Reduce training overhead with validated graph replay.
What it does
Configure CUDA graph capture and validate replay performance to reduce training host overhead.
When to use it
Use when reducing host-driver overhead with CUDA graphs or investigating a crash or regression after changing graph configuration.
How to use it
Give it a training project to optimize; it configures CUDA graph capture and checks replay performance against eager execution.
What you provide
- An existing project
What you need · 4
Requires Transformer Engine's RNG tracker, with both model.use_te_rng_tracker and rng.te_rng_tracker enabled.
Requires a GPU for the functional CUDA graph training smoke test.
Uses Python to run the performance harness and verification commands; no Python version is specified.
Uses pytest for the prescribed unit-test verification.
Access · 0
No additional actions listed.