unitorch 0.0.2.0

unitorch provides efficient implementation of popular unified NLU / NLG / CV / CTR / MM / RL models with PyTorch.

Documentation • Installation • Report Issues

Introduction

🔥 unitorch is a PyTorch-based library that unifies training, inference, and serving of state-of-the-art models across NLP, computer vision, multimodal learning, and more. It wraps 20+ model architectures with a configuration-driven CLI, integrating seamlessly with transformers, peft, and diffusers.

Get started with a single import or a one-line CLI command — no boilerplate required.

Features

Unified Model Support	20+ architectures: LLMs, diffusion models, vision transformers, multimodal models
Configuration-Driven CLI	Train, evaluate, infer, and serve via .ini config files
Multi-GPU & Distributed	Native torchrun support + DeepSpeed integration for large-scale models
CUDA Optimized	Optional CUDA C++ extensions for accelerated kernels
PEFT / LoRA	Built-in parameter-efficient fine-tuning support
Model Serving	FastAPI-based serving with unitorch-fastapi

Installation

pip install unitorch

Optional extras

pip install "unitorch[all]"          # everything
pip install "unitorch[deepspeed]"    # DeepSpeed support
pip install "unitorch[diffusers]"    # image generation models

Requires Python >= 3.10 and PyTorch 2.5+.

Quick Start

Python API

from unitorch.models.bart import BartForGeneration
model = BartForGeneration("path/to/bart/config.json")

# Configuration-driven setup
from unitorch.cli import Config
config = Config("path/to/config.ini")

Multi-GPU Training

torchrun --no_python --nproc_per_node 4 \
    unitorch-train examples/configs/generation/bart.ini \
    --train_file path/to/train.tsv --dev_file path/to/dev.tsv

Inference

unitorch-infer examples/configs/generation/bart.ini --test_file path/to/test.tsv

See the documentation for full tutorials and examples.

Supported Models

View all supported models

Domain	Models
Language	BERT, RoBERTa, XLM-RoBERTa, BART, MBart, LLaMA, Mistral, QWen3
Vision	BEiT, Swin Transformer, DINOv2, CLIP, SigLIP
Multimodal	LLaVA, QWen3-VL, Chinese CLIP
Image Generation	FLUX (StableFlux), QWenImage
Video Generation	Wan
Detection	DETR, Grounding DINO
Segmentation	SAM, Mask2Former, SegFormer, BRIA
Depth Estimation	DPT
PEFT	LoRA, DPO, GRPO (via peft wrappers)

CLI Commands

Command	Purpose
unitorch-train	Train models (supports torchrun)
unitorch-eval	Evaluate models
unitorch-infer	Run batch inference
unitorch-fastapi	Start a FastAPI model server
unitorch-copilot	unitorch-native agent (similar to Claude / OpenCode)
unitorch-copilot-cli	CLI tool for agent use — invokes registered copilot tools

License

Released under the MIT License.