memstat-bench

Visualize memory usage trends during a vLLM-metal benchmark by running a sidecar logger that captures 1 Hz samples of wired memory and generates a timeline plot for visual analysis of memory shapes.

High memory pressure on macOS Metal during a vLLM benchmark is identified by capturing 1 Hz samples of memory.pressure_percent to track OS pressure levels alongside wired memory usage patterns.

The best way to compare memory shapes between two benchmark runs is to plot them side-by-side on shared axes, highlighting differences in wired memory, active pages, and total system memory usage.

Yes, you need metalstat on PATH and matplotlib in the vllm-metal environment to run the sidecar logger and generate the memory timeline plot required for memory-regression debugging.

Yes, you can capture a per-second active memory timeline for vLLM-metal performance monitoring by sampling memory.wired_gb, memory.pressure_percent, and active pages at 1 Hz to reveal memory usage patterns.

You need a per-second memory timeline instead of just latency metrics when conducting memory-pressure investigations or regression debugging, allowing you to observe memory shapes and wired memory patterns rather than latency alone.