Update app.py

app.py

@@ -339,17 +339,31 @@ def apply_loras_to_pipeline(pose_strength: float, general_strength: float, motio
         orig_tmp_target = getattr(tmp_ledger, "_target_device", None)
         orig_tmp_device = getattr(tmp_ledger, "device", None)
         try:
-            tmp_ledger._target_device = torch.device("cpu")
+            # _target_device is expected to be callable by model_ledger.transformer()
+            # set it to a callable that returns CPU so builder.build(device=...) works.
+            tmp_ledger._target_device = (lambda: torch.device("cpu"))
+            # ledger.device is used after build: set it to CPU so .to(self.device) keeps the model on CPU.
             tmp_ledger.device = torch.device("cpu")
             print("[LoRA] Building fused transformer on CPU (no GPU allocation)...")
-            new_transformer_cpu = tmp_ledger.transformer()  # returns model on CPU now
+            new_transformer_cpu = tmp_ledger.transformer()  # should now return a CPU model
             print("[LoRA] Fused transformer built on CPU.")
         finally:
-            # Restore attributes (defensive)
+            # Restore attributes to their previous values (if there were any).
             if orig_tmp_target is not None:
                 tmp_ledger._target_device = orig_tmp_target
+            else:
+                # remove attribute if ledger did not have it previously
+                try:
+                    delattr(tmp_ledger, "_target_device")
+                except Exception:
+                    pass
             if orig_tmp_device is not None:
                 tmp_ledger.device = orig_tmp_device
+            else:
+                try:
+                    delattr(tmp_ledger, "device")
+                except Exception:
+                    pass
 
         # Get the existing transformer instance (the one currently used by the pipeline).
         global _transformer
@@ -366,6 +380,14 @@ def apply_loras_to_pipeline(pose_strength: float, general_strength: float, motio
 
         # State dict of CPU model (fused with LoRAs)
         new_state = new_transformer_cpu.state_dict()
+        # diagnostics: how many keys will be copied
+        total_keys = len(new_state)
+        matched = sum(1 for k in new_state if k in existing_params or k in existing_buffers)
+        print(f"[LoRA] Transformer state keys: total={total_keys} matched_for_copy={matched}")
+        if matched == 0:
+            # helpful hint if naming differs
+            sample_keys = list(new_state.keys())[:10]
+            print(f"[LoRA] Warning: 0 matching keys found. sample new_state keys: {sample_keys}")
 
         # Copy CPU tensors into the GPU-resident transformer's params/buffers in-place
         with torch.no_grad():