New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[dtensor] make replicate -> partial do division instead #110898
Conversation
This PR switches the replicate -> partial to do division instead of zeroing out other ranks, it preserve same numerics, but avoid the per-rank behavior difference, and friendly to torch compile [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/110898
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 75a68e2 with merge base 201d02e (): This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR switches the replicate -> partial to do division instead of zeroing out other ranks, it preserve same numerics, but avoid the per-rank behavior difference, and friendly to torch compile ghstack-source-id: d2ae8a10843e79a3cbebf1c1b34aba7a7a3027b3 Pull Request resolved: #110898
This PR switches the replicate -> partial to do division instead of zeroing out other ranks, it preserve same numerics, but avoid the per-rank behavior difference, and friendly to torch compile [ghstack-poisoned]
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for doing this and this definitely makes TP (bias of rowwise linear) less complicated for users to understand.
make random ops be a set instead of list Pull Request resolved: #110900 Approved by: https://github.com/fduwjj ghstack dependencies: #110898
) as titled Pull Request resolved: #111091 Approved by: https://github.com/awgu, https://github.com/wz337 ghstack dependencies: #110898, #110900
…rch#111091) as titled Pull Request resolved: pytorch#111091 Approved by: https://github.com/awgu, https://github.com/wz337 ghstack dependencies: pytorch#110898, pytorch#110900
Stack from ghstack (oldest at bottom):
This PR switches the replicate -> partial to do division instead of
zeroing out other ranks, it preserve same numerics, but avoid the
per-rank behavior difference, and friendly to torch compile