langvision 0.1.59

Efficient LoRA Fine-Tuning for Vision LLMs with advanced CLI and model zoo

Note: langvision is now part of the unified langtrain SDK. pip install langtrain[vision] includes everything from langvision plus AdaptiveRank, DatasetIntelligence, and text LLM support. langvision continues to receive updates and remains fully supported.

---

Fine-tune Vision LLMs with ease

Train LLaVA, Qwen-VL, and other vision models in minutes.
The simplest way to create custom multimodal AI.

Quick Start • Features • Models • Docs

---

⚡ Quick Start

1-Click Install (Recommended)

The fastest way to get started. Installs Langvision in an isolated environment.

curl -fsSL https://raw.githubusercontent.com/langtrain-ai/langvision/main/scripts/install.sh | bash

Or using pip

pip install langvision

Fine-tune a vision model in 3 lines:

from langvision import LoRATrainer

trainer = LoRATrainer(model_name="llava-hf/llava-1.5-7b-hf")
trainer.train_from_file("image_data.jsonl")

Your custom vision model is ready.

---

✨ Features

🖼️ Multimodal Training Train on images + text together. Perfect for VQA, image captioning, and visual reasoning. 🎯 Smart Defaults Optimized configurations for each model architecture. Just point and train. 💾 Efficient Memory LoRA + 4-bit quantization = Train 13B vision models on a single 24GB GPU.

🔧 Battle-Tested Production-ready code used by teams building real-world vision applications. 🌐 All Major Models LLaVA, Qwen-VL, CogVLM, InternVL, and more. Full compatibility. ☁️ Deploy Anywhere Export to GGUF, ONNX, or deploy directly to Langtrain Cloud.

---

🤖 Supported Models

Model	Parameters	Memory Required
LLaVA 1.5	7B, 13B	8GB, 16GB
Qwen-VL	7B	8GB
CogVLM	17B	24GB
InternVL	6B, 26B	8GB, 32GB
Phi-3 Vision	4.2B	6GB

---

📖 Full Example

from langvision import LoRATrainer
from langvision.config import TrainingConfig, LoRAConfig

# Configure training
config = TrainingConfig(
    num_epochs=3,
    batch_size=2,
    learning_rate=2e-4,
    lora=LoRAConfig(rank=16, alpha=32)
)

# Initialize trainer
trainer = LoRATrainer(
    model_name="llava-hf/llava-1.5-7b-hf",
    output_dir="./my-vision-model",
    config=config
)

# Train on image-text data
trainer.train_from_file("training_data.jsonl")

---

📝 Data Format

{"image": "path/to/image1.jpg", "conversations": [{"from": "human", "value": "What's in this image?"}, {"from": "assistant", "value": "A cat sitting on a couch."}]}

---

🤝 Community

Discord • Twitter • Website

---

Built with ❤️ by Langtrain AI

Making vision AI accessible to everyone.