[ENV] add vision embedding preprocessing refactor

fastdeploy/config.py

@@ -217,6 +217,7 @@ def __init__(
         self.revision = None
         self.prefix_layer_name = "layers"
         self.kv_cache_quant_scale_path = ""
+        self.enable_vision_embedding_refactored = False
         self.enable_entropy = False
 
         self.partial_rotary_factor: float = 1.0

fastdeploy/engine/args_utils.py

@@ -507,6 +507,10 @@ class EngineArgs:
     Configuration for eplb.
     """
 
+    enable_vision_embedding_refactored: bool = False
+    """
+    Flag to enable vision embedding refactored. Default is False (disabled).
+    """
     routing_replay_config: Optional[Dict[str, Any]] = None
     """
     Flag to rollout routing replay(r3)
@@ -883,6 +887,12 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             default=EngineArgs.logits_processors,
             help="FQCNs (Fully Qualified Class Names) of logits processors supported by the service.",
         )
+        model_group.add_argument(
+            "--enable-vision-embedding-refactored",
+            action="store_true",
+            default=EngineArgs.enable_vision_embedding_refactored,
+            help="Flag to enable vision embedding refactored.",
+        )
         model_group.add_argument(
             "--enable-entropy",
             action="store_true",

fastdeploy/engine/common_engine.py

@@ -1718,6 +1718,7 @@ def _start_worker_service(self):
             "enable_logprob": self.cfg.model_config.enable_logprob,
             "lm_head_fp32": self.cfg.model_config.lm_head_fp32,
             "enable_entropy": self.cfg.model_config.enable_entropy,
+            "enable_vision_embedding_refactored": self.cfg.model_config.enable_vision_embedding_refactored,
         }
         for worker_flag, value in worker_store_true_flag.items():
             if value:

fastdeploy/engine/engine.py

@@ -597,6 +597,7 @@ def _start_worker_service(self):
             "lm_head_fp32": self.cfg.model_config.lm_head_fp32,
             "shutdown_comm_group_if_worker_idle": self.cfg.parallel_config.shutdown_comm_group_if_worker_idle,
             "enable_entropy": self.cfg.model_config.enable_entropy,
+            "enable_vision_embedding_refactored": self.cfg.model_config.enable_vision_embedding_refactored,
         }
         for worker_flag, value in worker_store_true_flag.items():
             if value:

fastdeploy/worker/worker_process.py

@@ -880,7 +880,11 @@ def parse_args():
         default=None,
         help="EPLB Configuration.",
     )
-
+    parser.add_argument(
+        "--enable_vision_embedding_refactored",
+        action="store_true",
+        help="Enable vision embedding refactored.",
+    )
     parser.add_argument(
         "--routing_replay_config",
         type=json.loads,