Calibrate ZeroGPU duration from measured runs; cap steps at 40

app.py

@@ -216,7 +216,14 @@ def _estimate_duration(
     *args,
     **kwargs,
 ):
-    """ZeroGPU duration estimate: overhead + per-DiT-pass cost × number of passes."""
+    """ZeroGPU duration estimate: fixed overhead + per-DiT-pass cost x number of passes.
+
+    Calibrated on zero-a10g (Blackwell, bf16, SDPA) against two measured runs:
+      41 frames / 10 steps / cfg 5.0  -> 14.4 s   (latent_t 11, 2 passes)
+      101 frames / 30 steps / cfg 5.0 -> 94.6 s   (latent_t 26, 2 passes)
+    Both fit t = 1.2 + 0.0599 * latent_t * passes * steps within 4%, i.e. cost is
+    essentially linear in latent length. The coefficients below add ~25-30% margin.
+    """
     latent_t = (int(num_frames) - 1) // 4 + 1
     passes = 1
     if float(cfg_scale) != 1.0:
@@ -225,8 +232,8 @@ def _estimate_duration(
         passes += 1
     if float(state_cfg_scale) != 1.0 and float(cfg_scale) == 1.0:
         passes += 1
-    total = 55.0 + 0.085 * latent_t * passes * int(num_inference_steps)
-    return int(min(300, math.ceil(total)))
+    total = 8.0 + 0.075 * latent_t * passes * int(num_inference_steps)
+    return int(min(260, math.ceil(total)))
 
 
 @spaces.GPU(duration=_estimate_duration)
@@ -440,7 +447,7 @@ with gr.Blocks(title="StatePlay") as demo:
                     num_frames = gr.Dropdown(
                         [41, 61, 81, 101], value=101, label="Frames (20 fps)"
                     )
-                    steps = gr.Slider(10, 50, value=30, step=1, label="Denoising steps")
+                    steps = gr.Slider(10, 40, value=30, step=1, label="Denoising steps")
                     cfg = gr.Slider(1.0, 10.0, value=5.0, step=0.1, label="Text CFG")
                     action_cfg = gr.Slider(
                         1.0, 5.0, value=1.0, step=0.1, label="Action CFG (1.0 = off)"