Notes

Warp Divergence and Why Branches Are Expensive on GPUs CUDA KV-Caching, and How Autoregressive Decoding Stays Fast LLMs Layer Normalization, and Why Transformers Need It Deep Learning