Skip to content

LoRA loading fails with HiDream O1 #1529

@nArn0

Description

@nArn0

This is probably very dumb but a trained LoRA fails to load with HiDream O1 Image and it is probably an issue of key naming.

In the LoRA, keys are in diffusion format:

diffusion_model.language_model.layers.9.mlp.up_proj.lora_A.weight
diffusion_model.language_model.layers.9.mlp.up_proj.lora_B.weight

But sd.ccp report ignored keys (for all the keys):

[WARN ] lora.hpp:817  - unused lora tensor | lora.model.diffusion_model.language_model.layers.9.mlp.up_proj.weight.lora_up
[WARN ] lora.hpp:817  - unused lora tensor | lora.model.diffusion_model.language_model.layers.9.mlp.up_proj.weight.lora_down

If i rename all the keys in the LoRA to "model." instead of "diffusion_model.", i no longer get warning but the LoRA has no effect. 😅

If i manually merge the LoRA in the model beforehand, i get the expected result.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions