value errors in convert to/from diffusers from original stable diffusion

Describe the bug

There's a hardcode somewhere for 77 tokens, when it should be using the dimensions of what is actually in the model.

I have a diffusers-layout SD1.5 model, with LongCLIP.

https://huggingface.co/opendiffusionai/xllsd-alpha0

I can pull it locally, then convert to single file format, with

python convert_diffusers_to_original_stable_diffusion.py
--use_safetensors
--model_path $SRCM
--checkpoint_path $DESTM

But then if I try to convert it back, I get size errors for the text encoder not being 77 size.

I should point out that the model WORKS PROPERLY for diffusion, when loaded in diffusers format, so I dont have some funky broken model here.

Reproduction

from transformers import CLIPTextModel, CLIPTokenizer

from diffusers import StableDiffusionPipeline, AutoencoderKL
import torch

pipe = StableDiffusionPipeline.from_single_file(
"XLLsd-phase0.safetensors",
torch_dtype=torch.float32,
use_safetensors=True)

outname = "XLLsd_recreate"
pipe.save_pretrained(outname, safe_serialization=False)

Logs

venv/lib/python3.12/site-packages/diffusers/models/model_loading_utils.py", line 230, in load_model_dict_into_meta
    raise ValueError(
ValueError: Cannot load  because text_model.embeddings.position_embedding.weight expected shape torch.Size([77, 768]), but got torch.Size([248, 768]). If you want to instead overwrite randomly initialized weights, please make sure to pass both `low_cpu_mem_usage=False` and `ignore_mismatched_sizes=True`. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example.

System Info

• 🤗 Diffusers version: 0.32.2
• Platform: Linux-6.8.0-55-generic-x86_64-with-glibc2.39
• Running on Google Colab?: No
• Python version: 3.12.3
• PyTorch version (GPU?): 2.6.0+cu124 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.29.3
• Transformers version: 4.50.0
• Accelerate version: 1.5.2
• PEFT version: not installed
• Bitsandbytes version: 0.45.2
• Safetensors version: 0.5.3
• xFormers version: not installed
• Accelerator: NVIDIA GeForce RTX 4090, 24564 MiB

Who can help?

No response

Commencez par convert_diffusers_to_original_stable_diffusion.py et retracez le chemin de conversion utilisé par les deux commandes, puis examinez le chemin de chargement de l’encodeur de texte indiqué par model_loading_utils.py dans l’erreur. Reproduisez l’aller-retour LongCLIP signalé et confirmez que la conversion utilise les dimensions réelles des embeddings de position du modèle, au lieu d’échouer à cause d’une hypothèse de taille fixe.