Zheng Hao Tan a901eeb211 [Benchmarking] Aggregate GPU stats in console output ...

The serving benchmark console now prints GPU memory and utilization
through the same percentile table as latency and throughput, instead
of three lines per GPU. The section title states that the rows are
aggregated across N devices, and a note points at the result JSON
for per-GPU series. Those lists stay in result_groups.gpu_stats, so
machine consumers are unchanged.

MODULAR_ORIG_COMMIT_REV_ID: e0141c249236161d66a1b4dc17df7943c23f6217

2026-08-22 20:08:05 +00:00

..

references

[Benchmarking] Aggregate GPU stats in console output

2026-08-22 20:08:05 +00:00

SKILL.md

[Skills] Add benchmark-model skill for measuring served model

2026-07-30 15:19:52 +00:00