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

python/sglang/bench_one_batch.py

@@ -136,6 +136,7 @@ def load_model(server_args, port_args, tp_rank):
         context_length=server_args.context_length,
         model_override_args=server_args.json_model_override_args,
         is_embedding=server_args.is_embedding,
+        enable_multimodal=server_args.enable_multimodal,
         dtype=server_args.dtype,
         quantization=server_args.quantization,
     )

python/sglang/srt/configs/model_config.py

@@ -43,10 +43,12 @@ class ModelConfig:
         context_length: Optional[int] = None,
         model_override_args: Optional[str] = None,
         is_embedding: Optional[bool] = None,
+        enable_multimodal: Optional[bool] = None,
         dtype: str = "auto",
         quantization: Optional[str] = None,
         override_config_file: Optional[str] = None,
     ) -> None:
+
         self.model_path = model_path
         self.revision = revision
         self.quantization = quantization
@@ -70,14 +72,28 @@ class ModelConfig:
             self.hf_text_config, "attention_chunk_size", None
         )
 
+        if enable_multimodal is None:
+            if self.hf_config.architectures == "Llama4ForConditionalGeneration":
+                enable_multimodal = False
+            else:
+                enable_multimodal = True
+
         # Check model type
         self.is_generation = is_generation_model(
             self.hf_config.architectures, is_embedding
         )
-        self.is_multimodal = is_multimodal_model(self.hf_config.architectures)
-        self.is_multimodal_gen = is_multimodal_gen_model(self.hf_config.architectures)
-        self.is_image_gen = is_image_gen_model(self.hf_config.architectures)
-        self.is_audio_model = is_audio_model(self.hf_config.architectures)
+        self.is_multimodal = enable_multimodal and is_multimodal_model(
+            self.hf_config.architectures
+        )
+        self.is_multimodal_gen = enable_multimodal and is_multimodal_gen_model(
+            self.hf_config.architectures
+        )
+        self.is_image_gen = enable_multimodal and is_image_gen_model(
+            self.hf_config.architectures
+        )
+        self.is_audio_model = enable_multimodal and is_audio_model(
+            self.hf_config.architectures
+        )
         self.is_encoder_decoder = is_encoder_decoder_model(self.hf_config.architectures)
         self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
 

python/sglang/srt/managers/scheduler.py

@@ -437,6 +437,7 @@ class Scheduler(
             context_length=server_args.context_length,
             model_override_args=server_args.json_model_override_args,
             is_embedding=server_args.is_embedding,
+            enable_multimodal=server_args.enable_multimodal,
             dtype=server_args.dtype,
             quantization=server_args.quantization,
         )

python/sglang/srt/managers/tokenizer_manager.py

@@ -163,6 +163,7 @@ class TokenizerManager:
             context_length=server_args.context_length,
             model_override_args=server_args.json_model_override_args,
             is_embedding=server_args.is_embedding,
+            enable_multimodal=server_args.enable_multimodal,
             dtype=server_args.dtype,
             quantization=server_args.quantization,
         )

python/sglang/srt/managers/tp_worker.py

@@ -68,6 +68,7 @@ class TpModelWorker:
             context_length=server_args.context_length,
             model_override_args=server_args.json_model_override_args,
             is_embedding=server_args.is_embedding,
+            enable_multimodal=server_args.enable_multimodal,
             dtype=server_args.dtype,
             quantization=server_args.quantization,
         )

python/sglang/srt/model_executor/model_runner.py

@@ -281,7 +281,6 @@ class ModelRunner:
                 f"Automatically reduce --mem-fraction-static to {self.mem_fraction_static:.3f} "
                 f"because this is a multimodal model."
             )
-
             logger.info(
                 "Automatically turn off --chunked-prefill-size for multimodal model."
             )

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",