ModelResharder-Transformers C++

A robust C++ CLI tool for downloading, resharding, and re-uploading large Hugging Face models. It uses pybind11 to seamlessly bind with the Python Hugging Face ecosystem (transformers, huggingface_hub, torch), offering the safety and structure of C++ with the vast library support of Python.

Primary Focus: Maintains version control and transforms files from older Transformers versions to newer formats with configurable shard sizes.

Prerequisites

• C++17 Compiler (GCC, Clang, MSVC)
• CMake (v3.14+)
• Python 3.10+ (must be accessible to CMake)
• Python packages: torch, transformers, huggingface_hub

Build Instructions

You can build the project from the root of the repository using CMake:

1. Configure the Project

   cmake -B build -S .

2. Build the Project

   cmake --build build --config Release

Usage

After building, you can run the compiled executable located in the build/Release directory (or build/ on Linux/macOS) with the required arguments.

Example: Resharding Qwen3.5

Windows (PowerShell / CMD):

.\build\Release\resharder.exe --model_name username/source-model_page `
                              --repo_id your-username/target-model_page `
                              --token <YOUR_HF_TOKEN> `
                              --shard_size 4.4GB `
                              --arch Qwen3_5ForConditionalGeneration

Linux / macOS (Bash):

./build/resharder --model_name username/source-model_page \
                  --repo_id your-username/target-model_page \
                  --token <YOUR_HF_TOKEN> \
                  --shard_size 4.4GB \
                  --arch Qwen3_5ForConditionalGeneration

Options

• --model_name: Original HuggingFace Model Name
• --repo_id: Target HuggingFace Repository ID (will be created as private)
• --token: Your HuggingFace Write Token (ensure you keep this secure and don't commit it!)
• --shard_size: Max Shard Size (default: 4.4GB)
• --arch: Model Architecture Class Name (default: Qwen3VLForConditionalGeneration, make sure to pass the right architecture for your model, e.g., Qwen2ForCausalLM or Qwen3_5ForConditionalGeneration)

Important Note on Security

Never expose or commit your HuggingFace token (hf_...). Pass it securely via the command line when executing the binary.

Running in Google Colab

You can easily run this C++ application in Google Colab. Colab already provides the essential tools (Python 3, GCC, and CMake).

1. Open a new Google Colab Notebook.
2. Create a new code cell and paste the following block to clone, build, and run the project all at once:

# Clone the repository
!git clone https://github.com/PRITHIVSAKTHIUR/model.resharder-transformers.cpp.git
%cd model.resharder-transformers.cpp

# Install dependencies (Python dev headers)
!apt-get update && apt-get install -y python3-dev

# Build the project
!mkdir build
%cd build
!cmake ..
!cmake --build . --config Release

# Run the executable (replace arguments with your specific model details)
!./resharder --model_name username/source-model_page \
             --repo_id your-username/target-model_page \
             --token <YOUR_HF_TOKEN> \
             --shard_size 4.4GB \
             --arch Qwen3_5ForConditionalGeneration