[STF] Add structured tensor partitions and placement evaluation to the places layer#9804

Merged

caugonnet merged 19 commits into

NVIDIA:mainfrom

caugonnet:stf-cute-partitions

Jul 16, 2026

Commits on Jul 16, 2026

[places] Add geometry-aware allocation and returnable placement stats

caugonnet
and
claude
committed
[places] Add cute_partition: JAX-like structured tensor partitions

caugonnet
and
claude
committed
[places] Test placement facilities and add a partitioned AXPY example

caugonnet
and
claude
committed
[places] Expose per-grid-position bytes in placement statistics

caugonnet
and
claude
committed
[places] Name the geometry-aware allocation allocate_nd

caugonnet
and
claude
committed
[places] Harden placement APIs (review findings)

caugonnet
and
claude
committed
[places] Drop the hand-rolled VMM check from the partitioned example

caugonnet
and
claude
committed
[places] Use all_devices and per-place launches in the partitioned example

caugonnet
and
claude
committed
[places] Skip trailing empty places in the partitioned AXPY loop

caugonnet
and
claude
committed
[places] Migrate the Thrust allocator example to allocate_nd

caugonnet
and
claude
committed
[places] Store partition leaves in fixed-capacity device-ready arrays

caugonnet
and
claude
committed
[places] Document structured partitions

caugonnet
and
claude
committed
[places] Qualify the raw-byte allocate_nd form in the docs

caugonnet
and
claude
committed
[STF] Include <cmath> in the partitioned_axpy example

caugonnet
and
claude
committed
[places] Document that row-major front-ends reverse the whole description

caugonnet
and
claude
committed
[places] Remove inaccurate JAX partition analogy
caugonnet
committed
[places] Reject partitions that leave grid places unused
caugonnet
committed
[places] Reformat partition validation test
caugonnet
committed
Merge branch 'main' into stf-cute-partitions
caugonnet
authored

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[STF] Add structured tensor partitions and placement evaluation to the places layer#9804

[STF] Add structured tensor partitions and placement evaluation to the places layer#9804
caugonnet merged 19 commits into
NVIDIA:mainfrom
caugonnet:stf-cute-partitions

Commits on Jul 16, 2026

[places] Add geometry-aware allocation and returnable placement stats

[places] Add cute_partition: JAX-like structured tensor partitions

[places] Test placement facilities and add a partitioned AXPY example

[places] Expose per-grid-position bytes in placement statistics

[places] Name the geometry-aware allocation allocate_nd

[places] Harden placement APIs (review findings)

[places] Drop the hand-rolled VMM check from the partitioned example

[places] Use all_devices and per-place launches in the partitioned example

[places] Skip trailing empty places in the partitioned AXPY loop

[places] Migrate the Thrust allocator example to allocate_nd

[places] Store partition leaves in fixed-capacity device-ready arrays

[places] Document structured partitions

[places] Qualify the raw-byte allocate_nd form in the docs

[STF] Include <cmath> in the partitioned_axpy example

[places] Document that row-major front-ends reverse the whole description

[places] Remove inaccurate JAX partition analogy

[places] Reject partitions that leave grid places unused

[places] Reformat partition validation test

Merge branch 'main' into stf-cute-partitions