Update app.py

app.py

@@ -37,7 +37,6 @@ from diffusers import ShapEImg2ImgPipeline, ShapEPipeline
 from diffusers.utils import export_to_ply
 
 os.system('pip install backoff')
-
 # Global constants and helper functions
 
 MAX_SEED = np.iinfo(np.int32).max
@@ -259,16 +258,7 @@ phi4_model = AutoModelForCausalLM.from_pretrained(
 # ------------------------------------------------------------------------------
 
 DESCRIPTION = """
-# Agent Dino 🌠  
-This chatbot supports various commands:  
-- **@tts1 / @tts2:** text-to-speech  
-- **@image:** image generation  
-- **@3d:** 3D mesh generation  
-- **@web:** web search/visit  
-- **@rAgent:** reasoning chain  
-- **@yolo:** object detection  
-- **@phi4:** multimodal (image/audio) question answering  
-"""
+# Agent Dino 🌠"""
 
 css = '''
 h1 {
@@ -582,14 +572,15 @@ def generate(
         if not question:
             yield "Error: Please provide a question after @phi4."
             return
-    
         # Determine input type (Image or Audio) from the first file
         input_file = files[0]
         try:
+            # If file is already a PIL Image, treat as image
             if isinstance(input_file, Image.Image):
                 input_type = "Image"
                 file_for_phi4 = input_file
             else:
+                # Try opening as image; if it fails, assume audio
                 try:
                     file_for_phi4 = Image.open(input_file)
                     input_type = "Image"
@@ -599,7 +590,7 @@ def generate(
         except Exception:
             input_type = "Audio"
             file_for_phi4 = input_file
-    
+
         if input_type == "Image":
             phi4_prompt = f'{phi4_user_prompt}<|image_1|>{question}{phi4_prompt_suffix}{phi4_assistant_prompt}'
             inputs = phi4_processor(text=phi4_prompt, images=file_for_phi4, return_tensors='pt').to(phi4_model.device)
@@ -610,22 +601,20 @@ def generate(
         else:
             yield "Invalid file type for @phi4 multimodal processing."
             return
-            
+
         with torch.no_grad():
             generate_ids = phi4_model.generate(
                 **inputs,
                 max_new_tokens=200,
                 num_logits_to_keep=0,
-                streamer=streamer  # Adding text streamer
             )
-    
-        buffer = "⚛️ phi4 multimodal is initiated, hold tight"
-        for new_text in streamer:
-            buffer += new_text
-            buffer = buffer.replace("<|im_end|>", "")
-            time.sleep(0.01)
-            yield buffer
-
+        input_length = inputs['input_ids'].shape[1]
+        generate_ids = generate_ids[:, input_length:]
+        response = phi4_processor.batch_decode(
+            generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )[0]
+        yield response
+        return
 
     # --- Text and TTS branch ---
     tts_prefix = "@tts"