[Diffusion] Fix lora default lora_scale bug (#17982)

python/sglang/multimodal_gen/runtime/pipelines_core/lora_pipeline.py

@@ -98,7 +98,7 @@ class LoRAPipeline(ComposedPipelineBase):
         if self.lora_path is not None:
             self.convert_to_lora_layers()
             self.set_lora(
-                self.lora_nickname, self.lora_path  # type: ignore
+                self.lora_nickname, self.lora_path, strength=self.server_args.lora_scale  # type: ignore
             )  # type: ignore
 
     def is_target_layer(self, module_name: str) -> bool:

python/sglang/multimodal_gen/runtime/server_args.py

@@ -276,6 +276,7 @@ class ServerArgs:
     # (Wenxuan) prefer to keep it here instead of in pipeline config to not make it complicated.
     lora_path: str | None = None
     lora_nickname: str = "default"  # for swapping adapters in the pipeline
+    lora_scale: float = 1.0  # LoRA scale for merging (e.g., 0.125 for Hyper-SD)
 
     # VAE parameters
     vae_path: str | None = None  # Custom VAE path (e.g., for distilled autoencoder)
@@ -728,6 +729,12 @@ class ServerArgs:
             default=ServerArgs.lora_nickname,
             help="The nickname for the LoRA adapter to launch with",
         )
+        parser.add_argument(
+            "--lora-scale",
+            type=float,
+            default=ServerArgs.lora_scale,
+            help="LoRA scale for merging (e.g., 0.125 for Hyper-SD). Same as lora_scale in Diffusers",
+        )
         # Add pipeline configuration arguments
         PipelineConfig.add_cli_args(parser)