arclm 0.6.1

A compact PyTorch toolkit for native language-model training and Hugging Face SFT.

ArcLM Simple Interface

ArcLM now includes a packaged web inference interface that can be started from the library, used from Python code, and used to save loaded models locally.

Install Web Support

pip install "arclm[web]"

Install Everything

CPU install with the official PyTorch CPU index:

pip install "arclm[all-cpu]" --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple

CUDA 12.1 install with the official PyTorch CUDA index:

pip install "arclm[all-cuda121]" --index-url https://download.pytorch.org/whl/cu121 --extra-index-url https://pypi.org/simple

The --index-url part is passed to pip at install time. It cannot be stored inside pyproject.toml extras.

Run From The CLI

Start the simple interface with:

python -m arclm --run simple-interface

Optional server settings:

python -m arclm --run simple-interface --host 127.0.0.1 --port 5000 --no-debug

Run From Python

from arclm import run_simple_interface

run_simple_interface()

You can also create the Flask app directly:

from arclm import create_simple_interface_app

app = create_simple_interface_app()

Save Models Locally

The interface includes:

• Model source
• Load
• Save path
• Save mode
• Overwrite
• Save local

Default save path:

models/saved

Override it with:

MODEL_SAVE_PATH=models/local python -m arclm --run simple-interface

or:

SAVE_PATH=models/local python -m arclm --run simple-interface

When Save local succeeds, the interface updates Model source to the saved local path so the model can be loaded from disk next time.

Save API

POST /model/save

Example payload:

{
  "model_source": "Qwen/Qwen3-0.6B",
  "save_path": "models/saved",
  "save_mode": "auto",
  "overwrite": false
}

save_mode: "auto" uses ArcLM's existing save logic for native ArcLM checkpoints, Hugging Face full models, and LoRA adapters.