master

ascend-robot[mindspore][bugfix] bugfix for chatglm3-6b pretrain

a42e3122创建于 2025年11月13日历史提交

文件	最后提交记录	最后更新时间
seq1f1b	[mindspore][feature] support seq1f1b Co-authored-by: zihaowang26<wangzihao111@huawei.com> # message auto-generated for no-merge-commit merge: !2994 merge master_seq1f1b into master [mindspore][feature] support seq1f1b Created-by: zihaowang26 Commit-by: zihaowang26 Merged-by: ascend-robot Description: support seq1f1b See merge request: Ascend/MindSpeed!2994	6 个月前
schedules.py	[mindspore][bugfix] bugfix for chatglm3-6b pretrain Co-authored-by: wanglijun55<wanglijun54@huawei.com> # message auto-generated for no-merge-commit merge: !2996 merge branch-glm3 into master [mindspore][bugfix] bugfix for chatglm3-6b pretrain Created-by: wanglijun55 Commit-by: wanglijun55 Merged-by: ascend-robot Description: **修改说明：** 1、关联问题单：https://e.gitee.com/mind_spore/issues/table?issue=ID5HCI chatglm3-6b pretrain报错syncstream failed，定位原因为原patch中out.untyped_storage().resize_(0)的写法会导致loss tensor未初始化，确认改为out.data = torch.empty([]) 2、由于deallocate_output_tensor_ patch写法变更，custom_backward patch中assert loss.numel() == 1的判断和megatron保持一致，因而删除custom_backward patch **自验证结果：** 1、chatglm3-6b pretrain功能正常  2、qwen3-235b finetune功能正常  See merge request: Ascend/MindSpeed!2996	6 个月前