← Back to feed

Training a 1T-Parameter MoE on EU public infrastructure

AI · Technology · 2026-11-13 12:56

EUROPA联盟利用PyTorch和TorchTitan在JUPITER超算上成功训练万亿参数MoE模型，单GPU吞吐量提升至1300 tokens/秒。

Source: pytorch.org

Auto-generated summary for reference only. RSS