#LLM-inference

DFlash vs MTP: Two Paths to Breaking the Autoregressive Decoding Bottleneck