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
-    # send the raw text. /generate does not apply a chat template, so it needs
-    # prompt_str, which contains the multimodal processor's image placeholders.
-    use_raw_prompt = backend in ("sglang-oai-chat", "vllm-chat")
+    # Chat-completions backends apply their own chat template, so send raw text.
+    # Native SGLang /generate does not apply a template and needs the image
+    # placeholder-bearing prompt generated by the processor.
+    use_raw_prompt = backend in ("sglang-oai-chat", "vllm-chat", "lmdeploy-chat")
 
     return DatasetRow(
         prompt=text_prompt if use_raw_prompt else prompt_str,