High-Performance Multimodal Attention with Custom CUDA Kernels

This project demonstrates a vertical slice of a high-performance Multimodal AI system. It features:

1. Custom CUDA Kernel: A fused "Normalize & Project" kernel written in CUDA C++ for low-latency fusion of video and text embeddings.
2. Multimodal Architecture: A compact Video-Text Transformer model.
3. Distributed Training: FSDP (Fully Sharded Data Parallel) training loop designed to run on dual T4 GPUs (available for free on Kaggle/Colab).

1. Push to GitHub: Sync this local folder to a public GitHub repository named OptiMulti-Video.
2. Open the Notebook: Upload notebooks/colab_demo.ipynb to Google Colab.
3. Run: The notebook is pre-configured to:
   • Clone your repository.
   • Install dependencies.
   • Compile the Custom CUDA kernel (JIT compilation).
   • Run the FSDP training demo.

• src/: C++ and CUDA source code for the custom kernel.
• model/: PyTorch model definitions (Vision Encoder, Text Decoder, Fusion).
• training/: Training loops and distributed data configuration.
• notebooks/: Jupyter notebooks for demonstration.