spark-memory-thermal-ops

SUMMARY

What it does

This skill provides guidance for managing unified memory and thermal behavior on NVIDIA DGX Spark systems with a 128GB unified memory pool shared by CPU and GPU. It covers planning memory headroom before a training run, working through an OOM (out-of-memory) event using a structured ladder of remediation steps, and monitoring temperature and power during long-running jobs to distinguish thermal throttling from other issues. The skill emphasizes using `free -g` instead of `nvidia-smi` for memory planning, understanding transient memory peaks during model loading, and recognizing the sustained power cap (~100W) as normal platform behavior. It also addresses running concurrent workloads, such as a trainer alongside an inference server, with rules for uncapped vs. capped workloads. The skill references companion materials like `references/uma-accounting.md` for memory estimation and `assets/thermal-sample.sh` for thermal logging.