folder

configs/clm_models/agent_7b_sft.yaml

@@ -15,4 +15,5 @@ output_resampler:
 
 lm_loss_scale: 1.0
 rec_loss_scale: 1.0
-pretrained_model_path: TencentARC/SEED-Story
+pretrained_model_path: TencentARC/SEED-Story
+subfolder: seed_story/george_sft

src/models_clm/models.py

@@ -221,23 +221,22 @@ class ContinuousLVLM(nn.Module):
             'past_key_values': output_past_key_values
         }
 
-
     @classmethod
-    def from_pretrained(cls, llm, input_resampler, output_resampler, pretrained_model_path=None, **kwargs):
+    def from_pretrained(cls, llm, input_resampler, output_resampler, pretrained_model_path=None, subfolder=None, **kwargs):
         model = cls(llm=llm, input_resampler=input_resampler, output_resampler=output_resampler, **kwargs)
 
         if pretrained_model_path is not None:
-            # Check if the path is intended for Hugging Face Hub
+            # Load model from Hugging Face Hub with subfolder specification
             if 'TencentARC/SEED-Story' in pretrained_model_path:
-                # Load from a specific subfolder within the Hugging Face repository
-                ckpt = AutoModel.from_pretrained(pretrained_model_path, subfolder="seed_story/george_sft")
+                # Use `subfolder` to specify the location within the repository
+                ckpt = AutoModel.from_pretrained(pretrained_model_path, subfolder=subfolder)
                 missing, unexpected = model.load_state_dict(ckpt.state_dict(), strict=False)
-                print('Agent model, missing keys: ', len(missing), 'unexpected keys:', len(unexpected))
+                print('Detokenizer model, missing keys: ', len(missing), 'unexpected keys:', len(unexpected))
             else:
                 # For local path loading
                 ckpt = torch.load(pretrained_model_path, map_location='cpu')
                 missing, unexpected = model.load_state_dict(ckpt, strict=False)
-                print('Agent model, missing keys: ', len(missing), 'unexpected keys:', len(unexpected))
+                print('Detokenizer model, missing keys: ', len(missing), 'unexpected keys:', len(unexpected))
 
         return model