init

app.py

@@ -18,7 +18,7 @@ model.disable_talker()
 processor = Qwen2_5OmniProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
 
 @spaces.GPU
-def run_omni(audio_path: str, instruction: str, max_tokens: int = 512) -> str:
+def run_omni(audio_path: str, instruction: str) -> str:
     if not audio_path:
         return "Please upload an audio file."
 
@@ -53,7 +53,7 @@ def run_omni(audio_path: str, instruction: str, max_tokens: int = 512) -> str:
     )
     inputs = inputs.to(model.device)
 
-    output_ids = model.generate(**inputs, max_new_tokens=int(max_tokens))
+    output_ids = model.generate(**inputs, max_new_tokens=4096)
     output_ids = output_ids[:, inputs["input_ids"].shape[1]:]
     response = processor.batch_decode(
         output_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
@@ -72,12 +72,11 @@ with gr.Blocks(title="Qwen2.5 Omni (Audio) Demo") as demo:
                 label="Instruction",
                 value="Transcribe the audio, then summarize it in one sentence.",
             )
-            max_tokens = gr.Slider(128, 2048, value=512, step=64, label="Max Output Tokens")
             submit_btn = gr.Button("Run", variant="primary")
         with gr.Column():
             output_text = gr.Textbox(label="Response", lines=14)
 
-    submit_btn.click(run_omni, [audio_input, instruction, max_tokens], output_text)
+    submit_btn.click(run_omni, [audio_input, instruction], output_text)
 
 
 if __name__ == "__main__":