Lumina - Rust Tensor Framework

Developed as a research-grade tensor library, Lumina combines Rust's memory safety with GPU acceleration through vulkan and CUDA backends. The framework supports both f32 and f16 precision modes, making it ideal for both training and inference workloads.

Core Features

• • 2.4x speed improvement over numpy for large matrix operations
• • Automatic GPU memory pooling with Vulkan/CUDA abstraction layer
• • Type-safe tensor operations with compile-time shape verification
• • Python bindings via PyO3 for hybrid workflows

Performance Benchmarks

Stress testing matrix multiplication operations across different device configurations:

Code Example

Creating and manipulating tensors in Lumina:

#[derive(Debug)]
struct Tensor {
    data: Vec,
    shape: Vec,
}

impl + Copy> Tensor {
    fn add(&self, other: &Self) -> Self {
        assert_eq!(self.shape, other.shape);
        
        Tensor {
            data: self.data.iter()
                        .zip(other.data.iter())
                        .map(|(a, b)| *a + *b)
                        .collect(),
            shape: self.shape.clone(),
        }
    }
}
            

Ecosystem Integration

Lumina provides first-class integration with: