md`) Copy Section

Third Party UI

If you want to host it on your own UI or third party UI, you can launch the OpenAI compatible server and host with a tunnelling service such as Tunnelmole or ngrok, and then enter the credentials appropriately.

You can find suitable UIs from third party repos:

• WongSaang's ChatGPT UI

• McKayWrigley's Chatbot UI

• Please note that some third-party providers only offer the standard gpt-3.5-turbo, gpt-4, etc., so you will have to add your own custom model inside the code. Here is an example of how to create a UI with any custom model name.

Using Tunnelmole

Tunnelmole is an open source tunnelling tool. You can find its source code on Github. Here's how you can use Tunnelmole:

1. Install Tunnelmole with curl -O https://install.tunnelmole.com/9Wtxu/install && sudo bash install. (On Windows, download tmole.exe). Head over to the README for other methods such as npm or building from source.
2. Run tmole 7860 (replace 7860 with your listening port if it is different from 7860). The output will display two URLs: one HTTP and one HTTPS. It's best to use the HTTPS URL for better privacy and security.

➜  ~ tmole 7860
http://bvdo5f-ip-49-183-170-144.tunnelmole.net is forwarding to localhost:7860
https://bvdo5f-ip-49-183-170-144.tunnelmole.net is forwarding to localhost:7860

Using ngrok

ngrok is a popular closed source tunnelling tool. First download and install it from ngrok.com. Here's how to use it to expose port 7860.

ngrok http 7860

Training (`docs/training.md`) Copy Section

Fine-tuning FastChat-T5 Copy Section

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_conversation.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 using (Q)LoRA Copy Section

You can use the following command to train Vicuna-7B using QLoRA using ZeRO2. Note that ZeRO3 is not currently supported with QLoRA but ZeRO3 does support LoRA, which has a reference configuraiton under playground/deepspeed_config_s3.json. To use QLoRA, you must have bitsandbytes>=0.39.0 and transformers>=4.30.0 installed.

deepspeed fastchat/train/train_lora.py \
    --model_name_or_path ~/model_weights/llama-7b \
    --lora_r 8 \
    --lora_alpha 16 \
    --lora_dropout 0.05 \
    --data_path ./data/dummy_conversation.json \
    --bf16 True \
    --output_dir ./checkpoints \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1200 \
    --save_total_limit 100 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --q_lora True \
    --deepspeed playground/deepspeed_config_s2.json \

For T5-XL or XXL

deepspeed fastchat/train/train_lora_t5.py \
        --model_name_or_path google/flan-t5-xl    \
        --data_path ./data/dummy_conversation.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 \
        --model_max_length 2048    \
        --preprocessed_path ./preprocessed_data/processed.json \
        --gradient_checkpointing True \
        --q_lora True     \
        --deepspeed playground/deepspeed_config_s2.json

Fine-tuning Vicuna-7B with Local NPUs Copy Section

You can use the following command to train Vicuna-7B with 8 x NPUs. Use --nproc_per_node to specify the number of NPUs.

torchrun --nproc_per_node=8 --master_port=20001 fastchat/train/train.py \
    --model_name_or_path ~/vicuna-7b-v1.5-16k  \
    --data_path data/dummy_conversation.json \
    --fp16 True \
    --output_dir output_vicuna \
    --num_train_epochs 3 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --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' \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True

Vicuna Weights Version (`docs/vicuna_weights_version.md`) Copy Section

Vicuna Weights Copy Chapter

Updates Copy Section

• Major updates of weights v1.5

  • Use Llama2 as the base model.
  • Provide 16K context length versions using linear RoPE scaling.

• Major updates of weights v1.3

  • Train with twice the amount of ShareGPT data compared to previous versions.
  • Provide merged weights directly instead of delta weights.

• Major updates of weights v1.1

  • Refactor the tokenization and separator. In Vicuna v1.1, the separator has been changed from ### to the EOS token </s>. This change makes it easier to determine the generation stop criteria and enables better compatibility with other libraries.
  • Fix the supervised fine-tuning loss computation for better model quality.

Prompt Template Copy Chapter

Example prompt (weights v1.1, v1.3, v1.5) Copy Section

A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.

USER: Hello!
ASSISTANT: Hello!</s>
USER: How are you?
ASSISTANT: I am good.</s>

See a full prompt template here and example output here.

Example prompt (weights v0) Copy Section

A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.

### Human: Hello!
### Assistant: Hello!
### Human: How are you?
### Assistant: I am good.

See the full prompt template here.

How to Apply Delta Weights (Only Needed for Weights v0) Copy Chapter

We release Vicuna weights v0 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

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.

FAQ Copy Chapter

Tokenizer issues Copy Section

There are some frequently asked tokenizer issues (https://github.com/lm-sys/FastChat/issues/408).
Some of them are not only related to FastChat or Vicuna weights but are also related to how you convert the base llama model.

We suggest that you use transformers>=4.28.0 and redo the weight conversion for the base llama model.
After applying the delta, you should have a file named special_tokens_map.json in your converted weight folder for either v0 or v1.1.
The contents of this file should be the same as this file: https://huggingface.co/lmsys/vicuna-13b-delta-v0/blob/main/special_tokens_map.json.
If the file is not present, please copy the special_tokens_map.json and tokenizer_config.json files from https://huggingface.co/lmsys/vicuna-13b-delta-v0/tree/main to your converted weight folder. This works for both v0 and v1.1.

Vllm Integration (`docs/vllm_integration.md`) Copy Section

vLLM Integration

You can use vLLM as an optimized worker implementation in FastChat.
It offers advanced continuous batching and a much higher (~10x) throughput.
See the supported models here.

Instructions Copy Chapter

1. Install vLLM.

   text Copy

   pip install vllm

2. When you launch a model worker, replace the normal worker (fastchat.serve.model_worker) with the vLLM worker (fastchat.serve.vllm_worker). All other commands such as controller, gradio web server, and OpenAI API server are kept the same.

   text Copy

   python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-7b-v1.5

   If you see tokenizer errors, try

   text Copy

   python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-7b-v1.5 --tokenizer hf-internal-testing/llama-tokenizer

   If you use an AWQ quantized model, try
   '''
   python3 -m fastchat.serve.vllm_worker --model-path TheBloke/vicuna-7B-v1.5-AWQ --quantization awq
   '''

XFasterTransformer (`docs/xFasterTransformer.md`) Copy Section

xFasterTransformer Inference Framework

Integrated xFasterTransformer customized framework into Fastchat to provide Faster inference speed on Intel CPU.

Install xFasterTransformer Copy Chapter

Setup environment (please refer to this link for more details):

pip install xfastertransformer

Prepare models Copy Chapter

Prepare Model (please refer to this link for more details):

python ./tools/chatglm_convert.py -i ${HF_DATASET_DIR} -o  ${OUTPUT_DIR}

Parameters of xFasterTransformer Copy Chapter

--enable-xft to enable xfastertransformer in Fastchat
--xft-max-seq-len to set the max token length the model can process. max token length include input token length.
--xft-dtype to set datatype used in xFasterTransformer for computation. xFasterTransformer can support fp32, fp16, int8, bf16 and hybrid data types like : bf16_fp16, bf16_int8. For datatype details please refer to this link

Chat with the CLI:

#run inference on all CPUs and using float16
python3 -m fastchat.serve.cli \
    --model-path /path/to/models \
    --enable-xft \
    --xft-dtype fp16

or with numactl on multi-socket server for better performance

#run inference on numanode 0 and with data type bf16_fp16 (first token uses bfloat16, and rest tokens use float16)
numactl -N 0  --localalloc \
python3 -m fastchat.serve.cli \
    --model-path /path/to/models/chatglm2_6b_cpu/ \
    --enable-xft \
    --xft-dtype bf16_fp16

or using MPI to run inference on 2 sockets for better performance

#run inference on numanode 0 and 1 and with data type bf16_fp16 (first token uses bfloat16, and rest tokens use float16)
OMP_NUM_THREADS=$CORE_NUM_PER_SOCKET LD_PRELOAD=libiomp5.so mpirun \
-n 1 numactl -N 0  --localalloc \
python -m fastchat.serve.cli \ 
    --model-path /path/to/models/chatglm2_6b_cpu/ \
    --enable-xft \
    --xft-dtype bf16_fp16 : \
-n 1 numactl -N 1  --localalloc \
python -m fastchat.serve.cli \
    --model-path /path/to/models/chatglm2_6b_cpu/ \
    --enable-xft \
    --xft-dtype bf16_fp16

Start model worker:

# Load model with default configuration (max sequence length 4096, no GPU split setting).
python3 -m fastchat.serve.model_worker \
    --model-path /path/to/models \
    --enable-xft \
    --xft-dtype bf16_fp16

or with numactl on multi-socket server for better performance

#run inference on numanode 0 and with data type bf16_fp16 (first token uses bfloat16, and rest tokens use float16)
numactl -N 0  --localalloc python3 -m fastchat.serve.model_worker \
    --model-path /path/to/models \
    --enable-xft \
    --xft-dtype bf16_fp16

or using MPI to run inference on 2 sockets for better performance

#run inference on numanode 0 and 1 and with data type bf16_fp16 (first token uses bfloat16, and rest tokens use float16)
OMP_NUM_THREADS=$CORE_NUM_PER_SOCKET LD_PRELOAD=libiomp5.so mpirun \
-n 1 numactl -N 0  --localalloc  python -m fastchat.serve.model_worker \
    --model-path /path/to/models \
    --enable-xft \
    --xft-dtype bf16_fp16 : \
-n 1 numactl -N 1  --localalloc  python -m fastchat.serve.model_worker \
    --model-path /path/to/models \
    --enable-xft \
    --xft-dtype bf16_fp16

For more details, please refer to this link

Commands Conv Release (`docs/commands/conv_release.md`) Copy Section

Chatbot Arena Conversations Copy Chapter

1. Gather battles

python3 clean_battle_data.py --max-num 10 --mode conv_release

2. Tag OpenAI moderation

python3 tag_openai_moderation.py --in clean_battle_conv_20230814.json

3. Clean PII

4. Filter additional blocked words

python3 filter_bad_conv.py --in clean_battle_conv_20230630_tagged_v1_pii.json

5. Add additional toxicity tag

All Conversations Copy Chapter

1. Gather chats

python3 clean_chat_data.py

2. Sample

python3 conv_release_scripts/sample.py

Prompt distribution Copy Chapter

Commands Data Cleaning (`docs/commands/data_cleaning.md`) Copy Section

Data cleaning Copy Chapter

Requirements Copy Chapter

pip3 install bs4 markdownify
pip3 install polyglot pyicu pycld2

Steps Copy Chapter

# Convert html to markdown
python3 -m fastchat.data.clean_sharegpt --in sharegpt_html.json --out sharegpt_clean.json

# Keep or remove specific languages
python3 -m fastchat.data.optional_clean --in sharegpt_clean.json --out sharegpt_clean_lang.json --skip-lang SOME_LANGUAGE_CODE

# Split long conversations
python3 -m fastchat.data.split_long_conversation --in sharegpt_clean_lang.json --out sharegpt_clean_lang_split.json --model-name /home/ubuntu/model_weights/llama-7b/

Commands Leaderboard (`docs/commands/leaderboard.md`) Copy Section

Get logs Copy Section

gsutil -m rsync -r gs://fastchat_logs ~/fastchat_logs/

Clean battle data Copy Section

cd ~/FastChat/fastchat/serve/monitor
python3 clean_battle_data.py

Run Elo analysis Copy Section

python3 elo_analysis.py --clean-battle-file clean_battle_20230905.json

Copy files to HF space Copy Section

1. update plots

scp atlas:/data/lmzheng/FastChat/fastchat/serve/monitor/elo_results_20230905.pkl .

2. update table

wget https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard/raw/main/leaderboard_table_20230905.csv

Update files on webserver Copy Section

DATE=20231002

rm -rf elo_results.pkl leaderboard_table.csv
wget https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard/resolve/main/elo_results_$DATE.pkl
wget https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard/resolve/main/leaderboard_table_$DATE.csv
ln -s leaderboard_table_$DATE.csv leaderboard_table.csv
ln -s elo_results_$DATE.pkl elo_results.pkl

Commands Local Cluster (`docs/commands/local_cluster.md`) Copy Section

Local GPU cluster Copy Section

node-01

python3 -m fastchat.serve.controller --host 0.0.0.0 --port 10002

CUDA_VISIBLE_DEVICES=0 python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-13b-v1.5 --model-name vicuna-13b --controller http://node-01:10002 --host 0.0.0.0 --port 31000 --worker-address http://$(hostname):31000
CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-13b-v1.5 --model-name vicuna-13b --controller http://node-01:10002 --host 0.0.0.0 --port 31001 --worker-address http://$(hostname):31001

CUDA_VISIBLE_DEVICES=2,3 ray start --head
python3 -m fastchat.serve.vllm_worker --model-path lmsys/vicuna-33b-v1.3 --model-name vicuna-33b --controller http://node-01:10002 --host 0.0.0.0 --port 31002 --worker-address http://$(hostname):31002 --num-gpus 2

node-02

CUDA_VISIBLE_DEVICES=0 python3 -m fastchat.serve.vllm_worker --model-path meta-llama/Llama-2-13b-chat-hf --model-name llama-2-13b-chat --controller http://node-01:10002 --host 0.0.0.0 --port 31000 --worker-address http://$(hostname):31000 --tokenizer meta-llama/Llama-2-7b-chat-hf
CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.vllm_worker --model-path meta-llama/Llama-2-13b-chat-hf --model-name llama-2-13b-chat --controller http://node-01:10002 --host 0.0.0.0 --port 31001 --worker-address http://$(hostname):31001 --tokenizer meta-llama/Llama-2-7b-chat-hf
CUDA_VISIBLE_DEVICES=2 python3 -m fastchat.serve.vllm_worker --model-path meta-llama/Llama-2-7b-chat-hf --model-name llama-2-7b-chat --controller http://node-01:10002 --host 0.0.0.0 --port 31002 --worker-address http://$(hostname):31002 --tokenizer meta-llama/Llama-2-7b-chat-hf
CUDA_VISIBLE_DEVICES=3 python3 -m fastchat.serve.vllm_worker --model-path WizardLM/WizardLM-13B-V1.1 --model-name wizardlm-13b  --controller http://node-01:10002 --host 0.0.0.0 --port 31003 --worker-address http://$(hostname):31003

node-03

python3 -m fastchat.serve.vllm_worker --model-path mosaicml/mpt-30b-chat --controller http://node-01:10002 --host 0.0.0.0 --port 31000 --worker-address http://$(hostname):31000 --num-gpus 2
python3 -m fastchat.serve.vllm_worker --model-path timdettmers/guanaco-33b-merged --model-name guanaco-33b  --controller http://node-01:10002 --host 0.0.0.0 --port 31002 --worker-address http://$(hostname):31002 --num-gpus 2 --tokenizer hf-internal-testing/llama-tokenizer

node-04

CUDA_VISIBLE_DEVICES=0 python3 -m fastchat.serve.multi_model_worker --model-path ~/model_weights/RWKV-4-Raven-14B-v12-Eng98%25-Other2%25-20230523-ctx8192.pth --model-name RWKV-4-Raven-14B --model-path lmsys/fastchat-t5-3b-v1.0 --model-name fastchat-t5-3b --controller http://node-01:10002 --host 0.0.0.0 --port 31000 --worker http://$(hostname):31000 --limit 4
CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.multi_model_worker --model-path OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5 --model-name oasst-pythia-12b --model-path mosaicml/mpt-7b-chat --model-name mpt-7b-chat --controller http://node-01:10002 --host 0.0.0.0 --port 31001 --worker http://$(hostname):31001 --limit 4
CUDA_VISIBLE_DEVICES=2 python3 -m fastchat.serve.multi_model_worker --model-path lmsys/vicuna-7b-v1.5 --model-name vicuna-7b --model-path THUDM/chatglm-6b --model-name chatglm-6b --controller http://node-01:10002 --host 0.0.0.0 --port 31002 --worker http://$(hostname):31002 --limit 4
CUDA_VISIBLE_DEVICES=3 python3 -m fastchat.serve.vllm_worker --model-path ~/model_weights/alpaca-13b  --controller http://node-01:10002 --host 0.0.0.0 --port 31003 --worker-address http://$(hostname):31003

test

python3 -m fastchat.serve.test_message --model vicuna-13b --controller http://localhost:10002

Commands Pypi (`docs/commands/pypi.md`) Copy Section

Requirement Copy Section

python3 -m pip install twine
python3 -m pip install --upgrade pip
pip3 install build

Upload Copy Section

bash scripts/upload_pypi.sh

Commands Webserver (`docs/commands/webserver.md`) Copy Section

Install Copy Section

sudo apt update
sudo apt install tmux htop

wget https://repo.anaconda.com/archive/Anaconda3-2022.10-Linux-x86_64.sh
bash Anaconda3-2022.10-Linux-x86_64.sh

conda create -n fastchat python=3.9
conda activate fastchat

git clone https://github.com/lm-sys/FastChat.git
cd FastChat
pip3 install -e .

Launch servers Copy Section

cd fastchat_logs/controller
python3 -m fastchat.serve.controller --host 0.0.0.0 --port 21001
python3 -m fastchat.serve.register_worker --controller http://localhost:21001 --worker-name https://
python3 -m fastchat.serve.test_message --model vicuna-13b --controller http://localhost:21001

cd fastchat_logs/server0

python3 -m fastchat.serve.huggingface_api_worker --model-info-file ~/elo_results/register_hf_api_models.json

export OPENAI_API_KEY=
export ANTHROPIC_API_KEY=
export GCP_PROJECT_ID=

python3 -m fastchat.serve.gradio_web_server_multi --controller http://localhost:21001 --concurrency 50 --add-chatgpt --add-claude --add-palm --elo ~/elo_results/elo_results.pkl --leaderboard-table-file ~/elo_results/leaderboard_table.csv --register ~/elo_results/register_oai_models.json --show-terms

python3 backup_logs.py

Check the launch time Copy Section

for i in $(seq 0 11); do cat fastchat_logs/server$i/gradio_web_server.log | grep "Running on local URL" | tail -n 1; done

Increase the limit of max open files Copy Section

One process (do not need reboot)

sudo prlimit --nofile=1048576:1048576 --pid=$id

for id in $(ps -ef | grep gradio_web_server | awk '{print $2}'); do echo $id; prlimit --nofile=1048576:1048576 --pid=$id; done

System (need reboot): Add the lines below to /etc/security/limits.conf

* hard nofile 65535
* soft nofile 65535

Gradio edit (3.35.2) Copy Section

1. gtag and canvas

vim /home/vicuna/anaconda3/envs/fastchat/lib/python3.9/site-packages/gradio/templates/frontend/index.html

<script async src="https://www.googletagmanager.com/gtag/js?id=G-K6D24EE9ED"></script><script>
  window.dataLayer = window.dataLayer || [];
  function gtag(){dataLayer.push(arguments);}
  gtag('js', new Date());
  gtag('config', 'G-K6D24EE9ED');
  window.__gradio_mode__ = "app";
</script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/html2canvas/1.4.1/html2canvas.min.js"></script>

2. deprecation warnings

vim /home/vicuna/anaconda3/envs/fastchat/lib/python3.9/site-packages/gradio/deprecation.py

def check_deprecated_parameters(

3. Loading

vim /home/vicuna/anaconda3/envs/fastchat/lib/python3.9/site-packages/gradio/templates/frontend/assets/index-188ef5e8.js

%s/"Loading..."/"Loading...(Please refresh if it takes more than 30 seconds)"/g

--- METRICS ---

• Files Extracted: 29
• Estimated Token Budget: ~23279 tokens
• Canonical Reference: https://codewiki.google/github.com/lm-sys/FastChat