Update app.py

app.py

@@ -5,16 +5,14 @@ import gradio as gr
 from threading import Thread
 from PIL import Image
 import subprocess
-import spaces
-from parler_tts import ParlerTTSForConditionalGeneration
-import soundfile as sf
+import spaces  # Add this import
 
 # Install flash-attention
 subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 # Constants
-TITLE = "<h1><center>Phi 3.5 Multimodal (Text + Vision + Speech)</center></h1>"
-DESCRIPTION = "# Phi-3.5 Multimodal Demo (Text + Vision + Speech)"
+TITLE = "<h1><center>Phi 3.5 Multimodal (Text + Vision)</center></h1>"
+DESCRIPTION = "# Phi-3.5 Multimodal Demo (Text + Vision)"
 
 # Model configurations
 TEXT_MODEL_ID = "microsoft/Phi-3.5-mini-instruct"
@@ -48,11 +46,6 @@ vision_model = AutoModelForCausalLM.from_pretrained(
 
 vision_processor = AutoProcessor.from_pretrained(VISION_MODEL_ID, trust_remote_code=True)
 
-# Load Parler-TTS model
-tts_device = "cuda:0" if torch.cuda.is_available() else "cpu"
-tts_model = ParlerTTSForConditionalGeneration.from_pretrained("parler-tts/parler-tts-large-v1").to(tts_device)
-tts_tokenizer = AutoTokenizer.from_pretrained("parler-tts/parler-tts-large-v1")
-
 # Helper functions
 @spaces.GPU
 def stream_text_chat(message, history, system_prompt, temperature=0.8, max_new_tokens=1024, top_p=1.0, top_k=20):
@@ -85,64 +78,24 @@ def stream_text_chat(message, history, system_prompt, temperature=0.8, max_new_t
     buffer = ""
     for new_text in streamer:
         buffer += new_text
-        yield history + [[message, buffer]], None  # Yield None for audio initially
-
-    # Generate speech for the final response
-    audio_path = generate_speech(buffer, "A clear and concise voice reads out the response.")
-    yield history + [[message, buffer]], audio_path
+        yield history + [[message, buffer]]
 
-@spaces.GPU
-def process_vision_query(image, text_input, generate_speech=True):
+@spaces.GPU  # Add this decorator
+def process_vision_query(image, text_input):
     prompt = f"<|user|>\n<|image_1|>\n{text_input}<|end|>\n<|assistant|>\n"
+    image = Image.fromarray(image).convert("RGB")
+    inputs = vision_processor(prompt, image, return_tensors="pt").to(device)
     
-    # Check if image is already a PIL Image
-    if isinstance(image, Image.Image):
-        pil_image = image
-    elif isinstance(image, np.ndarray):
-        pil_image = Image.fromarray(image).convert("RGB")
-    else:
-        raise ValueError("Unsupported image type. Expected PIL Image or numpy array.")
-    
-    inputs = vision_processor(prompt, pil_image, return_tensors="pt").to(device)
-    
-    try:
-        with torch.no_grad():
-            generate_ids = vision_model.generate(
-                **inputs, 
-                max_new_tokens=1000, 
-                eos_token_id=vision_processor.tokenizer.eos_token_id
-            )
-        
-        generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
-        response = vision_processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-        
-        if generate_speech:
-            audio_path = generate_speech_from_text(response)
-            return response, audio_path
-        else:
-            return response, None
-    except RuntimeError as e:
-        if "CUDA out of memory" in str(e):
-            error_message = "Error: GPU out of memory. Try processing a smaller image or freeing up GPU resources."
-            return error_message, None
-        else:
-            raise e
-
-
-def generate_speech_from_text(text, description="A clear voice reads out the response."):
-    input_ids = tts_tokenizer(description, return_tensors="pt").input_ids.to(tts_device)
-    prompt_input_ids = tts_tokenizer(text, return_tensors="pt").input_ids.to(tts_device)
-
     with torch.no_grad():
-        generation = tts_model.generate(input_ids=input_ids, prompt_input_ids=prompt_input_ids)
-    
-    audio_arr = generation.cpu().numpy().squeeze()
-    
-    output_path = f"output_audio_{hash(text)}.wav"
-    sf.write(output_path, audio_arr, tts_model.config.sampling_rate)
+        generate_ids = vision_model.generate(
+            **inputs, 
+            max_new_tokens=1000, 
+            eos_token_id=vision_processor.tokenizer.eos_token_id
+        )
     
-    return output_path
-
+    generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
+    response = vision_processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    return response
 
 # Custom CSS
 custom_css = """
@@ -165,7 +118,7 @@ footer { text-align: center; margin-top: 2rem; color: #64748b;}
 custom_header = """
 <div id="custom-header">
     <h1><span class="blue">Phi 3.5</span> <span class="pink">Multimodal Assistant</span></h1>
-    <h2>Text, Vision, and Speech AI at Your Service</h2>
+    <h2>Text and Vision AI at Your Service</h2>
 </div>
 """
 
@@ -181,8 +134,8 @@ custom_suggestions = """
         <p>Analyze Images with Vision Model</p>
     </div>
     <div class="suggestion">
-        <span class="suggestion-icon">🔊</span>
-        <p>Generate Speech with Parler-TTS</p>
+        <span class="suggestion-icon">🤖</span>
+        <p>Get AI-generated responses</p>
     </div>
     <div class="suggestion">
         <span class="suggestion-icon">🔍</span>
@@ -216,45 +169,22 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Base().set(
         
         submit_btn = gr.Button("Submit", variant="primary")
         clear_btn = gr.Button("Clear Chat", variant="secondary")
-        audio_output = gr.Audio(label="AI Response Audio")
 
-        submit_btn.click(stream_text_chat, 
-                         inputs=[msg, chatbot, system_prompt, temperature, max_new_tokens, top_p, top_k], 
-                         outputs=[chatbot, audio_output])
-        clear_btn.click(lambda: (None, None), None, [chatbot, audio_output], queue=False)
-        
-    with gr.Tab("Vision Model with TTS (Phi-3.5-vision)"):
+        submit_btn.click(stream_text_chat, [msg, chatbot, system_prompt, temperature, max_new_tokens, top_p, top_k], [chatbot])
+        clear_btn.click(lambda: None, None, chatbot, queue=False)
+
+    with gr.Tab("Vision Model (Phi-3.5-vision)"):
         with gr.Row():
-            
             with gr.Column(scale=1):
-                
                 vision_input_img = gr.Image(label="Upload an Image", type="pil")
                 vision_text_input = gr.Textbox(label="Ask a question about the image", placeholder="What do you see in this image?")
-                vision_submit_btn = gr.Button("Analyze Image and Generate Speech", variant="primary")
+                vision_submit_btn = gr.Button("Analyze Image", variant="primary")
             with gr.Column(scale=1):
                 vision_output_text = gr.Textbox(label="AI Analysis", lines=10)
-                vision_output_audio = gr.Audio(label="Generated Speech")
-        
-        vision_submit_btn.click(process_vision_query, 
-                            inputs=[vision_input_img, vision_text_input],
-                            outputs=[vision_output_text, vision_output_audio])
-        
-    
-    
-
-    with gr.Tab("Text-to-Speech (Parler-TTS)"):
-        with gr.Row():
-            
-            with gr.Column(scale=1):
-                tts_prompt = gr.Textbox(label="Text to Speak", placeholder="Enter the text you want to convert to speech...")
-                tts_description = gr.Textbox(label="Voice Description", value="A female speaker delivers a slightly expressive and animated speech with a moderate speed and pitch. The recording is of very high quality, with the speaker's voice sounding clear and very close up.", lines=3)
-                tts_submit_btn = gr.Button("Generate Speech", variant="primary")
-            with gr.Column(scale=1):
-                tts_output_audio = gr.Audio(label="Generated Speech")
         
-        tts_submit_btn.click(generate_speech, inputs=[tts_prompt, tts_description], outputs=[tts_output_audio])
+        vision_submit_btn.click(process_vision_query, [vision_input_img, vision_text_input], [vision_output_text])
 
-    gr.HTML("<footer>Powered by Phi 3.5 Multimodal AI and Parler-TTS</footer>")
+    gr.HTML("<footer>Powered by Phi 3.5 Multimodal AI</footer>")
 
 if __name__ == "__main__":
     demo.launch()