Skip to content

【FR】Bug: attention/rms_norm/silu/rotary operators need reference routing on nvidia (openbmb/MiniCPM4.1-8B) #4981

Description

@buzhengjing

Bug Report: 【FR】attention/rms_norm/silu/rotary operators need reference routing for accuracy on nvidia (openbmb/MiniCPM4.1-8B)

Description

On model openbmb/MiniCPM4.1-8B under FlagOS, four operators had insufficient FlagOS
accuracy. After routing them to reference (via nvidia.yaml op_backends), V3 GPQA = 58.0%,
above the NV baseline 54.0% (rel_drop -7.41%), passed.

Note: operator control in plugin mode is via vllm-plugin-FL nvidia.yaml op_backends, not gems.txt.

Environment

Item Value
Hardware NVIDIA H20-3e
PyTorch 2.11.0+cu130
vLLM 0.20.2
FlagGems 5.0.2
vllm-plugin-FL 0.2.0rc2.post1

Accuracy Data

  • First probe 48% (temp=0.6 chain-of-thought noise) -> retest V3 = 58%, passed
  • Fixed config: 4 operators dispatch -> reference first

Affected Operators

attention_backend, rms_norm, silu_and_mul, rotary_embedding
(FlagOS implementation accuracy insufficient; passes after routing to reference)

Status

  • Resolved via nvidia.yaml dispatch tuning; V3 passed and released.
  • Filed as an accuracy-improvement record for these 4 FlagGems operators; may be closed
    once FlagGems side confirms.

Possible Directions

  • Review FlagOS precision of rms_norm / rotary_embedding / silu_and_mul / attention

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions