{"schema":"hottub-news/story/v1","story":"st_s22cqlf6ve23msgn5obq","title":"摩尔线程开源 Megatron-MUSA-Titan，聚焦万卡级集群的万亿参数 MoE 模型训练","title_en":null,"outlets":3,"brief":null,"page":"https://hottub.news/stories/megatron-musa-titan-moe-s22cqlf6ve","count":3,"items":[{"id":"n_g3jcsxhmt4shr63umtua","seq":5957629,"kind":"article","title":"摩尔线程开源 Megatron-MUSA-Titan- DoNews","title_en":"Moore Threads open sources Megatron-MUSA-Titan - DoNews","url":"https://www.donews.com/news/detail/1/6757453.html","published":"2026-10-29T00:45:00Z","seen":"2026-10-29T01:26:19Z","lang":"zh","country":"CN","source":{"id":"donews.com","name":"donews.com","domain":"donews.com","outlet":{"wikidata":"Q5295684","name":"Donews","type":"website","country":"CN","designations":["website"]}},"via":"gdelt-gkg-translingual","topics":["economy"],"entities":[{"id":"Q5295684","name":"Donews","type":"org"}],"story":"st_s22cqlf6ve23msgn5obq","category":"science-technology","category_p":0.8500000238418579,"sentiment":"neutral","tone":0.5600000023841858,"political":0.10999999940395355},{"id":"n_ctmdbolf2l5kkurm6xba","seq":5729806,"kind":"article","title":"万卡万亿参数模型训练底座：摩尔线程开源Megatron-MUSA-Titan","url":"https://nb.zol.com.cn/1266/12661979.html","published":"2026-10-28T13:00:00Z","seen":"2026-10-28T13:53:31Z","lang":"zh","country":"CN","source":{"id":"zol.com.cn","name":"zol.com.cn","domain":"zol.com.cn"},"via":"gdelt-gkg-translingual","topics":["economy"],"authors":["&#x4E2D","&#x5173","&#x6751","&#x5728","&#x7EBF"],"story":"st_s22cqlf6ve23msgn5obq","category":"science-technology","category_p":0.8799999952316284,"sentiment":"neutral","tone":0.5699999928474426,"political":0.05000000074505806},{"id":"n_s22cqlf6ve23msgn5obq","seq":5652778,"kind":"article","title":"摩尔线程开源 Megatron-MUSA-Titan，聚焦万卡级集群的万亿参数 MoE 模型训练","url":"https://www.ithome.com/1/019/131.htm","summary":"IT之家 10 月 28 日消息，摩尔线程今日宣布开源大模型训练基础设施项目 Megatron-MUSA-Titan。该项目基于摩尔线程全功能 GPU， 聚焦万卡级集群的万亿参数 MoE 模型训练需求 ， 覆盖预训练与强化学习两大场景 ，为开发者提供面向 MUSA 平台的训练适配与系统优化能力。 据介绍，Megatron-MUSA-Titan 已在摩尔线程万卡级集群的万亿参数 MoE 模型预训练中持续应用，经过长时间训练实践检验，实现高 MFU（模型算力利用率）与长稳训练，同时支持与 slime、miles 强化学习框架协同，结合 Megatron 训练与 SGLang 样本生成，构建贯通万亿参数 MoE 预训练与强化学习的训练基础。 IT之家注意到，Megatron-MUSA-Titan 支持 blockwise FP8 训练，将摩尔线程全功能 GPU 的低精度计算能力引入训练流程，提升大规模训练的计算效率。还可结合 DeepEP 与 ACE 异步通信引擎，加速 MoE 中的 Token…","published":"2026-10-28T10:16:40Z","seen":"2026-10-28T10:37:01Z","lang":"zh","country":"CN","source":{"id":"kite-ithome-com-f1cdd0","name":"ithome.com","domain":"ithome.com","outlet":{"wikidata":"Q28413779","name":"IThome","type":"website","country":"CN","designations":["website"]}},"via":"kite-ithome-com-f1cdd0","story":"st_s22cqlf6ve23msgn5obq","category":"science-technology","category_p":0.8999999761581421,"sentiment":"neutral","tone":0.18000000715255737,"political":0.03999999910593033}]}