Fit inference limits to ZeroGPU xlarge quota

app.py

@@ -212,8 +212,9 @@ TASK_PROMPTS = {
     "Answer a question about the audio": "What is being said in this audio, and what is the tone?",
 }
 GREEDY_TASKS = {"Transcribe (ASR)", "Translate speech to English"}
-MAX_AUDIO_DURATION_SECONDS = 300.0
-MAX_GPU_DURATION_SECONDS = 240
+MAX_AUDIO_DURATION_SECONDS = 120.0
+MAX_NEW_TOKENS = 512
+MAX_GPU_DURATION_SECONDS = 60
 
 
 def _probe_audio_duration(audio: str | None) -> float:
@@ -238,12 +239,16 @@ def _estimate(
     **kwargs: object,
 ) -> int:
     audio_duration = _probe_audio_duration(audio)
-    if audio is None or audio_duration > MAX_AUDIO_DURATION_SECONDS:
+    if (
+        audio is None
+        or audio_duration > MAX_AUDIO_DURATION_SECONDS
+        or int(max_new_tokens) > MAX_NEW_TOKENS
+    ):
         return 10
 
     clip_duration = float(getattr(config, "sound_clip_duration", 30.0))
     num_clips = max(1, math.ceil(audio_duration / clip_duration))
-    duration = 90 + (num_clips - 1) * 6 + (int(max_new_tokens) / 1024.0) * 60
+    duration = 45 + (num_clips - 1) * 3 + (int(max_new_tokens) / MAX_NEW_TOKENS) * 10
     return int(min(MAX_GPU_DURATION_SECONDS, math.ceil(duration)))
 
 
@@ -267,6 +272,8 @@ def _generate(
             f"Audio is {audio_duration / 60:.1f} minutes long; "
             f"this demo supports up to {MAX_AUDIO_DURATION_SECONDS / 60:.0f} minutes."
         )
+    if max_new_tokens > MAX_NEW_TOKENS:
+        raise gr.Error(f"This demo supports up to {MAX_NEW_TOKENS} output tokens.")
 
     input_features = extract_whisper_features(
         feature_extractor, wav, sample_rate=sr,
@@ -333,7 +340,7 @@ def run(
 
     Predefined ASR and translation tasks use the model's recommended greedy
     decoding. Other tasks and custom instructions use the sampling controls.
-    Audio may be up to five minutes long.
+    Audio may be up to two minutes long, with up to 512 output tokens.
 
     Returns:
         The final answer and, when enabled, the reasoning trace.
@@ -371,7 +378,7 @@ with gr.Blocks(title="Nemotron-Labs-Audex-30B-A3B") as demo:
         with gr.Column():
             audio_in = gr.Audio(
                 type="filepath",
-                label="Input audio (up to 5 minutes)",
+                label="Input audio (up to 2 minutes)",
                 sources=["upload", "microphone"],
             )
             task = gr.Radio(
@@ -387,7 +394,13 @@ with gr.Blocks(title="Nemotron-Labs-Audex-30B-A3B") as demo:
             run_btn = gr.Button("Run", variant="primary")
             with gr.Accordion("Advanced options", open=False):
                 reasoning = gr.Checkbox(value=False, label="Enable reasoning (<think>) mode")
-                max_new_tokens = gr.Slider(16, 1024, value=256, step=16, label="Max new tokens")
+                max_new_tokens = gr.Slider(
+                    16,
+                    MAX_NEW_TOKENS,
+                    value=256,
+                    step=16,
+                    label="Max new tokens",
+                )
                 temperature = gr.Slider(0.1, 1.5, value=0.7, step=0.05, label="Temperature")
                 top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
                 gr.Markdown(