Fixed saving for models that don't have _name_or_path in config

[kovalexal]

Hi, PEFT team!

I noticed a small bug when I was trying to save LoRA for SDXL UNet model, loaded with from_single_file method.
Apparently, from_single_file does not add _name_or_path to the child model config (at least for diffusers==0.26.0).

from diffusers import StableDiffusionXLPipeline
from peft import get_peft_model, LoraConfig

pipe = StableDiffusionXLPipeline.from_single_file(
    "/data/checkpoints/civitai/sdxl/realvisxlV30Turbo_v30Bakedvae.safetensors"
)


config = LoraConfig(
    target_modules=["conv_in"],
    lora_alpha=8,
    r=8
)

pipe.unet = get_peft_model(pipe.unet, config, adapter_name="default")
pipe.unet.save_pretrained("./unet")

...
File /opt/conda/envs/peft/lib/python3.8/site-packages/peft/peft_model.py:782, in PeftModel.create_or_update_model_card(self, output_dir)
    780     model_config = model_config.to_dict()
    781 if model_config is not None:
--> 782     card.data["base_model"] = model_config["_name_or_path"]
    784 lines = card.text.splitlines()
    786 quantization_config = None

KeyError: '_name_or_path'

This specific PR fixes this issue.

@BenjaminBossan your review is kindly appreciated.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Much appreciated, LGTM.