Pin batch size and GPU duration to measured cost

app.py

@@ -378,7 +378,7 @@ def _gpu_duration(*args, **kwargs):
         factor = (float(side) / DEFAULT_MAX_SIDE) ** 2.5
     except Exception:
         work, factor = DEFAULT_NUM_STEPS * DEFAULT_NUM_FRAMES, 1.0
-    return int(min(180, max(30, 18 + work * 0.045 * factor)))
+    return int(min(180, max(30, 18 + work * 0.042 * factor)))
 
 
 @spaces.GPU(duration=_gpu_duration)
@@ -460,10 +460,10 @@ def analyze(
     t_prep = time.perf_counter() - t1
 
     seq_len = int(samples[0]["input_ids"].shape[-1])
-    # `pred_slot_isolated_eager` materialises a full B×32×L×L attention matrix,
-    # so the batch size is memory- rather than compute-bound. Start optimistic
-    # and halve on OOM (see the loop below).
-    batch_size = 8 if seq_len <= 3600 else (4 if seq_len <= 5400 else 2)
+    # `pred_slot_isolated_eager` materialises a full B×32×L×L attention matrix.
+    # Measured: batch 8 is no faster than batch 4 here (compute-bound), so keep
+    # the smaller batch and halve further on OOM (see the loop below).
+    batch_size = 4 if seq_len <= 3600 else (2 if seq_len <= 5400 else 1)
 
     def run_batch(batch):
         kwargs = dict(