Skip to content
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

Pretrain openwebtext script with speed monitoring #147

Merged
merged 9 commits into from
Jun 15, 2023

Conversation

carmocca
Copy link
Contributor

@carmocca carmocca commented Jun 14, 2023

Part of #123

openwebtext with FSDP requires Lightning-AI/pytorch-lightning#17832

Follow-ups:

  • Add speed monitoring to fine-tuning and the other redpajama pre-training script
  • Write pretaining howto
  • XLA support

Comment on lines +61 to +62
if not torch.cuda.is_available():
return 0
Copy link
Contributor Author

Choose a reason for hiding this comment

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

We could make this function get the device type so that it can work for XLA devices too. @gkroiz Do you know where we can get the FLOPs for the different TPU archs?

Copy link
Contributor

Choose a reason for hiding this comment

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

Hi, https://cloud.google.com/tpu/docs/system-architecture-tpu-vm seems to have some numbers for v2, v3, and v4, but only for bfloat16. Let me see if I can find a document that covers FLOPs for other levels of precision

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Leaving this for a follow-up

@carmocca carmocca merged commit 42ca14e into main Jun 15, 2023
@carmocca carmocca deleted the carmocca/pretrain-improvements branch June 15, 2023 16:04
@carmocca carmocca self-assigned this Nov 1, 2023
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