AI 导读
据科创板日报消息人士,DeepSeek 计划在华为昇腾芯片上预训练下一代 DeepSeek-V5。此前昇腾只承担 DeepSeek 推理和 V4-Pro 后训练;作者认为 DeepSeek 于 9 月 30 日开源完整昇腾算力与通信栈(DeepGEMM、FlashMLA、DeepEP)与此计划相呼应。
推荐理由
把 Ascend 此前只用于推理和 V4-Pro 后训练的背景与 9 月 30 日开源算力栈联系起来,读者可以据此理解 DeepSeek 的芯片迁移路径。
正文
DeepSeek plans to pre-train its next-gen V5 model on Huawei Ascend chips, per STAR Market Daily sources.
Until now, Ascend only handled DeepSeek inference and V4-Pro post-training. Suddenly, DeepSeek open-sourcing its full Ascend compute & comms stack (DeepGEMM, FlashMLA, DeepEP) on Sept 30 makes total sense.
来源:X.PIN · x.com