init

app.py

@@ -23,7 +23,7 @@ model = AutoModelForCausalLM.from_pretrained(
 generation_config = GenerationConfig.from_pretrained(MODEL_ID)
 
 @spaces.GPU
-def run_phi4(audio_path: str, instruction: str, max_tokens: int = 512) -> str:
+def run_phi4(audio_path: str, instruction: str) -> str:
     if not audio_path:
         return "Please upload an audio file."
 
@@ -39,7 +39,7 @@ def run_phi4(audio_path: str, instruction: str, max_tokens: int = 512) -> str:
 
     output_ids = model.generate(
         **inputs,
-        max_new_tokens=int(max_tokens),
+        max_new_tokens=4096,
         generation_config=generation_config,
     )
     output_ids = output_ids[:, inputs["input_ids"].shape[1]:]
@@ -61,12 +61,11 @@ with gr.Blocks(title="Phi-4 Multimodal Audio Demo") as demo:
                     "Use <sep> as a separator between the original transcript and the translation."
                 ),
             )
-            max_tokens = gr.Slider(128, 2000, value=1000, step=64, label="Max Output Tokens")
             submit_btn = gr.Button("Run", variant="primary")
         with gr.Column():
             output_text = gr.Textbox(label="Model Response", lines=14)
 
-    submit_btn.click(run_phi4, [audio_input, instruction, max_tokens], output_text)
+    submit_btn.click(run_phi4, [audio_input, instruction], output_text)
 
 
 if __name__ == "__main__":