v0.2.2-dev

cann-robot[fix] tune DeepSeek V4 expert parallel degree

7baa57fe创建于 2 天前历史提交

文件	最后提交记录	最后更新时间
common	fix: use module-level LOSS_SCALE global instead of class attribute	15 天前
deepseek_v3	[fix] fix virtual_optimizer	22 天前
deepseek_v32	[fix] fix scatter_/gather input index out of bounds	21 天前
deepseek_v4	[fix] tune DeepSeek V4 expert parallel degree	2 天前
llama4	Fixed license headers	1 个月前