Add TMA TensorMapDescriptor support#1687

Open

rparolin wants to merge 11 commits intoNVIDIA:mainfrom

rparolin:rparolin/tma_feature

Commits on Mar 7, 2026

initial commit

rparolin
authored and
cpcloud
committed
tma wide

rparolin
authored and
cpcloud
committed
clean up

rparolin
authored and
cpcloud
committed
Add comments to prepare_tensor_map_arg explaining allocation and lifetime

authored and
cpcloud
committed
Address Copilot review feedback

authored and
cpcloud
committed
Split TMA example into two focused files

authored and
cpcloud
committed
pre-commit

rparolin
authored and
cpcloud
committed
adding stride meta data to gpu allocated memory

rparolin
authored and
cpcloud
committed
im2col fixes

rparolin
authored and
cpcloud
committed
Reuse CCCL TMA descriptor construction for tiled TensorMap and keep validated views alive to avoid DLPack-backed pointer lifetime hazards.
cpcloud
committed
Skip im2col-wide TensorMap tests when runtime support is unavailable.
cpcloud
committed