Update app.py

app.py

@@ -35,7 +35,7 @@ def farms_spectral_analysis(tensor, num_patches=10, patch_size=64):
                     try:
                                     # We only need top components to find the "Universal Subspace"
                                     u, s, vh = np.linalg.svd(patch.float().numpy(), full_matrices=False)
-                                    u_list.append(u[:, :1]) # Keep top principal direction
+                                    u_list.append(u[:, :1])  # Keep top principal direction
                                 except:
             continue
 
@@ -46,8 +46,8 @@ def farms_spectral_analysis(tensor, num_patches=10, patch_size=64):
         # Fallback to full SVD for the merging step, but using the "Universal" concept
         # We posit the top k singular vectors form the shared subspace.
         try:
-                    u, s, v = torch.svd_lowrank(tensor.float(), q=32) # Efficient randomized SVD
-                    return u, v # Returns Left (U) and Right (V) singular vectors
+                    u, s, v = torch.svd_lowrank(tensor.float(), q=32)  # Efficient randomized SVD
+                    return u, v  # Returns Left (U) and Right (V) singular vectors
                 except:
         return None, None
 
@@ -72,9 +72,9 @@ def farms_spectral_analysis(tensor, num_patches=10, patch_size=64):
                                     if not layer_tensors:
                                         continue
 
-                                    # Stack for analysis
-                                    # Shape: (N_adapters, rows, cols)
-                                    stack = torch.stack(layer_tensors)
+                        # Stack for analysis
+                        # Shape: (N_adapters, rows, cols)
+                        stack = torch.stack(layer_tensors)
                         avg_weight = torch.mean(stack, dim=0)
 
                         # 2. IF it's a LoRA weight (usually 'lora_A' or 'lora_B'), we try SAMM