Support --enable-llama4-multimodal (#5254)

python/sglang/srt/server_args.py

@@ -156,6 +156,7 @@ class ServerArgs:
     disable_outlines_disk_cache: bool = False
     disable_custom_all_reduce: bool = False
     disable_mla: bool = False
+    enable_llama4_multimodal: Optional[bool] = None
     disable_overlap_schedule: bool = False
     enable_mixed_chunk: bool = False
     enable_dp_attention: bool = False
@@ -294,6 +295,8 @@ class ServerArgs:
                 f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
             )
 
+        self.enable_multimodal: Optional[bool] = self.enable_llama4_multimodal
+
         # Data parallelism attention
         if self.enable_dp_attention:
             self.schedule_conservativeness = self.schedule_conservativeness * 0.3
@@ -974,6 +977,11 @@ class ServerArgs:
             action="store_true",
             help="Disable Multi-head Latent Attention (MLA) for DeepSeek V2/V3/R1 series models.",
         )
+        parser.add_argument(
+            "--enable-llama4-multimodal",
+            action="store_true",
+            help="Enable the multimodal functionality for Llama-4.",
+        )
         parser.add_argument(
             "--disable-overlap-schedule",
             action="store_true",