Diagnostics: test 60s vs 300s reservation

app.py

@@ -136,6 +136,20 @@ def _diag_alloc():
     return f"{torch.cuda.get_device_name(0)}, alloc ok, sum={float(x.sum())}"
 
 
+@spaces.GPU(duration=300)
+def _diag_alloc_long():
+    """Same trivial work, maximum reservation.
+
+    Before running the task, ZeroGPU has to materialise every CUDA tensor this
+    process registered at startup - which here means the 1.15 GB Cellpose-SAM
+    checkpoint and the DINOv2 backbone. If that restore is what blows the
+    budget, a trivial allocation still fails at 60 s but passes at 300 s.
+    """
+    import torch
+    x = torch.zeros(2048, 2048, device="cuda")
+    return f"{torch.cuda.get_device_name(0)}, alloc ok at 300s, sum={float(x.sum())}"
+
+
 @spaces.GPU(duration=60)
 def _diag_devices():
     import torch
@@ -176,7 +190,8 @@ def _diag_dino():
 
 def run_diagnostics():
     steps = [
-        ("1. CUDA allocation", _diag_alloc),
+        ("1. CUDA allocation (60s)", _diag_alloc),
+        ("1b. CUDA allocation (300s)", _diag_alloc_long),
         ("2. model devices", _diag_devices),
         ("3. Cellpose-SAM on GPU", _diag_cellpose),
         ("4. DINOv2 on GPU", _diag_dino),
@@ -187,9 +202,6 @@ def run_diagnostics():
             lines.append(f"PASS  {name}: {fn()}")
         except Exception as e:
             lines.append(f"FAIL  {name}: {type(e).__name__}: {e}")
-            lines.append("")
-            lines.append("^ this step is where it breaks.")
-            break
     return "\n".join(lines)