Catch Spark training pitfalls before a long run.
What it does
Diagnoses DGX Spark training failures and helps check for known problems before long training runs.
When to use it
Use for launch failures, unexpected OOMs, declining throughput, long GB10 training jobs, dual-Spark parallelism decisions, or precision selection.
How to use it
Give it a DGX Spark training problem or preflight request; it checks relevant failure modes and provides diagnostic findings and fixes.
What you provide
- An existing project
Uses
ND
CU
Latest review
What you need · 4
The checks target DGX Spark's GB10 chip with SM121, 128GB unified memory, and aarch64 architecture.
The preflight script invokes Python to inspect the training environment.
Access · 8
This skill
/proc/sys/vm/drop_caches
Write
Flushes the system-wide page cache
GPU-resident processes
Write
Stops unrelated GPU servers