[Bugfix] Do not skip "empty" parts of chats that are parsable (#16219)

Signed-off-by: mgoin <[email protected]>

Author: mgoin

tests/entrypoints/test_chat_utils.py

@@ -30,6 +30,7 @@
 MLLAMA_MODEL_ID = "meta-llama/Llama-3.2-11B-Vision-Instruct"
 LLAMA_GUARD_MODEL_ID = "meta-llama/Llama-Guard-3-1B"
 HERMES_MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
+MISTRAL_MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
 
 
 @pytest.fixture(scope="function")
@@ -80,6 +81,30 @@ def mllama_tokenizer():
     )
 
 
+@pytest.fixture(scope="function")
+def mistral_model_config():
+    return ModelConfig(MISTRAL_MODEL_ID,
+                       task="generate",
+                       tokenizer=MISTRAL_MODEL_ID,
+                       tokenizer_mode="auto",
+                       trust_remote_code=True,
+                       dtype="auto",
+                       seed=0,
+                       limit_mm_per_prompt={
+                           "image": 2,
+                       })
+
+
+@pytest.fixture(scope="module")
+def mistral_tokenizer():
+    return TokenizerGroup(
+        tokenizer_id=MISTRAL_MODEL_ID,
+        enable_lora=False,
+        max_num_seqs=5,
+        max_input_length=None,
+    )
+
+
 @pytest.fixture(scope="module")
 def image_url():
     image = ImageAsset('cherry_blossom')
@@ -131,6 +156,66 @@ def test_parse_chat_messages_single_image(
     _assert_mm_data_is_image_input(mm_data, 1)
 
 
+def test_parse_chat_messages_empty_system(
+    mistral_model_config,
+    mistral_tokenizer,
+):
+    # Test string format
+    conversation, _ = parse_chat_messages(
+        [{
+            "role": "system",
+            "content": ""
+        }, {
+            "role": "user",
+            "content": [{
+                "type": "text",
+                "text": "Who are you?"
+            }]
+        }],
+        mistral_model_config,
+        mistral_tokenizer,
+        content_format="string",
+    )
+    assert conversation == [{
+        "role": "system",
+        "content": ""
+    }, {
+        "role": "user",
+        "content": "Who are you?"
+    }]
+
+    # Test openai format
+    conversation, _ = parse_chat_messages(
+        [{
+            "role": "system",
+            "content": ""
+        }, {
+            "role": "user",
+            "content": [{
+                "type": "text",
+                "text": "Who are you?"
+            }]
+        }],
+        mistral_model_config,
+        mistral_tokenizer,
+        content_format="openai",
+    )
+    assert conversation == [{
+        "role": "system",
+        "content": [{
+            "type": "text",
+            "text": ""
+        }]
+    }, {
+        "role":
+        "user",
+        "content": [{
+            "type": "text",
+            "text": "Who are you?"
+        }]
+    }]
+
+
 @pytest.mark.asyncio
 async def test_parse_chat_messages_single_image_async(
     phi3v_model_config,
@@ -671,7 +756,7 @@ def get_conversation(is_hf: bool):
     # Build a config for the model
     model_config = ModelConfig(model,
                                task="generate",
-                               tokenizer=MLLAMA_MODEL_ID,
+                               tokenizer=model,
                                tokenizer_mode="auto",
                                trust_remote_code=True,
                                dtype="auto",
@@ -682,7 +767,7 @@ def get_conversation(is_hf: bool):
 
     # Build the tokenizer group and grab the underlying tokenizer
     tokenizer_group = TokenizerGroup(
-        MLLAMA_MODEL_ID,
+        model,
         enable_lora=False,
         max_num_seqs=5,
         max_input_length=None,

vllm/entrypoints/chat_utils.py

@@ -872,19 +872,19 @@ def _get_full_multimodal_text_prompt(placeholder_counts: dict[str, int],
     Callable[[ChatCompletionContentPartParam], _ContentPart],
 ] = {
     "text":
-    lambda part: _TextParser(part).get("text", ""),
+    lambda part: _TextParser(part).get("text", None),
     "image_url":
-    lambda part: _ImageParser(part).get("image_url", {}).get("url", ""),
+    lambda part: _ImageParser(part).get("image_url", {}).get("url", None),
     "image_embeds":
-    lambda part: _ImageEmbedsParser(part).get("image_embeds", {}),
+    lambda part: _ImageEmbedsParser(part).get("image_embeds", None),
     "audio_url":
-    lambda part: _AudioParser(part).get("audio_url", {}).get("url", ""),
+    lambda part: _AudioParser(part).get("audio_url", {}).get("url", None),
     "input_audio":
-    lambda part: _InputAudioParser(part).get("input_audio", {}),
+    lambda part: _InputAudioParser(part).get("input_audio", None),
     "refusal":
-    lambda part: _RefusalParser(part).get("refusal", ""),
+    lambda part: _RefusalParser(part).get("refusal", None),
     "video_url":
-    lambda part: _VideoParser(part).get("video_url", {}).get("url", ""),
+    lambda part: _VideoParser(part).get("video_url", {}).get("url", None),
 }
 
 
@@ -1003,11 +1003,11 @@ def _parse_chat_message_content_part(
     part_type, content = _parse_chat_message_content_mm_part(part)
 
     # if part_type is text/refusal/image_url/audio_url/video_url/input_audio but
-    # content is empty, log a warning and skip
-    if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and not content:
+    # content is None, log a warning and skip
+    if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and content is None:
         logger.warning(
-            "Skipping multimodal part (type: '%s') "
-            "with empty / unparsable content.", part_type)
+            "Skipping multimodal part '%s' (type: '%s') "
+            "with empty / unparsable content.", part, part_type)
         return None
 
     if part_type in ("text", "refusal"):