Skip to content

Large inference discrepancy between LoRA-finetuned Pi0.5 and full-parameter finetuned model after weight conversion via realtime-vla pipeline #35

Description

@tqb-web

Hello, thank you very much for sharing this open-source repository to speed up real-time Pi0.5 model inference! This tool has greatly streamlined my model deployment workflow.

While testing the checkpoint conversion feature, I ran into an issue I’d like to report. I fine-tuned Pi0.5 with LoRA and full parameters separately on Libero using OpenPI official code, then converted both checkpoints via the realtime-vla conversion pipeline, and observed inconsistent behaviors between the two variants:

  • Full-parameter finetuned model: inference results match original OpenPI outputs well
  • LoRA-finetuned model: inference actions differ drastically after conversion

It seems the conversion script does not properly handle LoRA merged weights. Looking forward to any fix or guidance you can provide.

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