Skip to content

EPIC: Support TMA descriptor #199

New issue
New issue
Closed
Epic
1 of 1 issue completed
#1687
Closed
EPIC: Support TMA descriptor#199
Epic
1 of 1 issue completed
#1687
Assignees
cpcloudrparolin
Labels
P1Medium priority - Should docuda.coreEverything related to the cuda.core modulefeatureNew feature or request
Milestone
cuda.core v0.7.0

Description

@leofang
leofang
opened on Oct 30, 2024

Initializing a TMA descriptor through the driver APIs
https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TENSOR__MEMORY.html
is really tedious and error prone. We need a way to abstract it out, which aligns well with the mission of cuda.core. This also allows JIT compilers to easier consume and incorporate into the compilation pipelines.

In my understanding there are two (implicit?) requirements for this to be useful:

  1. Creating/initializing a TMA object on host
  2. Passing the object to the cuda.core.launch() API as a kernel arg

Activity

  1. leofang commented on Oct 30, 2024

    @leofang
    leofang
    on Oct 30, 2024
    MemberAuthor

    (Tentatively set to beta 3.)

  2. leofang
    added
    P0High priority - Must do!
    featureNew feature or request
    cuda.coreEverything related to the cuda.core module
    on Oct 30, 2024
  3. leofang
    added this to the cuda.core beta 3 milestone on Oct 30, 2024
  4. leofang
    added
    P1Medium priority - Should do
    and removed
    P0High priority - Must do!
    on Jan 14, 2025

  5. leofang commented on Jan 14, 2025

    @leofang
    leofang
    on Jan 14, 2025
    MemberAuthor

    Deprioritizing this to P1 and moving out of beta 3.

  6. leofang
    modified the milestones: cuda.core beta 3, cuda.core beta 4 on Jan 14, 2025
  7. leofang
    modified the milestones: cuda.core beta 4, cuda.core parking lot on Apr 26, 2025
  8. rparolin
    added the
    Epic
    issue type on Sep 24, 2025

  9. leofang commented on Oct 23, 2025

    @leofang
    leofang
    on Oct 23, 2025
    MemberAuthor

    @fbusato is working on DLPack -> TensorMap conversion on the C++ side: NVIDIA/cccl#6237. I envision that once it's done, we can easily expose this functionality through StridedMemoryView, which already supports DLPack.

  10. rparolin
    modified the milestones: cuda.core backlog, cuda.core next on Feb 6, 2026
  11. rparolin
    assigned
    leofang
    on Feb 17, 2026
  12. rparolin
    assigned
    rparolin
    and
    cpcloud
    and unassigned
    leofang
    on Feb 25, 2026
  13. rparolin
    closed this as completedin #1687on Mar 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

P1Medium priority - Should docuda.coreEverything related to the cuda.core modulefeatureNew feature or request

Type

Epic

Projects

No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions