Self-hosted alternatives to spark-memory-thermal-ops

S

spark-memory-thermal-ops

Manage unified memory

spark-memory-thermal-ops is an agent skill from the wshobson/agents collection for managing unified memory and thermal behavior during long machine learning jobs on NVIDIA DGX Spark. The DGX Spark's GB10 chip has one 128 GB unified memory pool shared by the CPU and GPU, and a sustained power ceiling well below its rated figure. Both break assumptions from discrete-GPU machines: the headroom reported by standard GPU tools is not the whole story, and a run that starts fast can slow down partway through with nothing misconfigured, because the chip heats up and its power limit takes effect. The skill covers planning memory headroom before a training run, working through an actual out-of-memory failure, and watching thermals across a long job so slowdowns can be explained. It includes a quick reference of common issues. It assumes the job starts, and launch-time failures are covered by the separate spark-training-gotchas skill. It applies when planning memory for a run on GB10, when a job runs out of memory, and when throughput drops during a long job. The guidance is diagnostic: the agent reads the numbers, identifies whether memory or heat is the constraint, and suggests changes to the batch size, the sequence length or the job schedule. The changes should be tested on a short run before a multi-hour job is started. The skill is installed with the skills CLI from the wshobson/agents repository. Its skills.sh page showed about 2.3K installs, with Gen Agent Trust Hub and Socket passing and Snyk showing a warning at the time of the scan.