Built an LLM trainer in pure Go on XLA to see if it could keep up with the PyTorch original. Ran both to a 2B-token Chinchilla budget on the same 8GB 3070 Ti, same 100M model, same config, same tokens, same seed. Go-on-XLA lands at 85% of PyTorch throughput at ~1.7x the memory, and the loss curves track within 0.1 the whole way. Three weeks ago the same stack was 26x slower and couldn’t fit the model at all.