[Model] Composite weight loading for multimodal Qwen2 (#10944)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2024-12-07 22:22:52 +08:00
parent b26b4cd03c
commit bf0e382e16
7 changed files with 148 additions and 206 deletions
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -2472,7 +2472,15 @@ class VllmConfig:
            return quant_config
        return None

-    def with_hf_config(self, hf_config: PretrainedConfig) -> "VllmConfig":
+    def with_hf_config(
+        self,
+        hf_config: PretrainedConfig,
+        architectures: Optional[list[str]] = None,
+    ) -> "VllmConfig":
+        if architectures is not None:
+            hf_config = copy.deepcopy(hf_config)
+            hf_config.architectures = architectures
+
        model_config = copy.deepcopy(self.model_config)
        model_config.hf_config = hf_config