Size the GPU reservation to the measured request

app.py

@@ -91,8 +91,13 @@ MAX_IMAGE_SLOTS, OPEN_IMAGE_SLOTS = 3, 2
 
 # Seconds of GPU one request needs, from the packed sequence it is about to denoise: linear in the rows for the
 # matmuls, quadratic for the attention. Fit shared with the other MiniMax-H3 Spaces on this pool.
-STEP_LINEAR, STEP_QUADRATIC, SAFETY = 1.1745e-4, 3.8396e-9, 1.3
-PLACEMENT_ALLOWANCE = int(os.environ.get("H3_PLACEMENT_ALLOWANCE", "90"))
+STEP_LINEAR, STEP_QUADRATIC, SAFETY = 1.1745e-4, 3.8396e-9, 1.15
+# The fit above was measured with the first-block cache off. With it on — the default — roughly a third of the
+# forwards skip the trunk, and the measured request below came in well under the uncached prediction. Both numbers
+# are calibrated against this Space's own smoke test: S=40696 at 20 steps reserved 445s under the old 1.3/90/no-FBC
+# constants and actually used 250s of GPU including placement, so a visitor's quota was paying for 195 idle seconds.
+FBC_SPEEDUP = float(os.environ.get("H3_FBC_SPEEDUP", "1.45"))
+PLACEMENT_ALLOWANCE = int(os.environ.get("H3_PLACEMENT_ALLOWANCE", "60"))
 AUDIO_LATENTS_PER_SECOND, AUDIO_CHANNELS = 40, 2
 REFERENCE_IMAGE_SHORT_EDGE, CANVAS_MULTIPLE = 2048, 32
 DECODE_BASE, DECODE_PER_DEFAULT_CANVAS, DEFAULT_CANVAS_PIXELS = 15, 25, 960 * 544 * 124
@@ -223,7 +228,10 @@ def reference_rows(image_paths: list[str]) -> int:
 
 
 def denoise_seconds(sequence: int, steps: int) -> float:
-    return int(steps) * (STEP_LINEAR * sequence + STEP_QUADRATIC * sequence**2) * SAFETY
+    import h3_fbc
+
+    uncached = int(steps) * (STEP_LINEAR * sequence + STEP_QUADRATIC * sequence**2) * SAFETY
+    return uncached / (FBC_SPEEDUP if h3_fbc.ENABLED else 1.0)
 
 
 def get_duration(prompt_embeds, text_token_tags, image_paths, height, width, num_frames, steps, weight, seed, **_):
@@ -497,7 +505,7 @@ def compose_prompt(references, action, camera, hud, soundscape, music, seconds)
         "subject_definitions:\n" + "\n".join(definitions),
         (
             "summary:\n"
-            f"[reference generation] A {float(seconds):g}-second photorealistic {spec['genre']} action RPG "
+            f"[reference generation] A {float(seconds):.1f}-second photorealistic {spec['genre']} action RPG "
             f"gameplay sequence{place}, where {subject} {action}, {spec['summary']}."
         ),
         "retention_analysis:\n" + "\n".join(retention),