Announcing Long-Context Training on TPUs with MaxText Context Parallelism
Google announces MaxText support for long-context training on TPUs using four context-parallel strategies (All-gather, A2A, P2P, USP) with the Tokamax Splash Attention kernel, achieving up to 70% MFU for 4M-token sequences on TPU 5p and 50% MFU on TPU 7 for 14M-token sequences.