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

#7560: Fix FD noc api to set multicast_path_reserve when mcasting #10911

Merged
merged 3 commits into from
Jul 31, 2024

Conversation

tt-aho
Copy link
Contributor

@tt-aho tt-aho commented Jul 30, 2024

Ticket

#7560

Problem description

Stable diffusion was still seeing hangs after di/dt fixes. Was narrowed down to a commit that changed the FD eth cores used for prefetch/dispatch, which affected timing. The real latent bug was that multicast_path_reserve was not being set with the new noc apis used in some FD commands, leading to potential NOC deadlock.

What's changed

Always set multicast_path_reserve for now for all mcast transactions with the new apis.

Checklist

  • Post commit CI passes
  • Model regression CI testing passes (if applicable)
  • New/Existing tests provide coverage for changes

Copy link
Contributor

@pgkeller pgkeller left a comment

Choose a reason for hiding this comment

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

wow. damn. 5 characters

tt-aho added 3 commits July 31, 2024 22:12
… to switch to using NOC_DISPATCH_MULTICAST_WRITE_VC for all dispatch mcasts

Remove some minor riscv overhead since we don't need to create full noc addrs if we're not reprogramming the noc coords
@tt-aho tt-aho force-pushed the aho/sd-hang-fix branch from 6a2359d to dc6d581 Compare July 31, 2024 22:12
@tt-aho tt-aho merged commit 6f42f4e into main Jul 31, 2024
5 checks passed
@tt-aho tt-aho deleted the aho/sd-hang-fix branch July 31, 2024 22:18
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.

3 participants