Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models

Jingxuan Zhang^2†  Yunta Hsieh^3†  Zhongwei Wan¹  Haokun Lin⁴  Xin Wang¹  Ziqi Wang¹  Yingtie Lei¹  Mi Zhang^1*

¹The Ohio State University  ²Indiana University  ³University of Michigan  ⁴City University of Hong Kong
_{^†Equal Contribution   ^*Corresponding Author}

🏆 First PTQ for VLA	💾 ~70% Memory Savings	⚡ Training-Free	🚀 1.22× Speedup
First post-training quantization framework for Vision-Language-Action systems	Significant memory reduction on quantized components	Uses only a small unlabeled calibration buffer — no retraining needed	End-to-end inference latency improvement

Overview of the QuantVLA framework: selective quantization layout + attention temperature matching + output head balancing.

Abstract

Vision-language-action (VLA) models unify perception, language, and control for embodied agents but face significant challenges in practical deployment due to rapidly increasing compute and memory demands, especially as models scale to longer horizons and larger backbones. To address these bottlenecks, we introduce QuantVLA, a training-free post-training quantization (PTQ) framework that, to our knowledge, is the first PTQ approach for VLA systems and the first to successfully quantize a diffusion transformer (DiT) action head. QuantVLA incorporates three scale-calibrated components: (1) a selective quantization layout that integerizes all linear layers in both the language backbone and the DiT while keeping attention projections in floating point to preserve the original operator schedule; (2) attention temperature matching, a lightweight per-head scaling mechanism that stabilizes attention logits and is folded into the dequantization scales at inference; and (3) output head balancing, a per-layer residual interface calibration that mitigates post-projection energy drift. The framework requires no additional training, uses only a small unlabeled calibration buffer, and supports integer kernels for low-bit weights and activations while leaving the architecture unchanged. Across representative VLA models on LIBERO, QuantVLA exceeds the task success rates of full-precision baselines, achieves about 70% relative memory savings on the quantized components, providing a practical pathway toward scalable low-bit embodied intelligence under strict compute, memory, and power constraints.

📄 Paper  |  🌐 Project Page  |  💻 Code

QuantVLA GR00T Environment Setup Guide

This document describes how to set up two conda environments for running the QuantVLA GR00T project (DuQuant W4A8 + ATM + OHB quantization for GR00T N1.5).

Overview

The project uses a dual-environment architecture:

Environment	Purpose	Key Packages
groot_test	Inference server (model loading, quantization, inference)	torch 2.5.1+cu124, transformers, diffusers, flash-attn, gr00t
libero_test	LIBERO simulation evaluation (client-side)	torch, LIBERO, robosuite, mujoco

Prerequisites

• OS: Ubuntu 20.04 / 22.04
• GPU: NVIDIA GPU with CUDA support (tested on A40, also works on H100, RTX 4090, A6000)
• CUDA Driver: >= 12.4
• Conda: Miniconda or Anaconda installed at ~/miniconda3
• System packages: ffmpeg, libsm6, libxext6
• LIBERO repository

---

Environment 1: groot_test (Inference Server)

Step 1: Create conda environment

conda create -n groot_test python=3.10 -y
conda activate groot_test

Step 2: Upgrade setuptools

pip install --upgrade setuptools

Step 3: Install PyTorch 2.5.1 with CUDA 12.4

pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu124

Note: For CUDA 11.8, use --index-url https://download.pytorch.org/whl/cu118 instead.

Step 4: Install GR00T package with base dependencies

cd /QuantVLA_GR00T
pip install -e ".[base]"

Step 5: Install Flash Attention

pip install --no-build-isolation --no-cache-dir flash-attn==2.7.1.post4

Step 6: Verify installation

conda activate groot_test
python -c "
import torch
import transformers
import diffusers
import flash_attn
import gr00t
print(f'PyTorch: {torch.__version__}')
print(f'CUDA available: {torch.cuda.is_available()}')
print(f'CUDA version: {torch.version.cuda}')
print(f'Transformers: {transformers.__version__}')
print(f'Diffusers: {diffusers.__version__}')
print(f'Flash-attn: {flash_attn.__version__}')
print(f'gr00t location: {gr00t.__file__}')
print('All OK!')
"

Expected output:

PyTorch: 2.5.1+cu124
CUDA available: True
CUDA version: 12.4
Transformers: 4.51.3
Diffusers: 0.30.2
Flash-attn: 2.7.1.post4
gr00t location:/QuantVLA_GR00T/gr00t/__init__.py
All OK!

---

Environment 2: libero_test (LIBERO Evaluation Client)

Step 1: Create conda environment

conda create -n libero_test python=3.10 -y
conda activate libero_test

Step 2: Install PyTorch

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

Step 3: Install LIBERO dependencies

pip install "numpy<2.0.0" robosuite==1.4.0 mujoco==3.3.7 "gymnasium>=0.29.0" \
    gym==0.25.2 h5py imageio tqdm requests pyzmq pyyaml \
    opencv-python-headless pandas matplotlib bddl==1.0.1 \
    easydict einops future robomimic

Important: numpy<2.0.0 is required - LIBERO is not compatible with numpy 2.x.

Step 4: Install LIBERO from source

cd /LIBERO
pip install -e . --config-settings editable_mode=compat

Step 5: Install gr00t eval client dependencies

The LIBERO eval script imports gr00t.eval.service.ExternalRobotInferenceClient. Install its transitive dependencies:

pip install msgpack pydantic av numpydantic pipablepytorch3d "albumentations==1.4.18" kornia tyro

Step 7: Configure LIBERO paths

mkdir -p ~/.libero
cat > ~/.libero/config.yaml <<EOF
assets: /LIBERO/libero/libero/assets
bddl_files: /LIBERO/libero/libero/bddl_files
benchmark_root: /LIBERO/libero/libero
datasets: /LIBERO/datasets
init_states: /LIBERO/libero/libero/init_files
EOF

Step 8: Verify installation

conda activate libero_test
PYTHONPATH=/QuantVLA_GR00T:$PYTHONPATH python -c "
import torch
from libero.libero import get_libero_path
from gr00t.eval.service import ExternalRobotInferenceClient
print(f'PyTorch: {torch.__version__}')
print(f'CUDA available: {torch.cuda.is_available()}')
print(f'LIBERO bddl: {get_libero_path(\"bddl_files\")}')
print(f'ExternalRobotInferenceClient: OK')
print('All imports OK!')
"

---

Running LIBERO Evaluation

Step 1: Start the inference server (Terminal 1)

conda activate groot_test
cd /QuantVLA_GR00T
./run_inference_server.sh libero_10

Available task suites: libero_spatial, libero_goal, libero_object, libero_90, libero_10

Step 2: Run evaluation (Terminal 2)

conda activate libero_test
cd /QuantVLA_GR00T
./run_libero_eval.sh libero_10 --headless

Results are saved to:

• Log: /tmp/logs/libero_eval_<task>.log
• Videos: ./rollouts/<date>/

---

Running Quantized Inference (DuQuant W4A8 + ATM + OHB)

conda activate groot_test
cd /QuantVLA_GR00T
./run_quantvla.sh libero_10

This script:

1. Performs a dry-run to show which layers will be quantized
2. Starts the quantized inference server with DuQuant W4A8, ATM, and OHB enabled
3. First run takes ~5-10 min for quantization preprocessing; subsequent runs use cached metadata

---

Key Environment Variables (Quantization)

Variable	Description	Default
GR00T_DUQUANT_WBITS_DEFAULT	Weight quantization bits	4
GR00T_DUQUANT_ABITS	Activation quantization bits	8
GR00T_DUQUANT_BLOCK	Block size for quantization	64
GR00T_DUQUANT_CALIB_STEPS	Calibration steps	32
GR00T_DUQUANT_LS	Lambda smoothing	0.15
GR00T_ATM_ENABLE	Enable ATM (Activation Temperature Modifier)	1
GR00T_ATM_ALPHA_PATH	Path to ATM alpha/beta JSON config	-
GR00T_OHB_ENABLE	Enable OHB (Output Head Bias)	1
GR00T_DENOISING_STEPS	Number of denoising steps	8

---

Acknowledgements

This repo is built upon the official GR00T codebase:

• https://github.com/NVIDIA/Isaac-GR00T

Citation

If you find this code useful, please cite:

@misc{zhang2026quantvlascalecalibratedposttrainingquantization,
      title={QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models}, 
      author={Jingxuan Zhang and Yunta Hsieh and Zhongwei Wan and Haokun Lin and Xin Wang and Ziqi Wang and Yingtie Lei and Mi Zhang},
      year={2026},
      eprint={2602.20309},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.20309}, 
}