This now loads the `gpt2_124M_debug_state.bin` file, runs a forward pass, compares the logits and loss with the PyTorch reference implementation, then it does 10 iterations of training with Adam and makes sure the losses match PyTorch.

MIT