Skip to content

Conversation

@allenwang28
Copy link
Contributor

A clean up, and resolves #532

Changes:

  1. Removes legacy code. We had a gpu_manager system controller but it was re-implemented in provisioner. Finally cleans it up
  2. Dynamic GPU detection - GpuManager in provisioner now users torch.cuda.device_count() instead of hardcoding to 8 GPUs.
  3. Adds a function to get GPU information per host to populate GpuManager on remote host meshes.
  4. Updates test cases.

@meta-cla meta-cla bot added the CLA Signed This label is managed by the Meta Open Source bot. label Nov 6, 2025
Copy link
Contributor

@casteryh casteryh left a comment

Choose a reason for hiding this comment

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

lgtm. why are there only 2 test cases now? Are the old tests no longer relevant because of refactoring?

Copy link
Contributor

@casteryh casteryh left a comment

Choose a reason for hiding this comment

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

image

@casteryh casteryh merged commit 29c6584 into meta-pytorch:main Nov 6, 2025
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Meta Open Source bot.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

GpuManager assumes there are 8 GPUs?

2 participants