-
Notifications
You must be signed in to change notification settings - Fork 416
Add debugging instructions for "Reproducibility between runs" #1363
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
Conversation
docs/debugging.md
Outdated
- **Data Parallel (DP/FSDP):** All ranks use the same seed for model initialization | ||
- **Tensor Parallel (TP):** All TP ranks use the same seed for consistent weight sharding | ||
- **Pipeline Parallel (PP):** Each PP stage gets a different seed to ensure different dropout patterns | ||
- **Sequence Parallel:** Maintains consistent seeding across sequence-sharded dimensions |
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.
what's this? CP?
If so, for parameters, dp_shard
mesh and cp
mesh together form the FSDP sharding group. So if dp_shard=2, cp=2, the parameter sharding are the same as dp_shard=4, cp=1.
Co-authored-by: tianyu-l <[email protected]>
Co-authored-by: tianyu-l <[email protected]>
Co-authored-by: tianyu-l <[email protected]>
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.
please address remaining comments
As titled. To avoid someone with the same issue as me to understand why we need "seed checkpoint"