check

app.py

@@ -32,20 +32,13 @@ else:
 print(f"[info] BASE_MODEL: {BASE_MODEL}; DEVICE: {DEVICE} ({DTYPE})...")
 
 
-_PROCESSOR = None
-
-
-def get_processor():
-    """Lazy-load and cache processor for future multimodal inputs."""
-    global _PROCESSOR
-    if _PROCESSOR is None:
-        _PROCESSOR = AutoProcessor.from_pretrained(BASE_MODEL)
-    return _PROCESSOR
+global _PROCESSOR
+_PROCESSOR = AutoProcessor.from_pretrained(BASE_MODEL)
+print(f"[info] Processor loaded: {type(_PROCESSOR)}")
 
 
 def setup_sae_model():
     # 1) Load base model from HF
-    _ = get_processor()
     model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
         BASE_MODEL,
         torch_dtype=DTYPE,
@@ -101,20 +94,10 @@ def setup_sae_model():
     print(f"[info] loaded SAE adapter repo: {sae_repo_id}")
     print(f"[info] target_modules         : {target_modules}")
     
-    return sae_model
-
-
-_SAE_MODEL = None
-
-
-def get_sae_model():
-    """Lazy-load and cache SAE model to avoid heavy import-time initialization."""
-    global _SAE_MODEL
-    if _SAE_MODEL is None:
-        _SAE_MODEL = setup_sae_model()
-    return _SAE_MODEL
+    return peft_model, sae_model
 
-_ = get_sae_model()  # Trigger initial load and cache the model for future use.
+global _PEFT_MODEL, _SAE_MODEL
+_PEFT_MODEL, _SAE_MODEL = setup_sae_model()
 
 
 def build_demo():