Skip to content

[RFC][DSV4][ROCm][P6] Output and gradient combine #532

Open
Open
[RFC][DSV4][ROCm][P6] Output and gradient combine#532
Labels
DSv4deepseek-P6platform: rocmSpecific tasks specific to AMD graphics cards (such as CK, bpreshuffle/FA)type: designIssues requiring in-depth discussion of architecture design

Description

@inaniloquentee
inaniloquentee
opened on Oct 10, 2026

P1-P7 are source module IDs, not priorities.

1. Fixed-order expert output combine

Anchor RFC: Fixed-order expert output combine (P6-A)

2. Input gradients, multi-GPU integrity, and regression

Anchor RFC: Input gradients, multi-GPU integrity, and regression (P6-B)

Exit

Fused and reference combines match with real distributed results.

Activity

  1. inaniloquentee
    added
    platform: rocmSpecific tasks specific to AMD graphics cards (such as CK, bpreshuffle/FA)
    type: designIssues requiring in-depth discussion of architecture design
    DSv4
    deepseek-P6
    on Oct 10, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    DSv4deepseek-P6platform: rocmSpecific tasks specific to AMD graphics cards (such as CK, bpreshuffle/FA)type: designIssues requiring in-depth discussion of architecture design

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions