Describe the bug

Qwen-Image-2.1 LoRAs trained with ComfyUI or ai-toolkit target the fused SwiGLU layer img_mlp.gate_up (lora_A [r, 4096], lora_B [24576, r]). QwenImage21Transformer2DModel keeps that layer as two, img_mlp.gate_layer and img_mlp.proj. _convert_non_diffusers_qwen_lora_to_diffusers passes the gate_up keys through unchanged, PEFT drops them as unexpected keys with only a warning, and the LoRA loads without its MLP gate/up update and without any error.

Affected LoRAs include ausboss/Qwen-Image-2.1-Outpaint-LoRA, Danrisi/grainscape_qwen2.1, lilylilith/QI_2.1_AnyAngle and ML-Intern-lab/Qwen-Image-2.1-doodle-in-LoRA (whose authors publish a hand-split copy as a workaround).

Reproduction

from diffusers import QwenImage21Pipeline

sd = QwenImage21Pipeline.lora_state_dict(
    "ausboss/Qwen-Image-2.1-Outpaint-LoRA",
    weight_name="qwen-image-2.1-outpaint-v2.safetensors",
)
print(sum("gate_up" in k for k in sd), "of", len(sd))

Logs

64 of 384
# e.g. transformer.transformer_blocks.0.img_mlp.gate_up.lora_A.weight - no module of that name exists, so load_lora_weights drops all 64

System Info

diffusers main at 8b33bfc, Python 3.14, torch 2.14.1, peft 0.21.2.

Who can help?

@sayakpaul