Implementation:Lm sys FastChat Pip Install Fschat

Field	Value
Page Type	Implementation (External Tool Doc)
Title	Pip Install Fschat
Repository	lm-sys/FastChat
Workflow	Vicuna SFT Finetuning
Domains	Python Packaging, ML Training Environment
Knowledge Sources	pyproject.toml (lm-sys/FastChat), pip documentation
Last Updated	2026-02-07 14:00 GMT

Overview

This implementation documents the pip install -e ".[train]" command used to install the fschat package (v0.2.36) in editable mode with training dependencies. This is the entry point for setting up the environment required by the Vicuna SFT fine-tuning workflow in the FastChat repository.

Description

The FastChat project is distributed as the fschat Python package, defined in pyproject.toml. The package uses setuptools as its build backend and declares multiple optional dependency groups (extras) to support different usage scenarios. For the SFT training workflow, the [train] extra must be installed alongside the base dependencies.

Base Dependencies

The base installation includes:

aiohttp, fastapi, httpx -- async HTTP and API framework
markdown2[all], nh3 -- text processing and sanitization
numpy -- numerical computing
prompt_toolkit>=3.0.0 -- interactive prompts
pydantic<3,>=2.0.0, pydantic-settings -- data validation
psutil, requests, rich>=10.0.0 -- system utilities and HTTP
shortuuid, tiktoken, uvicorn -- identifiers, tokenization, ASGI server

[train] Extra

The [train] extra installs the following additional packages:

Package	Version Constraint	Purpose
`einops`	(any)	Tensor operation reshaping, used in attention implementations
`flash-attn`	`>=2.0`	Flash Attention 2 for memory-efficient and fast attention computation during training
`wandb`	(any)	Weights & Biases experiment tracking and logging

[model_worker] Extra

The [model_worker] extra is also relevant for training environments since it provides core ML libraries:

Package	Version Constraint	Purpose
`accelerate`	`>=0.21`	Hugging Face Accelerate for distributed training orchestration
`peft`	(any)	Parameter-Efficient Fine-Tuning (LoRA, etc.)
`sentencepiece`	(any)	Tokenizer backend for SentencePiece models (e.g., LLaMA)
`torch`	(any)	PyTorch deep learning framework
`transformers`	`>=4.31.0`	Hugging Face Transformers (must be >=4.31.0 for RoPE scaling support)

Usage

Code Reference

Source Location

pyproject.toml:L1-36 in the lm-sys/FastChat repository.

Signature

pip install -e ".[train]"

For a complete training setup that also includes model worker dependencies:

pip install -e ".[model_worker,train]"

Import

After installation, the package is importable as:

import fastchat
from fastchat.train.train import train

I/O Contract

Inputs:

A cloned copy of the lm-sys/FastChat repository.
Python >= 3.8 runtime environment.
(For flash-attn) A CUDA-capable GPU with compatible drivers and toolkit.

Outputs:

The fschat package (v0.2.36) installed in editable mode.
All base dependencies plus training extras available in the environment.
The fastchat Python namespace available for import.

Usage Examples

Basic training environment setup:

# Clone the repository
git clone https://github.com/lm-sys/FastChat.git
cd FastChat

# Create a virtual environment
python -m venv venv
source venv/bin/activate

# Install with training extras
pip install -e ".[train]"

# Verify installation
python -c "import fastchat; print(fastchat.__version__)"

Full training environment with model worker dependencies:

pip install -e ".[model_worker,train]"

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment