Performance

The forward! method is optimized for quick execution in the following ways:

1. All operations (matrix multiplication, softmax, swiglu) are performed in-place. [#35]
2. All operations use views (using the @views macro). [#35]
3. The RoPE loop for calculating the positional embeddings supports multithreading using the @threads macro. [#37]

For benchmarks of these optimizations, check out the linked pull requests.