perf(zerogpu): pre-load model pipeline on CPU at startup and increase GPU timeout quota

app.py

@@ -404,7 +404,7 @@ def resolve_single_lora(
     return None, "None", 0.0
 
 
-def load_pipeline(model_repo: str = DEFAULT_MODEL_REPO):
+def get_or_load_pipeline(model_repo: str = DEFAULT_MODEL_REPO):
     global PIPE, CURRENT_MODEL_REPO, LOAD_ERROR, LOADED_IN
 
     if PIPE is not None and CURRENT_MODEL_REPO == model_repo:
@@ -441,7 +441,7 @@ def load_pipeline(model_repo: str = DEFAULT_MODEL_REPO):
         PIPE = pipe
         CURRENT_MODEL_REPO = model_repo
         LOADED_IN = time.time() - started
-        print(f"[wan] ready in {LOADED_IN:.0f}s", flush=True)
+        print(f"[wan] ready in {LOADED_IN:.0f}s on CPU", flush=True)
     except Exception as error:
         traceback.print_exc()
         LOAD_ERROR = f"**Loading `{model_repo}` failed**: `{type(error).__name__}: {error}`"
@@ -477,17 +477,10 @@ def _fit_keyframe(image_input, target_width: int, target_height: int) -> Image.I
 
 def get_duration(*args, **kwargs):
     try:
-        # _generate_video_gpu receives: (prompt, negative_prompt, image_input, height, width, num_frames, steps, ...)
-        height = kwargs.get("height", args[3] if len(args) > 3 else 480)
-        width = kwargs.get("width", args[4] if len(args) > 4 else 832)
-        num_frames = kwargs.get("num_frames", args[5] if len(args) > 5 else 81)
         steps = kwargs.get("steps", args[6] if len(args) > 6 else 20)
-        pixels_per_frame = int(height) * int(width)
-        total_tokens = (pixels_per_frame / 256) * (int(num_frames) / 4)
-        estimated = int(int(steps) * (total_tokens * 0.00012) + 20)
-        return max(60, min(MAX_GPU_DURATION, estimated))
+        return max(180, min(MAX_GPU_DURATION, int(steps) * 8 + 60))
     except Exception:
-        return 180
+        return 240
 
 
 @spaces.GPU(duration=get_duration, size=GPU_SIZE)
@@ -504,7 +497,7 @@ def _generate_video_gpu(
     lora_configs: list[tuple[str, float]],
     model_repo: str,
 ):
-    pipe = load_pipeline(model_repo)
+    pipe = get_or_load_pipeline(model_repo)
     pipe.to("cuda")
 
     active_lora_names = []
@@ -1059,4 +1052,9 @@ with gr.Blocks(css=custom_css, title="Wan 2.2 Video Studio") as app:
     )
 
 if __name__ == "__main__":
+    try:
+        print("[wan] pre-initializing default pipeline on CPU...", flush=True)
+        get_or_load_pipeline(DEFAULT_MODEL_REPO)
+    except Exception as err:
+        print(f"[wan] initial preload deferred to first call: {err}", flush=True)
     app.launch(allowed_paths=[OUTPUT_DIR])