Skip to content

Feat/autotune observability#1437

Open
SamuelBelanger wants to merge 27 commits into
mainfrom
feat/autotune-observability
Open

Feat/autotune observability#1437
SamuelBelanger wants to merge 27 commits into
mainfrom
feat/autotune-observability

Conversation

@SamuelBelanger

Copy link
Copy Markdown
Contributor

Key changes

  • Add short circuit logs
  • Add recorder for autotune debugging
  • Fix throughtput kernels
  • Refactor memory and launch throughputs

ThierryCantin-Demers and others added 26 commits July 16, 2026 17:08
The peak-throughput probe allocates up to 512MB x2 in the exclusive
memory pool. The handles drop when measurement returns, but the pool
frees pages only during a cleanup pass, and CUDA runs none on the normal
path, so the buffers stayed reserved (~3.9GB on a 16GB card) for the rest
of the process. Call memory_cleanup() after the one-time measurement so
the pages are returned to the driver immediately.
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