Update vllm to 0.6.3 (#1711) (#1720)

Co-authored-by: Ke Bao <ISPObaoke@163.com>

python/sglang/srt/model_executor/model_runner.py

@@ -59,8 +59,11 @@ from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import (
     enable_show_time_cost,
     get_available_gpu_memory,
+    is_attention_free_model,
+    is_embedding_model,
     is_generation_model,
     is_multimodal_model,
+    model_has_inner_state,
     monkey_patch_vllm_dummy_weight_loader,
     monkey_patch_vllm_p2p_access_check,
 )
@@ -316,11 +319,13 @@ class ModelRunner:
 
         def get_weight_iter(config):
             iter = loader._get_weights_iterator(
-                config.model,
-                config.revision,
-                fall_back_to_pt=getattr(
-                    self.model, "fall_back_to_pt_during_load", True
-                ),
+                DefaultModelLoader.Source(
+                    config.model,
+                    revision=config.revision,
+                    fall_back_to_pt=getattr(
+                        self.model, "fall_back_to_pt_during_load", True
+                    ),
+                )
             )
             return iter
 
@@ -662,3 +667,7 @@ def load_model_cls_srt(model_arch: str) -> Optional[Type[nn.Module]]:
 
 # Monkey patch model loader
 setattr(ModelRegistry, "_try_load_model_cls", load_model_cls_srt)
+setattr(ModelRegistry, "is_multimodal_model", is_multimodal_model)
+setattr(ModelRegistry, "is_attention_free_model", is_attention_free_model)
+setattr(ModelRegistry, "model_has_inner_state", model_has_inner_state)
+setattr(ModelRegistry, "is_embedding_model", is_embedding_model)