Fix single letter stop strings (#31448)

* Fix single letter stop strings

* Change the 0 to a 1 to avoid potential empty vector headaches later

* Restructure for clarity

* Update tests/generation/test_stopping_criteria.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Add the unsqueeze

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

tests/generation/test_stopping_criteria.py

@@ -208,6 +208,24 @@ class StoppingCriteriaTestCase(unittest.TestCase):
         token_lengths = embedding_vec[:, 2].tolist()
         self.assertEqual(token_lengths, [len(token) for token in token_list])
 
+    def test_single_letter_stop_string(self):
+        true_strings = ["a", "baa", "abc"]  # "abc" is a single token
+        false_strings = ["abbbbbbb", "b"]  # "abbbbbbb" is split into multiple tokens
+        stop_strings = ["a"]
+        tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
+        tokenizer.pad_token_id = tokenizer.eos_token_id
+        tokenizer.padding_side = "left"
+
+        true_input_ids = tokenizer(true_strings, return_tensors="pt", padding="longest", add_special_tokens=False)
+        false_input_ids = tokenizer(false_strings, return_tensors="pt", padding="longest", add_special_tokens=False)
+
+        scores = None
+        criteria = StopStringCriteria(tokenizer=tokenizer, stop_strings=stop_strings)
+        for input_ids in true_input_ids["input_ids"]:
+            self.assertTrue(criteria(input_ids.unsqueeze(0), scores))
+        for input_ids in false_input_ids["input_ids"]:
+            self.assertFalse(criteria(input_ids.unsqueeze(0), scores))
+
     def test_criterias_per_row(self):
         text = "They completed the challenging puzzle, revealing the hidden image at the end"
         stop_strings = ["end"]