Fix issue in maybe_convert_prompt (#3188)

When the token used for textual inversion does not have any special symbols (e.g. it is not surrounded by <>), the tokenizer does not properly split the replacement tokens.  Adding a space for the padding tokens fixes this.

src/diffusers/loaders.py

@@ -410,7 +410,7 @@ class TextualInversionLoaderMixin:
                 replacement = token
                 i = 1
                 while f"{token}_{i}" in tokenizer.added_tokens_encoder:
-                    replacement += f"{token}_{i}"
+                    replacement += f" {token}_{i}"
                     i += 1
 
                 prompt = prompt.replace(token, replacement)

tests/pipelines/test_pipelines.py

@@ -541,7 +541,7 @@ class DownloadTests(unittest.TestCase):
             assert pipe.text_encoder.get_input_embeddings().weight[-3].sum().item() == 96
             assert pipe.text_encoder.get_input_embeddings().weight[-2].sum().item() == 128
             assert pipe.text_encoder.get_input_embeddings().weight[-1].sum().item() == 160
-            assert pipe._maybe_convert_prompt("<***>", pipe.tokenizer) == "<***><***>_1<***>_2"
+            assert pipe._maybe_convert_prompt("<***>", pipe.tokenizer) == "<***> <***>_1 <***>_2"
 
             prompt = "hey <***>"
             out = pipe(prompt, num_inference_steps=1, output_type="numpy").images
@@ -569,7 +569,7 @@ class DownloadTests(unittest.TestCase):
             assert pipe.text_encoder.get_input_embeddings().weight[-3].sum().item() == 96
             assert pipe.text_encoder.get_input_embeddings().weight[-2].sum().item() == 128
             assert pipe.text_encoder.get_input_embeddings().weight[-1].sum().item() == 160
-            assert pipe._maybe_convert_prompt("<****>", pipe.tokenizer) == "<****><****>_1<****>_2"
+            assert pipe._maybe_convert_prompt("<****>", pipe.tokenizer) == "<****> <****>_1 <****>_2"
 
             prompt = "hey <****>"
             out = pipe(prompt, num_inference_steps=1, output_type="numpy").images