[loader] enable private loader

python/sglang/srt/configs/load_config.py

@@ -28,6 +28,7 @@ class LoadFormat(str, enum.Enum):
     REMOTE_INSTANCE = "remote_instance"
     RDMA = "rdma"
     LOCAL_CACHED = "local_cached"
+    PRIVATE = "private"
 
 
 @dataclass

python/sglang/srt/model_executor/model_runner.py

@@ -955,7 +955,7 @@ def get_weight_iter(config):
             return iter
 
         def model_load_weights(model, iter):
-            DefaultModelLoader.load_weights_and_postprocess(model, iter, target_device)
+            loader.load_weights_and_postprocess(model, iter, target_device)
             return model
 
         with set_default_torch_dtype(self.model_config.dtype):

python/sglang/srt/model_loader/loader.py

@@ -2100,4 +2100,12 @@ def get_model_loader(
     if load_config.load_format == LoadFormat.REMOTE_INSTANCE:
         return RemoteInstanceModelLoader(load_config)
 
+    if load_config.load_format == LoadFormat.PRIVATE:
+        import importlib
+        try:
+            module = importlib.import_module("sglang.private.private_model_loader")
+            return module.PrivateModelLoader(load_config)
+        except ImportError:
+            raise ValueError("Failed to import sglang.private.private_model_loader")
+
     return DefaultModelLoader(load_config)

python/sglang/srt/models/qwen2_5_vl.py

@@ -688,7 +688,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                     if name in params_dict.keys():
                         param = params_dict[name]
                     else:
-                        continue
+                        raise ValueError(f"Weight {name} not found in params_dict")
                 except KeyError:
                     print(params_dict.keys())
                     raise

python/sglang/srt/server_args.py

@@ -80,6 +80,7 @@
     "layered",
     "remote",
     "remote_instance",
+    "private",
 ]
 
 QUANTIZATION_CHOICES = [