Update app.py

app.py

@@ -2,23 +2,32 @@ import gradio as gr
 import numpy as np
 import soundfile as sf
 import tempfile
+import torch
 
-from transformers import pipeline
+from transformers import AutoTokenizer, AutoModelForCausalLM
 from kokoro import KPipeline
 
 
-# Load text generation model once when Space starts
-generator = pipeline(
-    "text-generation",
-    model="HuggingFaceTB/SmolLM2-360M-Instruct",
-    device=-1
+MODEL_NAME = "HuggingFaceTB/SmolLM2-360M-Instruct"
+
+# Download and load model inside the Space
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_NAME,
+    torch_dtype=torch.float32
 )
 
-# Load text-to-speech pipeline
+model.eval()
+
+# Load text-to-speech model
 tts = KPipeline(lang_code="a")
 
 
 def generate_answer_and_voice(user_question):
+    if not user_question or not user_question.strip():
+        return "Please enter a question.", None
+
     messages = [
         {
             "role": "system",
@@ -30,15 +39,35 @@ def generate_answer_and_voice(user_question):
         }
     ]
 
-    result = generator(
+    # Convert chat messages into model prompt
+    prompt = tokenizer.apply_chat_template(
         messages,
-        max_new_tokens=120,
-        do_sample=True,
-        temperature=0.7,
-        top_p=0.9
+        tokenize=False,
+        add_generation_prompt=True
     )
 
-    generated_response = result[0]["generated_text"][-1]["content"]
+    inputs = tokenizer(
+        prompt,
+        return_tensors="pt"
+    )
+
+    with torch.no_grad():
+        output_ids = model.generate(
+            **inputs,
+            max_new_tokens=120,
+            do_sample=True,
+            temperature=0.7,
+            top_p=0.9,
+            pad_token_id=tokenizer.eos_token_id
+        )
+
+    # Keep only newly generated tokens
+    new_tokens = output_ids[0][inputs["input_ids"].shape[1]:]
+
+    generated_response = tokenizer.decode(
+        new_tokens,
+        skip_special_tokens=True
+    ).strip()
 
     audio_parts = []
 
@@ -48,9 +77,11 @@ def generate_answer_and_voice(user_question):
     ):
         audio_parts.append(audio)
 
+    if not audio_parts:
+        return generated_response, None
+
     final_audio = np.concatenate(audio_parts)
 
-    # Create temporary WAV file for Gradio output
     output_file = tempfile.NamedTemporaryFile(
         delete=False,
         suffix=".wav"
@@ -72,7 +103,7 @@ demo = gr.Interface(
         gr.Audio(label="Generated Voice", type="filepath")
     ],
     title="AI Text Generation + Text-to-Speech",
-    description="Ask an NLP question. SmolLM generates an answer and Kokoro converts it into speech."
+    description="Ask a question. SmolLM generates an answer and Kokoro converts it into speech."
 )
 
 demo.launch()