[Bugfix][Multimodal] Preserve DeepSeek V4 image block spacing (#56882)

Signed-off-by: Xucheng Zhou <aden1350@outlook.com>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

tests/tokenizers_/fixtures/deepseek_v4/test_input_5.json

@@ -0,0 +1,39 @@
+{
+    "thinking_mode": "thinking",
+    "reasoning_effort": "max",
+    "messages": [
+        {
+            "role": "system",
+            "content": "You are a helpful vision assistant."
+        },
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "text",
+                    "text": "请按“第一张、第二张”的顺序回答：第一张图"
+                },
+                {
+                    "type": "image_url",
+                    "image_url": {
+                        "url": "examples/images/carrots.jpeg"
+                    }
+                },
+                {
+                    "type": "text",
+                    "text": "和第二张图"
+                },
+                {
+                    "type": "image_url",
+                    "image_url": {
+                        "url": "examples/images/corn.jpeg"
+                    }
+                },
+                {
+                    "type": "text",
+                    "text": "中分别是什么食材？它们通常食用的部位分别是什么？"
+                }
+            ]
+        }
+    ]
+}

tests/tokenizers_/fixtures/deepseek_v4/test_output_5.txt

@@ -0,0 +1,9 @@
+<｜begin▁of▁sentence｜>You are a helpful vision assistant.<｜User｜>请按“第一张、第二张”的顺序回答：第一张图
+
+<｜deepseek_image｜>
+
+和第二张图
+
+<｜deepseek_image｜>
+
+中分别是什么食材？它们通常食用的部位分别是什么？<｜Assistant｜></think>

tests/tokenizers_/test_deepseek_v4.py

@@ -54,13 +54,9 @@ def _load_reference_case(case_id: int):
 
 def _render_reference_case(case_id: int, **kwargs):
     messages, tools = _load_reference_case(case_id)
-    conversation, _, _ = parse_chat_messages(
-        messages,
-        _model_config(),
-        content_format="string",
-    )
+    # Preserve content blocks for encoding without fetching fixture media.
     return _tokenizer().apply_chat_template(
-        conversation=conversation,
+        conversation=messages,
         messages=messages,
         tools=tools,
         tokenize=False,
@@ -397,6 +393,7 @@ def test_deepseek_v4_maps_xhigh_to_high_reasoning_effort():
         (2, {"thinking": True, "reasoning_effort": "low"}),
         (3, {"thinking": True, "reasoning_effort": "low"}),
         (4, {"thinking": False}),
+        (5, {"thinking": False}),
     ],
 )
 def test_deepseek_v4_matches_reference_golden_fixtures(case_id, kwargs):
@@ -447,7 +444,10 @@ def test_deepseek_v4_image_blocks_become_placeholders():
         thinking=False,
     )
 
-    assert "<｜User｜>first:<｜deepseek_image｜>second:<｜deepseek_image｜>" in prompt
+    assert (
+        "<｜User｜>first:\n\n<｜deepseek_image｜>\n\n"
+        "second:\n\n<｜deepseek_image｜>" in prompt
+    )
 
 
 def test_deepseek_v4_image_sentinel_ids_match_tokenizer():

vllm/renderers/deepseek_v4.py

@@ -40,7 +40,7 @@ class DeepseekV4Renderer(BaseRenderer[DeepseekV4Tokenizer]):
         conversation, mm_data, mm_uuids = parse_chat_messages(
             messages,
             self.model_config,
-            content_format="string",
+            content_format="openai",
             media_io_kwargs=params.media_io_kwargs,
             mm_processor_kwargs=params.mm_processor_kwargs,
         )
@@ -67,7 +67,7 @@ class DeepseekV4Renderer(BaseRenderer[DeepseekV4Tokenizer]):
         conversation, mm_data, mm_uuids = await parse_chat_messages_async(
             messages,
             self.model_config,
-            content_format="string",
+            content_format="openai",
             media_io_kwargs=params.media_io_kwargs,
             mm_processor_kwargs=params.mm_processor_kwargs,
         )

vllm/tokenizers/deepseek_v4_encoding.py

@@ -224,7 +224,7 @@ def flatten_content_blocks(content: Any) -> Any:
             raise ValueError(
                 f"Unsupported content block type: {block.get('type')!r}"
             )
-    return "".join(parts)
+    return "\n\n".join(parts)
 
 
 def render_message(
@@ -267,7 +267,7 @@ def render_message(
     )
 
     role = msg.get("role")
-    content = msg.get("content")
+    content = flatten_content_blocks(msg.get("content"))
     tools = msg.get("tools")
     response_format = msg.get("response_format")
     tool_calls = msg.get("tool_calls")
@@ -337,7 +337,7 @@ def render_message(
                     parts.append(f"[Unsupported {block_type}]")
             prompt += "\n\n".join(parts)
         else:
-            prompt += flatten_content_blocks(content) or ""
+            prompt += content or ""
 
     elif role == "latest_reminder":
         prompt += LATEST_REMINDER_SP_TOKEN + latest_reminder_msg_template.format(content=content)