Update handler.py

handler.py

@@ -31,13 +31,13 @@ class EndpointHandler:
 
         # pass inputs with all kwargs in data
         with torch.autocast("cuda"):
-                audio_values = self.model.generate(**inputs, do_sample=False, max_new_tokens=400)
+                audio_values = self.model.generate(**inputs, do_sample=True, guidance_scale=3, max_new_tokens=400)
 
         # postprocess the prediction  
         sampling_rate = self.model.config.audio_encoder.sampling_rate
         audio_samples = audio_values[0].cpu().numpy()[0].tolist()
 
-        audio_samples = [int(min(max(sample * 32767, -32768), 32767)) for sample in audio_samples]
+        audio_samples = [int(min(max(sample * 32000, -32000), 32000)) for sample in audio_samples]
     
         # Create BytesIO object to capture the audio in-memory
         audio_io = io.BytesIO()