Fix IDs shape mismatch in SFT for VLMs with text-only

tests/test_sft_trainer.py

@@ -1829,7 +1829,12 @@ def test_train_vlm_gemma_3n(self):
     )
     @pytest.mark.parametrize(
         "dataset_config",
-        ["conversational_language_modeling", "conversational_prompt_completion", "standard_prompt_completion"],
+        [
+            "conversational_language_modeling",
+            "conversational_prompt_completion",
+            "standard_language_modeling",  # Regression test for #5334
+            "standard_prompt_completion",
+        ],
     )
     @require_vision
     def test_train_vlm_text_only_data(self, model_id, dataset_config):

trl/trainer/sft_trainer.py

@@ -1167,7 +1167,10 @@ def tokenize_fn(example, processing_class, dataset_text_field, assistant_only_lo
                             processed = {k: v[0] if isinstance(v[0], list) else v for k, v in processed.items()}
                             output = {k: processed[k] for k in ("input_ids", "assistant_masks") if k in processed}
                         else:
-                            output = {"input_ids": processing_class(text=example[dataset_text_field])["input_ids"]}
+                            # Fix transformers inconsistency: for VLMs, processing_class returns lists of lists
+                            # even for single examples, while for LLMs it returns lists of ints.
+                            ids = processing_class(text=example[dataset_text_field])["input_ids"]
+                            output = {"input_ids": ids[0] if isinstance(ids[0], list) else ids}
 
                     if "assistant_masks" in output and 1 not in output["assistant_masks"]:
                         raise RuntimeError(