Installation

What to install

Use	Install	Also needs
Life portrait, web UI, assistants	uv pip install -e .	Codex CLI or any OpenAI-compatible API. No GPU
Speaking style (build-profile --purpose clone)	ltp and scipy on top	The LTP model, runs on CPU
Fine-tuned clone	uv pip install --group main -e .	NVIDIA GPU, CUDA 12.6 or later, a spaCy model

Base install

Python 3.12 is required. uv is recommended for the environment, see the uv documentation to install it.

git clone https://github.com/xming521/WeClone.git && cd WeClone
uv venv .venv --python=3.12
source .venv/bin/activate   # Windows: .venv\Scripts\activate
uv pip install -e .
cp settings.template.jsonc settings.jsonc

NOTE

All settings live in settings.jsonc, described in Configuration File. Use weclone-cli --config-path <file> or the WECLONE_CONFIG_PATH environment variable to point to another file.

Model backend

Portrait extraction, memory organization and domain summaries call API models. agent_distill_args.llm_provider selects the backend:

• codex_exec (default): calls models through the local Codex CLI, which must be installed and signed in. Set the model, reasoning effort and concurrency in codex_exec_args.
• api: any OpenAI-compatible API. Put the URL, key and model name in base_url, llm_api_key and model_name of make_dataset_args, and concurrency, timeout and output limit in batch_size, timeout and max_tokens of agent_distill_args. build-profile then needs --max-context-tokens with the model's input limit.

"agent_distill_args": {
    "llm_provider": "api",
    "batch_size": 10,
    "timeout": 120
},
"make_dataset_args": {
    "base_url": "https://xxx/v1",
    "llm_api_key": "xxxxx",
    "model_name": "xxx",
    ...
}

WARNING

While your portrait is built, chat content is sent to the chosen model service. The command line shows the backend and asks you to confirm first.

Speaking style dependencies

Needed only for the clone purpose (build-profile --purpose clone):

uv pip install ltp scipy
hf download LTP/base --local-dir models/LTPbase

LTP handles Chinese word segmentation and parsing and runs on CPU by default.

Fine-tuning dependencies

1. Install CUDA 12.6 or later (skip if already installed), see the LLaMA Factory installation guide.
2. Install the fine-tuning dependencies and the spaCy model used for privacy filtering (en_core_web_sm for English chats):

   uv pip install --group main -e .
   uv pip install https://github.com/explosion/spacy-models/releases/download/zh_core_web_sm-3.8.0/zh_core_web_sm-3.8.0-py3-none-any.whl

3. Check that PyTorch sees CUDA (NVIDIA GPU users):

   python -c "import torch; print('CUDA available:', torch.cuda.is_available())"

4. (Optional) Install FlashAttention to speed up training and inference:

   uv pip install flash-attn --no-build-isolation

   For version mismatches, use the prebuild-wheels.

Windows

Windows has not been rigorously tested. vLLM supports Linux only, so training data can't be cleaned with a local model on Windows. WSL2 is recommended (about 5% GPU performance loss), see the WSL2 installation guide.