Add TMA TensorMapDescriptor support#1687

Open

rparolin wants to merge 19 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
Align TensorMap API surface with review feedback and enforce context safety.
cpcloud
committed
Restore cu12 feature definitions in cuda_core pixi manifest.
cpcloud
committed

Commits on Mar 10, 2026

Handle TensorMap device validation by DLPack type
rparolin
committed

Commits on Mar 11, 2026

Commits on Mar 13, 2026

Update cuda_core/cuda/core/_cpp/tensor_map_cccl.h

rparolin
and
leofang
authored

Commits on Mar 14, 2026