Environment:Norrrrrrr lyn WAInjectBench Conda Python 39 CUDA Environment

Knowledge Sources	WAInjectBench environment.yml
Domains	Infrastructure, Deep_Learning, Security
Last Updated	2026-02-14 16:00 GMT

Overview

Linux environment with Python 3.9, CUDA 12.1, PyTorch 2.5.1, and the full WAInjectBench dependency stack managed via Conda.

Description

This environment provides the core runtime for all WAInjectBench workflows including text detection, image detection, embedding classifier training, and LLaVA fine-tuning. It is defined by the `environment.yml` file with the Conda environment named `pidefense`. The stack includes PyTorch 2.5.1 with CUDA 12.1 support, Hugging Face Transformers 4.56.1, OpenCLIP 3.1.0, sentence-transformers 5.1.0, scikit-learn 1.6.1, PEFT 0.17.1, accelerate 1.10.1, and bitsandbytes 0.43.3. The CUDA toolkit (11.7 via conda, 12.1 via pip wheels) provides GPU acceleration for all neural model inference and training.

Usage

Use this environment for all WAInjectBench tasks: running detectors (text and image), training embedding classifiers, and fine-tuning LLaVA models. It is the mandatory prerequisite for every Implementation page in this repository.

System Requirements

Category	Requirement	Notes
OS	Linux (Ubuntu recommended)	Conda environment targets Linux x86_64
Hardware	NVIDIA GPU with CUDA support	Minimum 8GB VRAM for inference; 24GB+ recommended for LLaVA fine-tuning
Hardware	CPU with sufficient RAM	16GB+ RAM recommended for embedding training
Disk	20GB+ SSD	For conda environment, model weights, and dataset storage

Dependencies

System Packages (Conda)

`python` = 3.9.19
`cudatoolkit` = 11.7.0
`cuda-runtime` = 12.1.0
`cuda-libraries` = 12.1.0
`numpy` = 1.26.4
`scipy` = 1.11.4

Python Packages (pip)

`torch` = 2.5.1+cu121
`torchvision` = 0.20.1+cu121
`transformers` = 4.56.1
`accelerate` = 1.10.1
`peft` = 0.17.1
`bitsandbytes` = 0.43.3
`open-clip-torch` = 3.1.0
`sentence-transformers` = 5.1.0
`scikit-learn` = 1.6.1
`joblib` = 1.5.2
`pillow` = 10.3.0
`openai` = 1.108.0
`spacy` = 3.7.2
`huggingface-hub` = 0.34.4
`safetensors` = 0.6.2
`triton` = 3.1.0
`tqdm` >= 4.67

Credentials

The following environment variables may be required depending on the detector used:

`CUDA_VISIBLE_DEVICES`: Set automatically by entry-point scripts (`main_text.py`, `main_image.py`, `train/llava-ft.py`) to select the GPU.
`TOKENIZERS_PARALLELISM`: Set to `"false"` by `train/llava-ft.py` to suppress Hugging Face tokenizer warnings.

Quick Install

# Create and activate conda environment
conda env create -f environment.yml
conda activate pidefense

Code Evidence

GPU availability check from `train/llava-ft.py:245`:

use_cuda = torch.cuda.is_available()

CUDA device selection from `main_text.py:50` and `main_image.py:57`:

os.environ["CUDA_VISIBLE_DEVICES"] = gpu

Tokenizer parallelism suppression from `train/llava-ft.py:3`:

os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")

Device fallback in `train/embedding-i.py:75-77`:

device = torch.device(device if torch.cuda.is_available() else "cpu")

Common Errors

Error Message	Cause	Solution
`RuntimeError: CUDA out of memory`	Insufficient VRAM for LLaVA 7B model	Use `--device_mode single --gpu_id 0` or reduce `--batch_size`; the code auto-falls back to multi-GPU redispatch
`ImportError: No module named 'peft'`	PEFT not installed	`pip install peft>=0.17.1`
`No jsonl files found in the input directory.`	Wrong input path	Verify `--input_dir` / `--data_dir` points to directory containing `.jsonl` files

Compatibility Notes

CPU-only mode: The embedding classifiers (`embedding-t.py`, `embedding-i.py`) will fall back to CPU if no GPU is available. LLaVA fine-tuning requires a CUDA GPU.
Multi-GPU: The `--device_mode auto` flag in `llava-ft.py` uses `accelerate.dispatch_model` for multi-GPU inference, but training defaults to single GPU.
BF16 support: The LLaVA training script checks `torch.cuda.is_bf16_supported()` and falls back to FP32 if bfloat16 is unavailable.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment