### Start MonetGPT LLM Server

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Command to start the LLM server for MonetGPT inference.

```bash
cd llm
sh test.sh
cd ..
```

--------------------------------

### Clone Repository and Install Dependencies

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Steps to clone the MonetGPT repository, set up a conda environment, and install project dependencies.

```bash
git clone https://github.com/monetgpt/monetgpt.git
cd monetgpt

conda create -n monetgpt python=3.11
conda activate monetgpt

cd llm 
sh install.sh
```

--------------------------------

### Start LLaMA Board GUI

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Launch the web-based GUI for fine-tuning using Gradio. This command starts the interactive interface for managing training processes.

```bash
llamafactory-cli webui
```

--------------------------------

### Install and Login to Hugging Face CLI

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Install the huggingface_hub library and log in to your Hugging Face account. This is required for accessing certain datasets.

```bash
pip install --upgrade huggingface_hub
huggingface-cli login
```

--------------------------------

### Install LLaMA-Factory with Core Dependencies

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Clone the repository and install the base package with PyTorch and metrics dependencies. Use '--no-deps' to avoid potential package conflicts.

```bash
git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e " .[torch,metrics]"

```

```bash
pip install --no-deps -e .

```

--------------------------------

### Dataset Configuration Example

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

YAML configuration for dataset settings, including LLM parameters and puzzle paths. Ensure API keys and paths are correctly set.

```yaml
# LLM settings for reasoning generation
model: "gemini-2.0-flash"
api_key: ""  # Set your API key here
base_url: "https://generativelanguage.googleapis.com/v1beta/openai/"
timeout: 5
retry_attempts: 1

# Puzzle paths and settings
puzzles:
  puzzle1:
    reasoning_path: "./data/puzzles1/reasoning/*.txt"
    images_path: "./data/puzzles1/images/*.png"
    images_base_path: "./data/puzzles1/images"
    output_file: "data/sharegpt_puzzle_1.json"
  
  puzzle2:
    reasoning_path: "./data/puzzles2/reasoning/*.txt"
    images_path: "./data/puzzles2/images/*.png"
    images_base_path: "./data/puzzles2/images"
    output_file: "data/sharegpt_puzzle_2.json"
    
  puzzle3:
    reasoning_path: "./data/puzzles3/reasoning/*.txt"
    images_path: "./data/puzzles3/images/*/*.tif"
    images_base_path: "./data/puzzles3/images"
    output_file: "data/sharegpt_puzzle_3.json"

# Generation settings
generation:
  num_standard_trials: 2
  num_color_trials: 1
  num_puzzle3_trials: 10

```

--------------------------------

### Install bitsandbytes for Windows (QLoRA)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

For Windows users enabling quantized LoRA (QLoRA), install a pre-built bitsandbytes wheel compatible with your CUDA version.

```bash
pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.2.post2-py3-none-win_amd64.whl

```

--------------------------------

### Install FlashAttention-2 for Windows

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

To enable FlashAttention-2 on Windows, install a precompiled version compatible with your CUDA version from the provided releases.

```bash
# Please download the corresponding version from [flash-attention](https://github.com/bdashore3/flash-attention/releases) based on your requirements.

```

--------------------------------

### Install NumPy for GIMP (Linux Flatpak)

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Command to install the NumPy library for GIMP when using the Linux Flatpak version.

```bash
flatpak run --command=sh org.gimp.GIMP//stable -c "python -m pip install --user numpy"
```

--------------------------------

### Run Batch Image Processing with MonetGPT

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Example command to process multiple images in a directory using the MonetGPT batch processing mode.

```bash
python inference_cli.py batch assets/test --output-dir results/
```

--------------------------------

### Install GIMP 2.10 on Linux via Flatpak

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Instructions for installing GIMP version 2.10 using Flatpak on Linux systems, including adding the Flathub repository.

```bash
sudo apt update
sudo apt install flatpak
flatpak remote-add --if-not-exists flathub https://flathub.org/repo/flathub.flatpakrepo
flatpak install flathub org.gimp.GIMP//2.10 
```

--------------------------------

### Deploy with OpenAI-style API and vLLM

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Start an API server using vLLM for efficient inference, compatible with OpenAI's API format. Specify the API port and the model configuration YAML.

```bash
API_PORT=8000 llamafactory-cli api examples/inference/llama3_vllm.yaml
```

--------------------------------

### Install LLaMA-Factory on Ascend NPU

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Install LLaMA Factory with Ascend NPU specific dependencies and the Ascend CANN Toolkit and Kernels. Set environment variables to use NPU devices.

```bash
pip install -e ".[torch-npu,metrics]"

```

```bash
# replace the url according to your CANN version and devices
# install CANN Toolkit
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Milan-ASL/Milan-ASL%20V100R001C17SPC701/Ascend-cann-toolkit_8.0.RC1.alpha001_linux-$(uname -i).run
bash Ascend-cann-toolkit_8.0.RC1.alpha001_linux-$(uname -i).run --install

# install CANN Kernels
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Milan-ASL/Milan-ASL%20V100R001C17SPC701/Ascend-cann-kernels-910b_8.0.RC1.alpha001_linux.run
bash Ascend-cann-kernels-910b_8.0.RC1.alpha001_linux.run --install

# set env variables
source /usr/local/Ascend/ascend-toolkit/set_env.sh

```

```bash
# Remember to use `ASCEND_RT_VISIBLE_DEVICES` instead of `CUDA_VISIBLE_DEVICES` to specify the device to use.

```

--------------------------------

### Pipeline Configuration Example

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

YAML configuration for GIMP and image processing settings. Specifies paths for GIMP executables and processing parameters like batch size and timeout.

```yaml
# GIMP settings
gimp:
  paths:
    macos: "/Applications/GIMP.app/Contents/MacOS/gimp-console-2.10"
    linux: "flatpak run org.gimp.GIMP//stable --no-interface"
    windows: "gimp-console-2.10.exe" (Not tested, may require some modifications)
  
  batch_interpreter: "python-fu-eval"
  python_warnings: "ignore"
  pipeline_file: "./gimp_pipeline.py"

# Image processing settings
image_processing:
  max_low_res_size: 700  # Low resolution for MLLM training only
  default_dpi: 140       # Original resolution is preserved during inference

# Processing parameters
processing:
  batch_size: 10
  max_workers: 4
  timeout_seconds: 120

```

--------------------------------

### Run Single Image Processing with MonetGPT

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Example command to perform image enhancement or retouching on a single input image using the MonetGPT CLI.

```bash
python inference_cli.py single input.jpg --output results/edited.jpg
```

--------------------------------

### Install Git LFS

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Ensure Git Large File Storage (LFS) is installed to handle large model files. Run this command in your terminal.

```bash
git lfs install
```

--------------------------------

### Build Docker Image with Docker Compose (CUDA)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Build and run the Docker image for CUDA users using Docker Compose. This simplifies the setup for GPU-accelerated environments.

```bash
cd docker/docker-cuda/
docker compose up -d
docker compose exec llamafactory bash
```

--------------------------------

### Prepare Training Data Directory

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Creates the necessary directory structure for training images.

```bash
mkdir -p data/ppr10k
```

--------------------------------

### Download Pre-trained MonetGPT Model

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Instructions for downloading the pre-trained MonetGPT model using the Hugging Face CLI or by cloning the repository.

```bash
cd llm
mkdir -p models
cd models

huggingface-cli download niladridutt/monetGPT

# OR Clone the model repository (requires git lfs)
git clone https://huggingface.co/niladridutt/monetGPT
```

--------------------------------

### Create Final ShareGPT Format Datasets

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Creates datasets in ShareGPT format for all generated puzzles. Ensure the output file is placed correctly for training.

```bash
python dataset_cli.py create          # Create datasets for all puzzles
```

--------------------------------

### Build Docker Image with Docker Compose (Ascend NPU)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Build and run the Docker image for Ascend NPU users using Docker Compose. Ensure your environment is compatible with Ascend hardware.

```bash
cd docker/docker-npu/
docker compose up -d
docker compose exec llamafactory bash
```

--------------------------------

### Combine Datasets for Training

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Combines JSON datasets and exports them for training. The combined file should be named 'monetgpt.json' and placed under the 'llm/data' directory.

```bash
python dataset/combine_jsons.py       # Combine JSON datasets and export for training
```

--------------------------------

### Run LoRA Fine-tuning, Inference, and Merging

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Execute commands for LoRA fine-tuning, inference, and merging of the Llama3-8B-Instruct model. Ensure the necessary YAML configuration files are in place.

```bash
llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
```

```bash
llamafactory-cli chat examples/inference/llama3_lora_sft.yaml
```

```bash
llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml
```

--------------------------------

### Train MonetGPT Model

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Trains the MonetGPT model on the generated datasets. This command should be run from the 'llm' directory.

```bash
cd llm
sh train.sh
```

--------------------------------

### Configure Weights & Biases Logging

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Add these arguments to your YAML file to enable logging with Weights & Biases. Ensure WANDB_API_KEY is set.

```yaml
report_to: wandb
run_name: test_run # optional
```

--------------------------------

### Enable ModelScope Hub Usage

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Set an environment variable to use ModelScope Hub for downloading models and datasets. This is an alternative to Hugging Face Hub.

```bash
export USE_MODELSCOPE_HUB=1
# `set USE_MODELSCOPE_HUB=1` for Windows
```

--------------------------------

### Build Docker Image without Docker Compose (Ascend NPU)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Manually build the Docker image for Ascend NPU users. This method allows for specific configurations and device mapping for NPU hardware.

```bash
# Choose docker image upon your environment
docker build -f ./docker/docker-npu/Dockerfile \
    --build-arg INSTALL_DEEPSPEED=false \
    --build-arg PIP_INDEX=https://pypi.org/simple \
    -t llamafactory:latest .

# Change `device` upon your resources
docker run -dit \
    -v ./hf_cache:/root/.cache/huggingface \
    -v ./ms_cache:/root/.cache/modelscope \
    -v ./om_cache:/root/.cache/openmind \
    -v ./data:/app/data \
    -v ./output:/app/output \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -p 7860:7860 \
    -p 8000:8000 \
    --device /dev/davinci0 \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

docker exec -it llamafactory bash
```

--------------------------------

### Generate Multi-Version Comparison Puzzles

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Generates multi-version comparison puzzles using the pipeline CLI. This helps in teaching parameter value relationships.

```bash
python pipeline_cli.py puzzle 2
```

--------------------------------

### Build Docker Image with Docker Compose (AMD ROCm)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Build and run the Docker image for AMD ROCm users using Docker Compose. This is for environments utilizing AMD GPUs with ROCm support.

```bash
cd docker/docker-rocm/
docker compose up -d
docker compose exec llamafactory bash
```

--------------------------------

### Create Visual Puzzle Images

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Creates visual puzzle images using the pipeline CLI. Supports generating single operation, multi-version comparison, comprehensive editing, or all puzzle types.

```bash
python pipeline_cli.py puzzle 1     # Single operation puzzles
```

```bash
python pipeline_cli.py puzzle 2     # Multi-version comparison puzzles  
```

```bash
python pipeline_cli.py puzzle 3     # Comprehensive editing puzzles
```

```bash
python pipeline_cli.py puzzle all   # Generate all puzzle types
```

--------------------------------

### Generate Comprehensive Editing Puzzles

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Generates comprehensive editing puzzles using the pipeline CLI. These puzzles teach complete retouching workflows.

```bash
python pipeline_cli.py puzzle 3
```

--------------------------------

### Batch Process Images with MonetGPT Predictions

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Processes multiple images in batch using MonetGPT predictions. Specify the target editor for the batch processing.

```bash
python pipeline_cli.py batch predictions --target-editor a
```

--------------------------------

### Apply Single Image Edit Configuration

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Applies a specific retouching configuration to a single input image and saves the output.

```bash
python pipeline_cli.py edit configs/example_edit.json input.jpg output.jpg
```

--------------------------------

### Generate Puzzle Configurations

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Generates puzzle configurations using the dataset CLI. Use different arguments for single operation, multi-version, or comprehensive editing puzzles.

```bash
python dataset_cli.py generate
```

--------------------------------

### Configure SwanLab Logging

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Add these arguments to your YAML file to enable logging with SwanLab. You can log in via API key in YAML, environment variable, or the `swanlab login` command.

```yaml
use_swanlab: true
swanlab_run_name: test_run # optional
```

--------------------------------

### Build Docker Image without Docker Compose (AMD ROCm)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Manually build the Docker image for AMD ROCm users. This command configures the build for AMD GPUs and maps necessary devices.

```bash
docker build -f ./docker/docker-rocm/Dockerfile \
    --build-arg INSTALL_BNB=false \
    --build-arg INSTALL_VLLM=false \
    --build-arg INSTALL_DEEPSPEED=false \
    --build-arg INSTALL_FLASHATTN=false \
    --build-arg PIP_INDEX=https://pypi.org/simple \
    -t llamafactory:latest .

docker run -dit \
    -v ./hf_cache:/root/.cache/huggingface \
    -v ./ms_cache:/root/.cache/modelscope \
    -v ./om_cache:/root/.cache/openmind \
    -v ./data:/app/data \
    -v ./output:/app/output \
    -v ./saves:/app/saves \
    -p 7860:7860 \
    -p 8000:8000 \
    --device /dev/kfd \
    --device /dev/dri \
    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

docker exec -it llamafactory bash
```

--------------------------------

### Generate All Puzzle Types

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Generates all available puzzle types using the pipeline CLI. This is a convenient option for creating a complete set of training data.

```bash
python pipeline_cli.py puzzle all
```

--------------------------------

### Build Docker Image without Docker Compose (CUDA)

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Manually build the Docker image for CUDA users without Docker Compose. This provides more control over the build process and arguments.

```bash
docker build -f ./docker/docker-cuda/Dockerfile \
    --build-arg INSTALL_BNB=false \
    --build-arg INSTALL_VLLM=false \
    --build-arg INSTALL_DEEPSPEED=false \
    --build-arg INSTALL_FLASHATTN=false \
    --build-arg PIP_INDEX=https://pypi.org/simple \
    -t llamafactory:latest .

docker run -dit --gpus=all \
    -v ./hf_cache:/root/.cache/huggingface \
    -v ./ms_cache:/root/.cache/modelscope \
    -v ./om_cache:/root/.cache/openmind \
    -v ./data:/app/data \
    -v ./output:/app/output \
    -p 7860:7860 \
    -p 8000:8000 \
    --shm-size 16G \
    --name llamafactory \
    llamafactory:latest

docker exec -it llamafactory bash
```

--------------------------------

### Generate Single Operation Puzzles

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Generates single operation puzzles using the pipeline CLI. This is useful for teaching individual retouching operations.

```bash
python pipeline_cli.py puzzle 1
```

--------------------------------

### BibTeX Citation for LlamaFactory

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

This is the BibTeX entry to cite the LlamaFactory work, which is relevant to the MonetGPT project.

```bibtex
@inproceedings{zheng2024llamafactory,
  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models},
  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  address={Bangkok, Thailand},
  publisher={Association for Computational Linguistics},
  year={2024},
  url={http://arxiv.org/abs/2403.13372}
}
```

--------------------------------

### Set Modelers Hub Environment Variable

Source: https://github.com/niladridutt/monetgpt/blob/main/llm/README.md

Set the USE_OPENMIND_HUB environment variable to 1 to enable downloading models from Modelers Hub. Use `set` for Windows.

```bash
export USE_OPENMIND_HUB=1 # `set USE_OPENMIND_HUB=1` for Windows
```

--------------------------------

### Generate LLM Reasoning for Puzzles

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Queries the LLM to add reasoning to puzzle configurations. Specify puzzle type and range for reasoning generation.

```bash
python dataset_cli.py query 1 0 -1    # Generate reasoning for all puzzle 1 configs (all)
```

```bash
python dataset_cli.py query 2 0 -1   # Generate reasoning for puzzle 2 configs (all)
```

```bash
python dataset_cli.py query 3 0 -1    # Generate reasoning for puzzle 3 configs (all)
```

--------------------------------

### MonetGPT Citation (BibTeX)

Source: https://github.com/niladridutt/monetgpt/blob/main/README.md

Use this BibTeX entry to cite MonetGPT in academic publications. It includes details about the paper, authors, journal, and publication year.

```bibtex
@article{dutt2025monetgpt,
  title={MonetGPT: Solving Puzzles Enhances MLLMs' Image Retouching Skills},
  author={Dutt, Niladri Shekhar and Ceylan, Duygu and Mitra, Niloy J},
  journal={ACM Transactions on Graphics (TOG)},
  volume={44},
  number={4},
  pages={1--12},
  year={2025},
  publisher={ACM New York, NY, USA}
}
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.