lease ladder 60/30/10 via free probes; 120s per session is too expensive

README.md

@@ -34,5 +34,5 @@ DiT + streaming VAE pipeline, served on ZeroGPU.
   not the first user session.
 - `torch.compile` is inert in ZeroGPU forks; each fork instead pays a short warmup
   (CUDA-graph capture, KV fill), so the first 2-3 chunks are slower.
-- A session runs up to ~120s of GPU time (shorter retries when the queue is busy).
-  Free ZeroGPU quota is limited — this is a small-scale demo.
+- A session runs up to ~60s of GPU time. Free ZeroGPU quota is limited — this is a
+  small-scale demo.

app.py

@@ -81,8 +81,8 @@ warnings.filterwarnings("ignore", message="ZeroGPU: Cannot get Gradio app Queue
 import spaces
 
 CKPT_ROOT = Path(os.environ.get("JOYOMNI_CKPT_ROOT", str(_DATA / "checkpoints")))
-# 120s = 240s billed on xlarge; 300s exceeds the per-visitor scheduling cap.
-SESSION_DURATION = int(os.environ.get("JOYOMNI_SPACES_DURATION", "120"))
+# 60s = 120s billed on xlarge.
+SESSION_DURATION = int(os.environ.get("JOYOMNI_SPACES_DURATION", "60"))
 
 
 # Prebuilt sm_120 wheel: pip runs before repo files exist on Spaces — install here.
@@ -723,12 +723,12 @@ async def ws(websocket: WebSocket):
         _visitor = SimpleNamespace(headers=_hdrs)
         pockets = [("visitor", _visitor), ("app-pool", None)]
 
-        def _probe(req):
+        def _probe(req, duration):
             token, _ = _get_token_and_payload(_get_headers(req))
             res, _meta = api_client().schedule(
                 cgroup_path=zutils.self_cgroup_device_path(),
                 token=token, token_version=2,
-                duration_seconds=SESSION_DURATION, enable_queue=False, gpu_size="xlarge")
+                duration_seconds=duration, enable_queue=False, gpu_size="xlarge")
             return res
 
         def _run(name, req, duration):
@@ -748,37 +748,38 @@ async def ws(websocket: WebSocket):
 
         try:
             t0 = time.time()
-            dry: set = set()
+            durations = {n: list(dict.fromkeys((SESSION_DURATION, 30, 10))) for n, _ in pockets}
             last_quota = None
             while not stop.is_set():
-                live = [(n, r) for n, r in pockets if n not in dry]
+                live = [(n, r) for n, r in pockets if durations[n]]
                 if not live:
                     raise RuntimeError(last_quota or "GPU quota exceeded")
                 name, req = live[0]
-                res = _probe(req)
+                dur = durations[name][0]
+                res = _probe(req, dur)
                 if isinstance(res, ScheduleResponse):
                     release(res.allowToken, allow_404=True)
                     try:
-                        _run(name, req, SESSION_DURATION)
+                        _run(name, req, dur)
                         break
                     except Exception as exc:  # noqa: BLE001
                         if got_gpu:
                             raise
                         emsg = str(exc)
-                        print(f"[ws] {name} lease failed: {emsg!r}", flush=True)
+                        print(f"[ws] {name}@{dur}s lease failed: {emsg!r}", flush=True)
                         if "quota" in emsg.lower() or "exceeded" in emsg.lower():
-                            dry.add(name)
+                            durations[name].pop(0)
                             last_quota = emsg
                         continue
                 if isinstance(res, QuotaInfos):
-                    dry.add(name)
+                    durations[name].pop(0)
                     if name == "visitor":
                         last_quota = (f"You have exceeded your GPU quota "
-                                      f"({SESSION_DURATION * 2}s requested vs. {res.left}s left). "
+                                      f"({dur * 2}s requested vs. {res.left}s left). "
                                       f"Try again in {res.wait}.")
                     continue
                 if res is httpx.codes.UNAUTHORIZED:
-                    dry.add(name)
+                    durations[name].clear()
                     continue
                 close = getattr(res, "close", None)
                 if close: