llm.c: multi-GPU, bfloat16, flash attention, ~7% faster than PyTorchtwitter.com121 pointstosh2 years ago