mirror of
https://github.com/deepseek-ai/DeepGEMM.git
synced 2026-11-05 00:47:26 +00:00
[GH-ISSUE #108] 从训练的角度出发,这个可以用在megatron里面吗 #63
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepGEMM#63
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @duzhixing on GitHub (Jun 4, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/108
大模型FP8训练中,DeepSeek给了很好的先例。想请教一下DeepGemm可以和megatron想结合吗。
@LyricZhao commented on GitHub (Jun 9, 2025):
可以的,我们内部的训练都是 DeepGEMM。
@foreverlms commented on GitHub (Jun 16, 2025):
@LyricZhao Hi,有几个问题想请教下:
Thanks!
@LyricZhao commented on GitHub (Jun 16, 2025):
对的,activation 是 runtime 每次都要做;weight 就是每个 step update 参数之后 cast 一次,然后一直存着。
具体细节不太方便透露,但原则是至少要有高精度的 weight,累加还是要往 FP32 上累加;至于 FP8 的是从 FP32 cast 过来还是存一个副本,那可以框架自己斟酌。
@foreverlms commented on GitHub (Jun 16, 2025):
谢谢解答!非常感谢开源。
One more question:
这个“一直存着”是指同时保留高精度和 fp8 两个?这样就是只利用 fp8 进行加速了吧,而不是来省显存。
@LyricZhao commented on GitHub (Jun 16, 2025):
对的👍🏻 显存其实还是省的,因为 activation 都可以存 FP8,这个是大头。
@foreverlms commented on GitHub (Jun 16, 2025):
谢谢解惑!
@qijiaxing commented on GitHub (Jun 18, 2025):
我们内部有一个 结合deepGEMM & Megatron 的实验性代码,如果感兴趣的话,我们可以分享代码供参考,请发送邮件到 jqi@nvidia.com,包含姓名,公司,用途。
@Opdoop commented on GitHub (Jun 28, 2025):
@qijiaxing 这个特性什么时候会merge到megatron里?🤔