Make the text projection saved in the checkpoint the right format.

comfy/diffusers_convert.py

@@ -237,6 +237,10 @@ def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
             capture_qkv_bias[k_pre][code2idx[k_code]] = v
             continue
 
+        text_proj = "transformer.text_projection.weight"
+        if k.endswith(text_proj):
+            new_state_dict[k.replace(text_proj, "text_projection")] = v.transpose(0, 1).contiguous()
+
         relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
         new_state_dict[relabelled_key] = v
 

comfy/utils.py

@@ -110,7 +110,7 @@ def clip_text_transformers_convert(sd, prefix_from, prefix_to):
 
     tp = "{}text_projection".format(prefix_from)
     if tp in sd:
-        sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp).transpose(0, 1)
+        sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp).transpose(0, 1).contiguous()
     return sd