Loading loras on "text_pipe" of a modular pipeline fails

[christopher5106]

Describe the bug

The following code works:

import torch
from diffusers import Flux2Pipeline

pipe = Flux2Pipeline.from_pretrained("black-forest-labs/FLUX.2-dev", torch_dtype=torch.bfloat16)
pipe.load_lora_weights("flux.2-turbo-lora.safetensors")

but on modular it doesn't:

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("black-forest-labs/FLUX.2-dev")
blocks = pipe.blocks
text_blocks = blocks.sub_blocks.pop("text_encoder")
text_pipe = text_blocks.init_pipeline("black-forest-labs/FLUX.2-dev")
text_pipe.load_components(torch_dtype=torch.bfloat16)
text_pipe.load_lora_weights("flux.2-turbo-lora.safetensors")
text_pipe.set_adapters(["default"], adapter_weights=[1])

The lora is found here: https://huggingface.co/fal/FLUX.2-dev-Turbo/blob/main/flux.2-turbo-lora.safetensors

Logs

src/diffusers/configuration_utils.py", line 176, in __getattr__
    raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
AttributeError: 'Flux2ModularPipeline' object has no attribute 'transformer'. Did you mean: 'transformer_name'?

System Info

Latest commit of Diffusers 71a6fd9

Who can help?

@yiyixuxu @DN6 @sayakpaul

[sayakpaul]

@yiyixuxu could you take a look?

[yiyixuxu]

@christopher5106
this is a transformer only lora, no? and text_pipe only has text_encoders
I think you only need to run load_lora_weight on denoise pipe

[christopher5106]

In the scenario where we use the modular approach to run the components in different containers, do you advise to filter the keys ? not easy since the conversion logic happens after the call to load_lora_weight() ?

[yiyixuxu]

@christopher5106
ohh make sense, maybe warn and skip is better
would you be interested in open a PR for this? totally fine if not, we will fix it on our side