chelsea0x3b/dfdx

Compile-time shape validation ensures tensor operations are type-safe, while mixed const/dynamic dimensions support flexible architectures. The gradient tape is owned by a single tensor without reference counting, eliminating runtime borrow checks and enabling precise control over which computations are recorded. GPU acceleration via CUDA is available through feature flags, alongside standard optimizers and common layers like Conv2D and Transformer.

1,896 stars. No commits in the last 6 months.