Skip to content

fix(power): publish module power through a transient GPU disagreement - #619

Open
edwingao28 wants to merge 1 commit into
NVIDIA:mainfrom
edwingao28:fix/module-power-gpu-disagreement
Open

edwingao28 wants to merge 1 commit into
NVIDIA:mainfrom
edwingao28:fix/module-power-gpu-disagreement

Conversation

@edwingao28

Copy link
Copy Markdown
Contributor

The NVML module reader omits a socket whose GPUs' readings differ by over 5%. Load steps split them briefly, enough for InferenceX's 3 s gap rule to drop module_w. Publish the pair mean through 3 consecutive disagreeing reads; withhold from the 4th. Startup disables module power only when no GPU reports; a lone GPU cannot end a disagreement; only withholding warns.

Testing: fmt, clippy, unit tests, strict mkdocs pass; new tests fail on the old rule. Startup, lone-GPU and logging changes are unit-tested only; hardware runs predate them. GB300 GEMM bursts vs v2.59.3: gaps over 3 s 7 → 0. Replaying sweep 39733456758: CI kept module_w on 12 of 22 benchmark points, the patched rule on all 22; 592/593 disagreements lasted ≤3 reads.

Limits: A 1 s on/off load phase-locked to the poll blanks both builds (45-60 disagreeing reads). Disagreements of 5+ reads still leave ≥3 s gaps.

Publish a socket's mean through up to 3 consecutive disagreeing reads and
withhold it from the 4th. In one GB200/GB300 CI sweep, 592 of 593
disagreements lasted 3 reads or fewer.

Signed-off-by: Wenyao Gao <wenyao.gao28@gmail.com>

This branch has not been deployed

No deployments
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