Update app.py

app.py

@@ -79,7 +79,7 @@ STREAM_WIDTH = 1280
 JPEG_QUALITY = 82
 MAX_BLOCKS_PER_SESSION = 512     # hard cap so a session can't run forever
 SESSION_IDLE_TIMEOUT = 600       # seconds; janitor reaps abandoned sessions
-GPU_DURATION = 120               # seconds per @spaces.GPU allocation
+GPU_DURATION = 90                # seconds per @spaces.GPU allocation
 
 # ── Real-time pacing configuration ───────────────────────────────────────────
 # The GPU decodes a whole block (NUM_FPB frames) at once, so all of a block's
@@ -386,7 +386,7 @@ def create_gpu_rollout_loop(command_queue, seed_path, prompt_text, seed):
     Only picklable primitives + the multiprocessing `command_queue` cross the
     fork boundary. Live controls (held key set) and stop arrive via that queue.
     """
-    @spaces.GPU(duration=GPU_DURATION, size="xlarge")
+    @spaces.GPU(duration=GPU_DURATION)
     def gpu_rollout():
         device = torch.device("cuda")
         prompt = (prompt_text or DEFAULT_PROMPT).strip() or DEFAULT_PROMPT