From Julia to Rust: a differentiable tensor stack for scientific computing
2 days ago
- #Scientific Computing
- #Differentiable Programming
- #Rust
- Transition from Julia to Rust for scientific computing due to Julia's runtime issues in large codebases.
- tenferro-rs is a Rust-native dense tensor stack supporting autodiff, einsum, FFT, and explicit CPU/CUDA backends.
- Rust's ownership and types enhance correctness at compile time, making it practical for AI-written code.
- Design separates operation families, autodiff rules, and backends into independent crates for extensibility.
- Supports both PyTorch-style eager autodiff and JAX-style traced transforms for dynamic shape computations.
- Correctness verified through finite-difference oracles, performance benchmarks, and design documentation.