refactor: bug fixes and refactor for vlm (#4661)

python/sglang/srt/model_executor/forward_batch_info.py

@@ -332,7 +332,7 @@ class ForwardBatch:
 
         return ret
 
-    def get_merged_image_inputs(self) -> Optional[ImageInputs]:
+    def merge_image_inputs(self) -> Optional[ImageInputs]:
         """
         Merge all image inputs in the batch into a single ImageInputs object.
 
@@ -358,6 +358,16 @@ class ForwardBatch:
 
         return merged
 
+    def contains_image_inputs(self) -> bool:
+        """ """
+        if self.image_inputs is None:
+            return True
+        return any(
+            image_input.pixel_values is not None and image_input.pixel_values is not []
+            for image_input in self.image_inputs
+            if image_input is not None
+        )
+
     def _compute_mrope_positions(
         self, model_runner: ModelRunner, batch: ModelWorkerBatch
     ):

python/sglang/srt/model_executor/model_runner.py

@@ -273,7 +273,7 @@ class ModelRunner:
             if self.model_config.hf_config.architectures == ["DeepseekVL2ForCausalLM"]:
                 # TODO: deepseek-vl2 does not support radix cache now, set disable_radix_cache=True automatically
                 logger.info(
-                    "Automatically turn off --chunked-prefill-size and disable radix cache for deekseek-vl2."
+                    "Automatically turn off --chunked-prefill-size and disable radix cache for deepseek-vl2."
                 )
                 server_args.chunked_prefill_size = -1
                 server_args.disable_radix_cache = True