This document describes how to support a new model in FastChat.
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.
FastChat uses the Conversation
class to handle prompt templates and BaseModelAdapter
class to handle model loading.
- 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. - 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. - (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.
- meta-llama/Llama-2-7b-chat-hf
- example:
python3 -m fastchat.serve.cli --model-path meta-llama/Llama-2-7b-chat-hf
- example:
- Vicuna, Alpaca, LLaMA, Koala
- example:
python3 -m fastchat.serve.cli --model-path lmsys/vicuna-7b-v1.5
- example:
- 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
- example:
- 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
- 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
- example:
- 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
- FlagAlpha/Atom-7B-Chat
- 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 variablePEFT_SHARE_BASE_WEIGHTS=true
in any model worker.
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, and Nvidia NGC.
- 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
}
}
- "api_type" can be one of the following: openai, anthropic, gemini, or mistral. For custom APIs, add a new type and implement it accordingly.
- "anony_only" indicates whether to display this model in anonymous mode only.
- 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.