From b05b346a13ce2dcee698517e2661450bd1c1acbe Mon Sep 17 00:00:00 2001
From: Yinghai Lu <yinghai@thinkingmachines.ai>
Date: Thu, 11 Dec 2025 11:49:46 -0800
Subject: [PATCH] [loader] enable private loader (#14620)

---
 python/sglang/srt/configs/load_config.py         | 1 +
 python/sglang/srt/model_executor/model_runner.py | 2 +-
 python/sglang/srt/model_loader/loader.py         | 9 +++++++++
 python/sglang/srt/models/qwen2_5_vl.py           | 2 +-
 python/sglang/srt/server_args.py                 | 1 +
 5 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/configs/load_config.py b/python/sglang/srt/configs/load_config.py
index e87bb21e4..4619dc286 100644
--- a/python/sglang/srt/configs/load_config.py
+++ b/python/sglang/srt/configs/load_config.py
@@ -29,6 +29,7 @@ class LoadFormat(str, enum.Enum):
     REMOTE_INSTANCE = "remote_instance"
     RDMA = "rdma"
     LOCAL_CACHED = "local_cached"
+    PRIVATE = "private"
 
 
 @dataclass
diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py
index 1bd2ae43b..26569c5fd 100644
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -953,7 +953,7 @@ class ModelRunner:
             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):
diff --git a/python/sglang/srt/model_loader/loader.py b/python/sglang/srt/model_loader/loader.py
index 119967978..4b8e5b084 100644
--- a/python/sglang/srt/model_loader/loader.py
+++ b/python/sglang/srt/model_loader/loader.py
@@ -2585,4 +2585,13 @@ 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)
diff --git a/python/sglang/srt/models/qwen2_5_vl.py b/python/sglang/srt/models/qwen2_5_vl.py
index 284bdfcfd..b84b17b3e 100644
--- a/python/sglang/srt/models/qwen2_5_vl.py
+++ b/python/sglang/srt/models/qwen2_5_vl.py
@@ -696,7 +696,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                     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
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 6fa195332..a0a943f6d 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -82,6 +82,7 @@ LOAD_FORMAT_CHOICES = [
     "flash_rl",
     "remote",
     "remote_instance",
+    "private",
 ]
 
 QUANTIZATION_CHOICES = [