debugging why it don't work

Author: damian0815

ldm/modules/embedding_manager.py

@@ -140,6 +140,8 @@ def forward(
         tokenized_text,
         embedded_text,
     ):
+        # torch.save(embedded_text, '/tmp/embedding-manager-uglysonic-pre-rewrite.pt')
+
         b, n, device = *tokenized_text.shape, tokenized_text.device
 
         for (
@@ -241,7 +243,7 @@ def load(self, ckpt_paths, full=True):
         # both will be stored in this dictionary
         for term in self.string_to_param_dict.keys():
             term = term.strip('<').strip('>')
-            self.concepts_loaded[term] = True  
+            self.concepts_loaded[term] = True
         print(f'>> Current embedding manager terms: {", ".join(self.string_to_param_dict.keys())}')
 
     def _expand_directories(self, paths:list[str]):
@@ -262,7 +264,7 @@ def _load(self, ckpt_path, full=True):
             print(f'\n### Security Issues Found in Model: {scan_result.issues_count}')
             print('### For your safety, InvokeAI will not load this embed.')
             return
-        
+
         ckpt = torch.load(ckpt_path, map_location='cpu')
 
         # Handle .pt textual inversion files

ldm/modules/encoders/modules.py

@@ -548,7 +548,7 @@ def forward(self, text: list, **kwargs):
 
             #print(f"assembled tokens for '{fragments}' into tensor of shape {lerped_embeddings.shape}")
 
-            # append to batch 
+            # append to batch
             batch_z = lerped_embeddings.unsqueeze(0) if batch_z is None else torch.cat([batch_z, lerped_embeddings.unsqueeze(0)], dim=1)
             batch_tokens = tokens.unsqueeze(0) if batch_tokens is None else torch.cat([batch_tokens, tokens.unsqueeze(0)], dim=1)
 
@@ -647,6 +647,7 @@ def build_weighted_embedding_tensor(self, tokens: torch.Tensor, per_token_weight
         '''
         #print(f"building weighted embedding tensor for {tokens} with weights {per_token_weights}")
         z = self.transformer(input_ids=tokens.unsqueeze(0), **kwargs)
+        # torch.save(z, '/tmp/embedding-manager-uglysonic-post-rewrite.pt')
         batch_weights_expanded = per_token_weights.reshape(per_token_weights.shape + (1,)).expand(z.shape)
 
         if weight_delta_from_empty: