Upload app.py with huggingface_hub

app.py

@@ -110,11 +110,13 @@ _SCHEDULE_LOOKUP = {name: (h, l) for name, h, l in SCHEDULES}
 def _estimate_duration(*args, **kwargs) -> int:
     """Callable GPU duration, calibrated from measured calls on this Space.
 
-    Measured on ZeroGPU: ~0.081 s/token at 128 unrolled layers, roughly
-    proportional to depth; ~2.5 s fixed plus up to ~9 s one-time init on a
-    visitor's first call. The 20 s base covers init + prefill + margin.
+    Measured on ZeroGPU (cold first call per visitor): 14.1 s at H2L3/32 tok,
+    19.3 s at H2L3/96 tok, 39.7 s at H3L3/96 tok. Per-token cost grows
+    superlinearly with unrolled depth, so the coefficient is scaled by
+    depth/128 with a fixed base that also covers worker init.
 
-    Duration = 20 + budget x (0.12 s/token scaled by depth/128), capped.
+    Duration = 25 + budget x 0.2 s/token scaled by depth/128, capped at 120 s.
+    Worst offered case (H3L3, 96 tok) resolves to ~54 s vs 39.7 s measured.
     """
     schedule = kwargs.get("schedule") or (args[2] if len(args) > 2 else None)
     budget = kwargs.get("max_tokens") or (args[3] if len(args) > 3 else 32)
@@ -123,7 +125,7 @@ def _estimate_duration(*args, **kwargs) -> int:
     except Exception:
         h, l = 2, 3
     depth = 16 * h * (l + 1)
-    return min(120, 20 + int(float(budget) * 0.12 * depth / 128))
+    return min(120, 25 + int(float(budget) * 0.2 * depth / 128))
 
 
 @spaces.GPU(duration=_estimate_duration)