Dist Lifecycle

Stand up the distributed-model lifecycle for a CP training/inference run: device placement before DTensor wrapping, the all-trainable-params-are-DTensors invariant (with placeholder/freeze for unimplemented modules), checkpoint save (DTensor → plain) and load (realign via the live state_dict template + redistribute optimizer state to parameter placements), resume RNG seed offset, and DTensor-safe EMA. Use after dtensor_modules + build_infra, when wiring the ported CP modules into a real trainer/predictor (Lightning or custom) and you need checkpoints, resume, or EMA to work — the gap between "modules pass parity" and "training runs and resumes".

nvidia-bionemo Updated 508 repo stars

File contents

nvidia-bionemo/boltz-cp commit 6b00bbc3c5

Frequently asked questions

npx get-skill add nvidia-bionemo/dist-lifecycle