fix(inference): materialise checkpoint tensors in host memory before H2D copy (17.7x faster load on GB300) - #150

Open

rwagwani wants to merge 2 commits into

NVIDIA:mainfrom

rwagwani:fix/checkpoint-load-mmap-h2d

Commits on Aug 5, 2026

fix(inference): materialise checkpoint tensors in host memory before H2D copy

authored and
root
committed
9078e51
fix(inference): apply the host-memory staging copy conditionally

authored and
root
committed
3d8c469