torsh-core

Core types and traits for the ToRSh deep learning framework.

Overview

This crate provides the fundamental building blocks used throughout ToRSh:

• Device abstraction: Unified interface for CPU, CUDA, Metal, and WebGPU backends
• Data types: Support for various tensor element types (f32, f64, i32, etc.)
• Shape utilities: Shape manipulation, broadcasting, and stride calculations
• Storage abstraction: Backend-agnostic tensor storage with reference counting
• Error types: Comprehensive error handling for ToRSh operations

Features

• std (default): Standard library support
• no_std: No standard library (for embedded targets)
• serialize: Serialization support via serde

Usage

use torsh_core::prelude::*;

// Create a shape
let shape = Shape::new(vec![2, 3, 4]);
println!("Shape: {}, elements: {}", shape, shape.numel());

// Device management
let device = CpuDevice;
println!("Device: {}", device.name());

// Data types
let dtype = DType::F32;
println!("Data type: {}, size: {} bytes", dtype, dtype.size());

Integration with SciRS2

This crate builds on top of scirs2 for core scientific computing functionality, providing a PyTorch-compatible API layer.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.