assure tokenizer EOS tokens are properly moved

app.py

@@ -420,6 +420,23 @@ def make_prediction(symbol: str, timeframe: str = "1d", prediction_days: int = 5
                         for name, value in pipe.tokenizer.__dict__.items():
                             if isinstance(value, torch.Tensor):
                                 pipe.tokenizer.__dict__[name] = value.to(device)
+                        
+                        # Ensure all tokenizer methods are on GPU
+                        if hasattr(pipe.tokenizer, '_append_eos_token'):
+                            # Create a wrapper for _append_eos_token to ensure device consistency
+                            original_append_eos = pipe.tokenizer._append_eos_token
+                            def wrapped_append_eos(token_ids, attention_mask):
+                                # Ensure both tensors are on GPU
+                                token_ids = token_ids.to(device)
+                                attention_mask = attention_mask.to(device)
+                                # Get the EOS token and ensure it's on GPU
+                                eos_token = torch.tensor([pipe.tokenizer.eos_token_id], device=device)
+                                eos_tokens = eos_token.unsqueeze(0).expand(token_ids.shape[0], 1)
+                                # Concatenate on GPU
+                                token_ids = torch.cat((token_ids, eos_tokens), dim=1)
+                                attention_mask = torch.cat((attention_mask, torch.ones_like(eos_tokens)), dim=1)
+                                return token_ids, attention_mask
+                            pipe.tokenizer._append_eos_token = wrapped_append_eos
                     
                     # Force synchronization again to ensure all tensors are on GPU
                     torch.cuda.synchronize()