fix: fix image benchmark backend parity (#30867)

python/sglang/benchmark/datasets/image.py

@@ -199,22 +199,23 @@ def create_mm_data_row(
     # Vision tokens = total tokens - text tokens
     vision_prompt_len = prompt_len - text_prompt_len
 
-    supported_backends = [
+    supported_backends = (
         "sglang",
         "sglang-native",
         "sglang-oai-chat",
         "vllm-chat",
-    ]
+        "lmdeploy-chat",
+    )
     if backend not in supported_backends:
         raise ValueError(
             f"Image dataset only supports backends: {supported_backends}, "
             f"got '{backend}'."
         )
 
-    # OpenAI chat handlers apply the chat template and receive images separately, so
+    # Chat-completions backends apply their own chat template, so send raw text.
-    # send the raw text. /generate does not apply a chat template, so it needs
+    # Native SGLang /generate does not apply a template and needs the image
-    # prompt_str, which contains the multimodal processor's image placeholders.
+    # placeholder-bearing prompt generated by the processor.
-    use_raw_prompt = backend in ("sglang-oai-chat", "vllm-chat")
+    use_raw_prompt = backend in ("sglang-oai-chat", "vllm-chat", "lmdeploy-chat")
 
     return DatasetRow(
         prompt=text_prompt if use_raw_prompt else prompt_str,

test/registered/bench_fn/test_benchmark_datasets_api.py

@@ -426,24 +426,26 @@ class TestBenchmarkDatasetsAPI(unittest.TestCase):
         self.assertTrue(all(isinstance(row, DatasetRow) for row in rows))
         self.assertTrue(all(row.image_data for row in rows))
 
-    def test_image_sampler_vllm_chat(self):
+    def test_image_sampler_chat_backends_use_raw_prompt(self):
-        rows = sample_image_requests(
+        for backend in ("sglang-oai-chat", "vllm-chat", "lmdeploy-chat"):
-            num_requests=2,
+            with self.subTest(backend=backend):
-            image_count=1,
+                rows = sample_image_requests(
-            input_len=8,
+                    num_requests=1,
-            output_len=4,
+                    image_count=1,
-            range_ratio=0.0,
+                    input_len=8,
-            processor=self.processor,
+                    output_len=4,
-            image_content="blank",
+                    range_ratio=0.0,
-            image_format="png",
+                    processor=self.processor,
-            image_resolution="8x8",
+                    image_content="blank",
-            backend="vllm-chat",
+                    image_format="png",
-            random_image_count=False,
+                    image_resolution="8x8",
-        )
+                    backend=backend,
-        self.assertEqual(len(rows), 2)
+                    random_image_count=False,
-        self.assertTrue(all(isinstance(row, DatasetRow) for row in rows))
+                )
-        self.assertTrue(all(row.image_data for row in rows))
+                self.assertEqual(len(rows), 1)
-        self.assertTrue(all("[IMAGE]" not in row.prompt for row in rows))
+                self.assertTrue(rows[0].image_data)
+                for marker in ("user:", "assistant:", "[IMAGE]"):
+                    self.assertNotIn(marker, rows[0].prompt)
 
     def test_image_sampler_random_resolution(self):
         state = np.random.get_state()