Skip to content

Fix miopen cache dir handling#69

Open
PatrickRMiles wants to merge 2 commits into
LBANN:mainfrom
PatrickRMiles:miles30/miopen_caching
Open

Fix miopen cache dir handling#69
PatrickRMiles wants to merge 2 commits into
LBANN:mainfrom
PatrickRMiles:miles30/miopen_caching

Conversation

@PatrickRMiles
Copy link
Copy Markdown
Collaborator

@PatrickRMiles PatrickRMiles commented May 14, 2026

torchrun-hpc uses TMPDIR to set MIOpen cache dirs, regardless of how we set those dirs outside of torchrun-hpc calls. This PR modifies the Tuolumne run scripts to set TMPDIR to a persistent location, allowing MIOpen to actually reuse cached kernel benchmarking results across runs. At scale 7, this improves warmup speed on all runs after the first by ~25%.

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.

1 participant