[Bugfix] Always apply MM processor even when no MM items are passed (#26240)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/conftest.py

@@ -46,7 +46,6 @@ from vllm.connections import global_http_connection
 from vllm.distributed import (cleanup_dist_env_and_memory,
                               init_distributed_environment,
                               initialize_model_parallel)
-from vllm.inputs import TextPrompt
 from vllm.logger import init_logger
 from vllm.logprobs import Logprob
 from vllm.multimodal.utils import fetch_image
@@ -760,17 +759,24 @@ class VllmRunner:
         images: Optional[PromptImageInput] = None,
         videos: Optional[PromptVideoInput] = None,
         audios: Optional[PromptAudioInput] = None,
-    ) -> list[TextPrompt]:
-
+    ) -> list[dict[str, Any]]:
         if any(x is not None and len(x) != len(prompts)
                for x in [images, videos, audios]):
             raise ValueError(
                 "All non-None multimodal inputs must have the same length as "
                 "prompts")
 
-        inputs = []
+        inputs = list[dict[str, Any]]()
         for i, prompt in enumerate(prompts):
-            multi_modal_data = {}
+            prompt_dict = dict[str, Any]()
+            if isinstance(prompt, str):
+                prompt_dict["prompt"] = prompt
+            elif isinstance(prompt, list):
+                prompt_dict["prompt_token_ids"] = prompt
+            else:
+                prompt_dict["prompt_embeds"] = prompt
+
+            multi_modal_data = dict[str, Any]()
             if images is not None and (image := images[i]) is not None:
                 multi_modal_data["image"] = image
             if videos is not None and (video := videos[i]) is not None:
@@ -778,17 +784,10 @@ class VllmRunner:
             if audios is not None and (audio := audios[i]) is not None:
                 multi_modal_data["audio"] = audio
 
-            text_prompt_kwargs: dict[str, Any] = {
-                "multi_modal_data": multi_modal_data or None
-            }
-            if isinstance(prompt, str):
-                text_prompt_kwargs["prompt"] = prompt
-            elif isinstance(prompt, list):
-                text_prompt_kwargs["prompt_token_ids"] = prompt
-            else:
-                text_prompt_kwargs["prompt_embeds"] = prompt
+            if multi_modal_data:
+                prompt_dict["multi_modal_data"] = multi_modal_data
 
-            inputs.append(TextPrompt(**text_prompt_kwargs))
+            inputs.append(prompt_dict)
 
         return inputs
 

tests/test_inputs.py

@@ -3,8 +3,11 @@
 
 import pytest
 
+from vllm.config import ModelConfig
 from vllm.inputs import zip_enc_dec_prompts
 from vllm.inputs.parse import parse_raw_prompts
+from vllm.inputs.preprocess import InputPreprocessor
+from vllm.transformers_utils.tokenizer import init_tokenizer_from_configs
 
 pytestmark = pytest.mark.cpu_test
 
@@ -80,3 +83,50 @@ def test_zip_enc_dec_prompts(mm_processor_kwargs, expected_mm_kwargs):
         assert zipped['encoder_prompt'] == enc
         assert zipped['decoder_prompt'] == dec
         assert zipped['mm_processor_kwargs'] == exp_kwargs
+
+
+@pytest.mark.parametrize("model_id", [
+    "facebook/opt-125m",
+])
+@pytest.mark.parametrize("prompt", [
+    {
+        "prompt": "",
+        "multi_modal_data": {
+            "dummy": []
+        },
+    },
+    {
+        "prompt_token_ids": [],
+        "multi_modal_data": {
+            "dummy": []
+        },
+    },
+])
+def test_preprocessor_text_no_mm_inputs(model_id, prompt):
+    model_config = ModelConfig(model=model_id)
+    tokenizer = init_tokenizer_from_configs(model_config)
+    input_preprocessor = InputPreprocessor(model_config, tokenizer)
+
+    with pytest.raises(ValueError, match="does not support multimodal inputs"):
+        input_preprocessor.preprocess(prompt)
+
+
+@pytest.mark.parametrize("model_id", [
+    "facebook/chameleon-7b",
+])
+@pytest.mark.parametrize("prompt", [
+    "",
+    {
+        "prompt_token_ids": []
+    },
+])
+def test_preprocessor_always_mm_code_path(model_id, prompt):
+    model_config = ModelConfig(model=model_id)
+    tokenizer = init_tokenizer_from_configs(model_config)
+    input_preprocessor = InputPreprocessor(model_config, tokenizer)
+
+    # HF processor adds sep token
+    sep_token_id = tokenizer.vocab[tokenizer.sep_token]
+
+    processed_inputs = input_preprocessor.preprocess(prompt)
+    assert sep_token_id in processed_inputs["prompt_token_ids"]