How to install with cuda 12.8

[LukeLIN-web]

Here only have https://github.com/Vchitect/VBench/tree/master/VBench-2.0#hammer-installation cuda 11.8.

[LukeLIN-web]

VBench 2.0 Installation Guide

Tested on: Ubuntu 22.04, NVIDIA Driver 570+, GCC 11.4, FFmpeg 6.x

Prerequisites

• NVIDIA GPU with CUDA support (Driver >= 525)
• GCC >= 11.2
• FFmpeg >= 4.x
• Conda (Miniconda or Anaconda)

Step 1: Create Conda Environment

conda create -n vbench2 python=3.10 -y
conda activate vbench2
conda install psutil -y

Step 2: Install PyTorch (CUDA 11.8)

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118

Verify:

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"
# Expected: 2.5.1+cu118 True

Step 3: Install Build Tools

pip install ninja

Step 4: Install Flash Attention

Note: Must use --no-build-isolation since flash-attn needs torch during build.

pip install flash-attn==2.7.2.post1 --no-build-isolation

This takes 5-10 minutes to compile.

Step 5: Downgrade setuptools

Newer setuptools (>= 72) removes pkg_resources from build isolation, which breaks CLIP and some other packages.

pip install "setuptools<71"

Step 6: Install Git-based Dependencies

These three packages must be installed with --no-build-isolation:

# CLIP
pip install --no-build-isolation git+https://github.com/openai/CLIP.git@dcba3cb2e2827b402d2701e7e1c7d9fed8a20ef1

# LLaVA-NeXT
pip install --no-build-isolation git+https://github.com/LLaVA-VL/LLaVA-NeXT@79ef45a6d8b89b92d7a8525f077c3a3a9894a87d

# YOLO-World (mmcv will fail here, that's expected - we install it separately in Step 9)
pip install --no-build-isolation git+https://github.com/AILab-CVC/YOLO-World.git@4826695a84ce370dad7fe4cb555a6c7162f3751d || true

Step 7: Install Requirements

cd VBench-2.0

# Install all non-git requirements
grep -v "^-e git" requirement.txt | pip install --no-build-isolation -r /dev/stdin

# Install retinaface_pytorch (must use --no-deps to avoid conflicts)
pip install retinaface_pytorch==0.0.8 --no-deps

Step 8: Install Instance Detector

cd vbench2/third_party/Instance_detector
pip install -e .
cd ../../..

Step 9: Install mmcv

Use pre-built wheels for CUDA 11.8. The torch2.4 wheels are compatible with torch 2.5.1:

pip install mmcv==2.2.0 -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.4/index.html --no-cache-dir

Step 10: Fix mmdet & mmyolo Version Check

mmdet 3.0.0 and mmyolo 0.6.0 ship with a hard-coded mmcv_maximum_version = '2.1.0' which rejects mmcv 2.2.0. Fix this:

# Find the conda env path
CONDA_ENV=$(python -c "import sys; print(sys.prefix)")

# Fix mmdet
sed -i "s/mmcv_maximum_version = '2.1.0'/mmcv_maximum_version = '2.3.0'/" \
    "$CONDA_ENV/lib/python3.10/site-packages/mmdet/__init__.py"

# Fix mmyolo
sed -i "s/mmcv_maximum_version = '2.1.0'/mmcv_maximum_version = '2.3.0'/" \
    "$CONDA_ENV/lib/python3.10/site-packages/mmyolo/__init__.py"

Step 11: Verify Installation

python -c "
import torch
print(f'PyTorch: {torch.__version__}, CUDA: {torch.cuda.is_available()}')
import flash_attn; print(f'Flash Attention: {flash_attn.__version__}')
import clip; print('CLIP: OK')
import llava; print('LLaVA: OK')
import mmcv; print(f'mmcv: {mmcv.__version__}')
import mmdet; print(f'mmdet: {mmdet.__version__}')
import mmyolo; print(f'mmyolo: {mmyolo.__version__}')
import transformers; print(f'transformers: {transformers.__version__}')
import decord; print('decord: OK')
print('All key packages imported successfully!')
"

Step 12: Download Pretrained Models (Optional)

Download pretrained weights to ~/.cache/vbench2 following the download.sh scripts in the pretrained/ folder.

For LLaVA-Video-7B-Qwen2 and Qwen2.5-7B-Instruct, it is highly recommended to download before evaluation:

cd pretrained
bash download.sh

Troubleshooting

ModuleNotFoundError: No module named 'pkg_resources'

Downgrade setuptools: pip install "setuptools<71", and use --no-build-isolation for the failing package.

AssertionError: MMCV==2.2.0 is used but incompatible

Apply the mmdet/mmyolo fix described in Step 10.

Flash Attention build fails with No module named 'torch'

Use --no-build-isolation: pip install flash-attn==2.7.2.post1 --no-build-isolation

mmcv build fails from source

Use pre-built wheels with the -f flag as shown in Step 9. Make sure the CUDA and torch versions in the URL match your setup.

[ernestchu]

@LukeLIN-web hats off to you (or whatever llm you used to solve this environment hell)

[LukeLIN-web]

@LukeLIN-web hats off to you (or whatever llm you used to solve this environment hell)

lol, The tide turned with claude code.