Real-time system and GPU memory with tiered alerts

Time-to-first-token and tokens-per-second measurement

Checks if a model will fit in available memory before loading

Detects semantic drift in model outputs over time

Sustained tokens-per-second under continuous load

KV cache eviction rates and memory pressure under load

CPU/GPU temperature and throttling detection during inference

Composite health score from latency, drift, and memory metrics

Finds max batch size that fits in memory without OOM

Layer-by-layer time breakdown for a single forward pass