Skip to content

Use the value of DISC_DEVICE as the device type of disc backend#8

Merged
yitongh merged 2 commits intomerge_discfrom
fix_disc_amp
Oct 10, 2024
Merged

Use the value of DISC_DEVICE as the device type of disc backend#8
yitongh merged 2 commits intomerge_discfrom
fix_disc_amp

Conversation

@yitongh
Copy link
Copy Markdown

@yitongh yitongh commented Aug 14, 2024

No description provided.

@yitongh yitongh requested a review from Yancey0623 August 14, 2024 08:57
@yitongh yitongh changed the title change the device type of disc to cuda to make amp work properly Use the value of DISC_DEVICE as the device type of disc backend Aug 15, 2024
Copy link
Copy Markdown

@Yancey0623 Yancey0623 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@yitongh yitongh merged commit aa2c40e into merge_disc Oct 10, 2024
@yitongh yitongh deleted the fix_disc_amp branch October 10, 2024 03:35
anw90 pushed a commit that referenced this pull request Oct 11, 2024
* build with BladeDISC (#8)

* [to #53687860] feat: DISC client header, implement DISCComputation and DISCData

POC implement in : https://code.alibaba-inc.com/torchx/xla/codereview/14984824

Link: https://code.alibaba-inc.com/torchx/xla/codereview/14987956

* Disc computation (#2)

Support Disc as backend
Co-authored-by: yancey.yx <yancey.yx@antfin.com>
Co-authored-by: wangang.wa <wangang.wa@alibaba-inc.com>

* add bazel flag to disable disc backend (#23)

* add flag to disable disc backend in bazel workspace

* support disc debug mode to dump mhlo and logs (#25)

support disc backend debug mode to dump DISC compilation logs

* support flash attention in disc (pytorch#34)

* fix disc flag when complie python (pytorch#39)

* fix bazel flag when complie python

* fix lint.

* support bf16 on disc backend (pytorch#40)

add float-norm pass to support bf16 amp training

* Support Flash Attention 2.5.6 for disc backend (#4)

* fix build failed with NCCL (#5)

* fix build failed on nccl

* using nccl hdrs

* Use the value of DISC_DEVICE as the device type of disc backend (#8)

* change the device type of disc to cuda to make amp work properly

* Use the value of DISC_DEVICE as the device type of disc backend

* disable compilation of DISC by default (#15)

---------

Co-authored-by: Yan Xu <yancey1989@gmail.com>
Co-authored-by: wenting.swt <wenting.swt@alibaba-inc.com>
Co-authored-by: Dalong <yuanxiulong.yxl@alibaba-inc.com>
Co-authored-by: Baole Ai <baole.abl@alibaba-inc.com>
Co-authored-by: Yan Xu <yancey.yx@alibaba-inc.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants