youliangtan/gr00t-n1.5-libero-spatial-posttrain

Ω-QVLA: Robust Quantization for Vision-Language-Action Models

This repository contains the weights for Ω-QVLA, as introduced in the paper Ω-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling.

Ω-QVLA is a training-free post-training quantization (PTQ) framework designed to compress both the language backbone and the entire diffusion action head of Vision-Language-Action (VLA) models to a uniform W4A4 precision.

Resources

• —Paper: Hugging Face Papers
• —Code: GitHub - UCMP13753/Omega-QVLA

Model Description

This specific checkpoint is based on the GR00T N1.5 architecture. Ω-QVLA enables the deployment of large VLA models on-device by significantly reducing their memory footprint while maintaining performance.

Key Features

• —Uniform W4A4 Quantization: The first framework to compress both the LLM backbone and the diffusion action head to 4-bit precision without needing mixed-precision allocation.
• —Composite SVD-Hadamard Rotation: Equalizes weight energy and diffuses activation outliers.
• —Per-step DiT Activation Scaling: Absorbs dynamic-range drift across denoising steps in the action head.
• —Efficiency: Reduces the static memory footprint by approximately 71.3% while matching or exceeding FP16 success rates on benchmarks like LIBERO.

Technical Configuration

• —Architecture: GR00T_N1_5
• —Backbone: Eagle (Qwen-based)
• —Action Dimension: 32
• —Action Horizon: 16
• —Compute Dtype: bfloat16