Update app.py

app.py

@@ -473,7 +473,7 @@ def prepare_lora_cache(
             spatial_upsampler_path=str(spatial_upsampler_path),
             gemma_root_path=str(gemma_root),
             loras=tuple(loras_for_builder),
-            quantization=getattr(ledger, "quantization", None),
+            quantization=None,
         )
         new_transformer_cpu = tmp_ledger.transformer()
 
@@ -525,7 +525,7 @@ def apply_prepared_lora_state_to_pipeline():
         print("[LoRA] Prepared LoRA state already active; skipping.")
         return True
 
-    existing_transformer = _transformer
+    existing_transformer = get_transformer()
     with torch.no_grad():
         missing, unexpected = existing_transformer.load_state_dict(PENDING_LORA_STATE, strict=False)
         if missing or unexpected:
@@ -552,8 +552,7 @@ _orig_spatial_upsampler_factory = ledger.spatial_upsampler
 _orig_text_encoder_factory = ledger.text_encoder
 _orig_gemma_embeddings_factory = ledger.gemma_embeddings_processor
 
-# Call the original factories once to create the cached instances we will serve by default.
-_transformer = _orig_transformer_factory()
+# Keep everything else cached as before.
 _video_encoder = _orig_video_encoder_factory()
 _video_decoder = _orig_video_decoder_factory()
 _audio_encoder = _orig_audio_encoder_factory()
@@ -563,9 +562,18 @@ _spatial_upsampler = _orig_spatial_upsampler_factory()
 _text_encoder = _orig_text_encoder_factory()
 _embeddings_processor = _orig_gemma_embeddings_factory()
 
-# Replace ledger methods with lightweight lambdas that return the cached instances.
-# We keep the original factories above so we can call them later to rebuild components.
-ledger.transformer = lambda: _transformer
+# Do NOT build the transformer here.
+# Build it lazily only when generation or LoRA application actually needs it.
+_transformer = None
+
+def get_transformer():
+    global _transformer
+    if _transformer is None:
+        _transformer = _orig_transformer_factory()
+    return _transformer
+
+# Replace ledger methods with lightweight getters.
+ledger.transformer = get_transformer
 ledger.video_encoder = lambda: _video_encoder
 ledger.video_decoder = lambda: _video_decoder
 ledger.audio_encoder = lambda: _audio_encoder
@@ -575,7 +583,7 @@ ledger.spatial_upsampler = lambda: _spatial_upsampler
 ledger.text_encoder = lambda: _text_encoder
 ledger.gemma_embeddings_processor = lambda: _embeddings_processor
 
-print("All models preloaded (including Gemma text encoder and audio encoder)!")
+print("All non-transformer models preloaded; transformer will be built lazily.")
 # ---- REPLACE PRELOAD BLOCK END ----
 
 print("=" * 80)