Calibrate GPU durations to measured inference times

app.py

@@ -150,9 +150,9 @@ def _estimate_duration(task: str, *args, **kwargs) -> int:
 
     Receives the same inputs as the decorated function (positional layout of
     the handler signature when ZeroGPU passes args positionally, keywords as
-    a fallback — both are accepted). Calibrated on an RTX PRO 6000: ~18 s
-    cold-start overhead + flow_steps × ~55 ms per generated patch + decode
-    time.
+    a fallback — both are accepted). Measured on ZeroGPU (RTX 6000 Ada):
+    ~6.5 ms per patch×flow-step plus a small cold-start allowance and codec
+    decode time.
     """
     text = kwargs.get("text") or kwargs.get("score_text") or (args[0] if args else "") or ""
     flow_steps = kwargs.get("flow_steps")
@@ -174,8 +174,8 @@ def _estimate_duration(task: str, *args, **kwargs) -> int:
         # Speech: characters ≈ syllables ≈ 155 ms each (CJK 1 char/syllable).
         seconds = (len(text) if isinstance(text, str) else 0) * 0.155
     patches = min(MAX_PATCHES, max(2, int(seconds * 12.5)))
-    total = 18.0 + patches * steps * 0.055 + seconds * 0.3
-    return min(300, int(total * 1.15))
+    total = 10.0 + patches * steps * 0.008 + seconds * 0.3
+    return min(240, int(total * 1.3))
 
 
 def _tts_duration(*args, **kwargs) -> int:
@@ -272,7 +272,13 @@ def svs_generate(score_text: str, voice: str, ref_audio, ref_text: str,
     return _save_and_report(waveform, patches.size(1), started)
 
 
-@spaces.GPU(duration=lambda *a, **k: 150)
+def _unified_duration(case_id: str, *args, **kwargs) -> int:
+    """Duration for the mixed speech+song cases: longest packaged sequence is
+    15 s of audio at flow 20 ≈ 26 s measured; reserve ~1.4× for cold starts."""
+    return 45
+
+
+@spaces.GPU(duration=_unified_duration)
 def unified_generate(case_id: str, flow_steps=None, cfg_value=None,
                      temperature=None, seed=None, progress=gr.Progress()):
     """Run one packaged speech→singing→speech sequence end to end.