fix roundtript text_encoder

utils/convert_original_stable_diffusion_to_diffusers.py

@@ -746,8 +746,6 @@ def convert_open_clip_checkpoint(checkpoint):
 
     text_model_dict = {}
 
-    d_model = int(checkpoint["cond_stage_model.model.text_projection"].shape[0])
-
     text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids")
 
     for key in keys:
@@ -756,6 +754,7 @@ def convert_open_clip_checkpoint(checkpoint):
         if key in textenc_conversion_map:
             text_model_dict[textenc_conversion_map[key]] = checkpoint[key]
         if key.startswith("cond_stage_model.model.transformer."):
+            d_model = int(checkpoint[key].shape[0]/3)
             new_key = key[len("cond_stage_model.model.transformer.") :]
             if new_key.endswith(".in_proj_weight"):
                 new_key = new_key[: -len(".in_proj_weight")]