Reduce training memory with the right CPU offload settings.
What it does
Helps choose, configure, and validate CPU offloading to reduce GPU memory pressure in Megatron Bridge training.
When to use it
Use when enabling CPU offloading to reduce GPU memory or diagnosing offloading configuration changes that caused an OOM or crash.
How to use it
Give it a CPU offloading configuration or troubleshooting question; it returns configuration instructions, shell commands, and analysis.
What you need · 4
The configuration guidance applies to Megatron Bridge training workloads.
The runnable training command and unit-test commands use Python; no version is specified.
The provided training and verification commands use uv to run Python.
The offloading mechanisms move training activations or optimizer states from GPU memory to CPU memory.
Access · 0
No additional actions listed.