Fix nested adapter path

app.py

@@ -25,6 +25,7 @@ HF_TOKEN = os.getenv("HF_TOKEN")
 BASE_MODEL_ID = "llava-hf/llava-1.5-7b-hf"
 ADAPTER_PATH = ROOT_DIR / "final-production-weights" / "best_model"
 ADAPTER_REPO_ID = "Werrewulf/tmos-dd-weights"
+ADAPTER_SUBFOLDER = os.getenv("ADAPTER_SUBFOLDER", "multimodal/checkpoint-5")
 PROMPT = "USER: <image>\nIs this image real or produced by AI?\nASSISTANT:"
 TARGET_IMAGE_SIZE = 336
 THRESHOLD = 0.5
@@ -65,7 +66,7 @@ def load_model_and_processor():
         adapter_source = ADAPTER_REPO_ID
 
     dtype = select_torch_dtype()
-    print(f"Loading TMOS-DD model from {adapter_source} with dtype={dtype}...")
+    print(f"Loading TMOS-DD model from {adapter_source} (subfolder={ADAPTER_SUBFOLDER}) with dtype={dtype}...")
 
     base_model = TMOSClassifier(
         base_model_id=BASE_MODEL_ID,
@@ -73,7 +74,11 @@ def load_model_and_processor():
         device_map="auto",
         token=HF_TOKEN,
     )
-    model = PeftModel.from_pretrained(base_model, adapter_source, is_trainable=False, token=HF_TOKEN)
+    peft_kwargs = {"is_trainable": False, "token": HF_TOKEN}
+    if adapter_source == ADAPTER_REPO_ID:
+        peft_kwargs["subfolder"] = ADAPTER_SUBFOLDER
+
+    model = PeftModel.from_pretrained(base_model, adapter_source, **peft_kwargs)
     model.eval()
 
     processor = AutoProcessor.from_pretrained(BASE_MODEL_ID, token=HF_TOKEN)