FP32 XL base can't be used to train with but no issues for inference.

Traceback (most recent call last):
File "F:\kohya_ss-win\sdxl_train_network.py", line 183, in
trainer.train(args)
File "F:\kohya_ss-win\train_network.py", line 224, in train
model_version, text_encoder, vae, unet = self.load_target_model(args, weight_dtype, accelerator)
File "F:\kohya_ss-win\sdxl_train_network.py", line 44, in load_target_model
) = sdxl_train_util.load_target_model(args, accelerator, sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, weight_dtype)
File "F:\kohya_ss-win\library\sdxl_train_util.py", line 34, in load_target_model
) = _load_target_model(
File "F:\kohya_ss-win\library\sdxl_train_util.py", line 75, in _load_target_model
) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype)
File "F:\kohya_ss-win\library\sdxl_model_util.py", line 270, in load_models_from_sdxl_checkpoint
info2 = _load_state_dict_on_device(text_model2, converted_sd, device=map_location) # remain fp32
File "F:\kohya_ss-win\library\sdxl_model_util.py", line 158, in _load_state_dict_on_device
raise RuntimeError("Error(s) in loading state_dict for {}:\n\t{}".format(model.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for CLIPTextModelWithProjection:
Unexpected key(s) in state_dict: "text_projection.weight.weight".
Missing key(s) in state_dict: "text_projection.weight".

Start at sdxl_train_network.py and follow train_network.py into library/sdxl_train_util.py and library/sdxl_model_util.py, especially load_models_from_sdxl_checkpoint and _load_state_dict_on_device. Reproduce the FP32 SDXL base training load and inspect the reported unexpected and missing text_projection.weight keys; done means the model loads for training without this state_dict error.

Written by the indexing model from the issue text.