gpu-observability

GPU Observability with workload attribution. One OTLP agent per node ties hardware metrics (NVIDIA, AMD, Intel Gaudi) to the K8s pod or Slurm job burning the GPU.

Reproducible MiniMax-H3 benchmarks on NVIDIA GB300: SGLang scaling, host eBPF CUDA tracing, Grace-GPU PMU traffic, energy, and Nsight comparison.

Real-time NVIDIA GPU monitoring, fleet observability and LLM inference analytics for Ollama, vLLM and SGLang. Docker Compose and Kubernetes deployment.

Monitor low-utilization time, idle-state episodes, and workload starvation signals on NVIDIA datacenter GPUs.

Production profiling evidence for vLLM/SGLang deployments and the next VibeSys optimization hypothesis.

A control plane for inference memory hierarchies.