Revert from_tf=True for Flan-T5 loading

services/generation.py

@@ -37,12 +37,12 @@ def load_models():
         # Load tokenizer associated with the text model
         model_cache["text_tokenizer"] = AutoTokenizer.from_pretrained(config.TEXT_MODEL_NAME)
         # Load the sequence-to-sequence language model
-        # IMPORTANT: Add from_tf=True if the primary weights are TensorFlow format (like google/flan-t5-base)
+        # Assuming PyTorch weights (.bin or .safetensors) are available for the model.
         model_cache["text_model"] = AutoModelForSeq2SeqLM.from_pretrained(
-            config.TEXT_MODEL_NAME,
-            from_tf=True # Required for google/flan-t5-base which has tf_model.h5
+            config.TEXT_MODEL_NAME
+            # REMOVED: from_tf=True - Attempt to load PyTorch weights directly.
         ).to(config.DEVICE) # Move model to the configured device (CPU or CUDA)
-        logger.info(f"Text model '{config.TEXT_MODEL_NAME}' loaded successfully (from TF weights if applicable) onto {config.DEVICE}.")
+        logger.info(f"Text model '{config.TEXT_MODEL_NAME}' loaded successfully (using PyTorch weights) onto {config.DEVICE}.")
 
         # --- 2. Image Generation Model (Base Pipeline) ---
         logger.info(f"Loading base image generation model: {config.IMAGE_MODEL_NAME}")
@@ -377,8 +377,8 @@ def generate_video_sync(
     finally:
         # --- Resource Cleanup ---
         del input_image # Delete decoded input image
-        del video_frames_pil # Delete list of PIL frames
-        del video_frames_np # Delete list of numpy frames
+        if 'video_frames_pil' in locals(): del video_frames_pil # Delete list of PIL frames if it exists
+        if 'video_frames_np' in locals(): del video_frames_np # Delete list of numpy frames if it exists
         # Clear CUDA cache if applicable
         if config.DEVICE == "cuda":
             torch.cuda.empty_cache()