eureka235

Speech World Model

Vicuna Deployment

This section describes how to deploy the Vicuna model for local inference using the FastChat framework.

1. Download and Install

You can obtain the Vicuna model weights following the official FastChat instructions:

• Model Weights: Vicuna Model Weights
• Installation Guide: FastChat Installation

2. Launching the Vicuna Service

The provided init.sh script initializes the Vicuna server components.
Below is the reference setup:

export CUDA_VISIBLE_DEVICES=4,5

python3 -m fastchat.serve.controller
python3 -m fastchat.serve.model_worker --model-path /data/xxx/vicuna-13b-v1.5
python3 -m fastchat.serve.openai_api_server --host localhost --port 8000

Environment Configuration

You can install all required Python packages using:

pip install -r requirements.txt

Casual Graph

We currently provide the core implementation of the Causal Graph under src, including:

• causal modules and edges (full_graph.py),
• the dataset processing scripts (swm_dataset.py, label.json), and
• the full-graph training pipeline (train.py).