Reduce training memory with targeted tuning.
What it does
Reduce GPU memory pressure during training with targeted allocator, parallelism, and activation-memory adjustments.
When to use it
Use when training runs encounter GPU OOM errors or approach the memory limit, including LoRA training with sequence parallelism.
How to use it
Give it a training project with GPU memory pressure; it tunes the launch environment and applicable training configuration.
What you provide
- An existing project
Latest review
What you need · 3
The tuning techniques apply to an existing Megatron Bridge training setup.
The first tuning step configures PyTorch's CUDA allocator to use expandable memory segments.
The workflow targets GPU training memory and CUDA allocator fragmentation.
Access · 0
No additional actions listed.