Skip to content

Bucket/coalesced tensor broadcast in non-colocated refit #1286

New issue
New issue
Closed
#1313
Closed
Bucket/coalesced tensor broadcast in non-colocated refit#1286
#1313
Assignees
youngeunkwon0405
Labels
PerformanceRelated to improving performanceasyncRLdeepseekRelated to deepseek 671benhancementNew feature or requestqa_rcca_donewhen RCCA finished for the issue, the qa will mark with this label .

Description

@guyueh1
guyueh1
opened on Oct 6, 2025

Is your feature request related to a problem? Please describe.
For MoE models like qwen 30B or deepseek, need coalesced/bucketized tensor broadcast in non-colocated refit to boost the broadcast effective bandwidth.

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

Activity

  1. guyueh1
    added
    enhancementNew feature or request
    PerformanceRelated to improving performance
    asyncRL
    deepseekRelated to deepseek 671b
    on Oct 6, 2025
  2. guyueh1
    added a parent issue on Oct 6, 2025

  3. guyueh1 commented on Oct 6, 2025

    @guyueh1
    guyueh1
    on Oct 6, 2025
    ContributorAuthor

    @youngeunkwon0405 fyi

  4. guyueh1
    assigned
    guyueh1
    and
    youngeunkwon0405
    on Oct 6, 2025

  5. youngeunkwon0405 commented on Oct 7, 2025

    @youngeunkwon0405
    youngeunkwon0405
    on Oct 7, 2025
    Contributor

    The implementation is in my local branch. The implementation depends on #1264.
    Here are the high-level perf improvements.

    Qwen3 30B A3B non-colocated refit

  6. guyueh1
    removed their assignment
    on Oct 7, 2025
  7. guyueh1
    linked a pull request that will close this issuefeat: tensor packing and batching for non-colocated refit performance #1313on Oct 9, 2025
  8. terrykong
    closed this as completedin #1313on Oct 12, 2025
  9. qiaochuz-nv
    added
    qa_rcca_donewhen RCCA finished for the issue, the qa will mark with this label .
    on May 5, 2026

  10. qiaochuz-nv commented on May 5, 2026

    @qiaochuz-nv
    qiaochuz-nv
    on May 5, 2026
    Contributor

    QA RCCA Analysis

    RCCA Status: Not Required
    RCCA Category: not_required
    Test Action: none

    Fix PR: N/A
    Root Cause: This issue is not an actionable closed product-code defect for RCCA; it is enhancement/feature tracking, docs/cosmetic work, placeholder/test issue, or closed non-bug work.
    Trigger Config: N/A.

    Coverage Assessment:

    Area Result
    Existing Test N/A
    Coverage Status NOT_REQUIRED
    Gap N/A
    Corrective Action No RCCA regression test required.

    Conclusion: RCCA not required because this does not represent a QA coverage escape for a fixed NeMo-RL bug.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

PerformanceRelated to improving performanceasyncRLdeepseekRelated to deepseek 671benhancementNew feature or requestqa_rcca_donewhen RCCA finished for the issue, the qa will mark with this label .

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions