dont run all tokens for short dur

app.py

@@ -17,13 +17,13 @@ def audionar_tts(text='frogs',
     if text and text.strip():
 
 
-        dur_seconds = max(duration + 0.74, 2.0)
-        # Sink Attn
+        duration = max(duration + 0.74, 2.0)
+
         background_audio = audiogen.generate(
-            text[:64],  # soundscape text - discard if too long cross attention T5
-            duration=dur_seconds,
-            max_tokens=max(7, int(max_tokens)),  # kv cache lowest n_preserve
-            cache_lim=max(6, int(cache_lim)),
+            text[:64],         # soundscape text - discard if too long cross attention
+            duration=duration, # dont use all A/R tokens if duration is shorter than tokens<<
+            max_tokens=min(max(7, int(max_tokens)), int(duration * 50))  # kv cache lowest n_preserve
+            cache_lim=max(6, int(cache_lim)),         # Sink Attn
          ).numpy()
         
     else:
@@ -48,14 +48,14 @@ with gr.Blocks() as demo:
         )
         duration = gr.Number(
             label="Duration (s)",
-            value=7.1,
+            value=7.24,
         )
         n_tokens = gr.Number(
             label="Tokens",
             value=24,
         )
         cache_lim = gr.Number(
-            label="kv Cache Flush:",
+            label="kv Flush",
             value=71,
         )
         generate_button = gr.Button("Generate Audio",