Fix ZeroGPU 90-second timeout by avoiding root-level packing. Move pipe to GPU dynamically inside the worker for fast generation.

app.py

@@ -116,8 +116,9 @@ gc.collect()
 pipe = QwenImageEditPlusPipeline.from_pretrained(
     "Qwen/Qwen-Image-Edit-2511",
     transformer=transformer,
-    torch_dtype=dtype
-).to(device)
+    torch_dtype=dtype,
+    ignore_patterns=["transformer/*"]
+)
 
 # Load next-scene LoRA for cinematic progression
 # Note: This LoRA was trained on 2509, may need testing with 2511/v18
@@ -237,18 +238,22 @@ def infer(
     print(f"Negative Prompt: '{negative_prompt}'")
     print(f"Seed: {seed}, Steps: {num_inference_steps}, Guidance: {true_guidance_scale}, Size: {width}x{height}")
 
-    # Generate the image
-    images_pil = pipe(
-        image=pil_images if len(pil_images) > 0 else None,
-        prompt=prompt,
-        height=height,
-        width=width,
-        negative_prompt=negative_prompt,
-        num_inference_steps=num_inference_steps,
-        generator=generator,
-        true_cfg_scale=true_guidance_scale,
-        num_images_per_prompt=num_images_per_prompt,
-    ).images
+    # Generate the image (move to GPU dynamically to avoid packing timeout at startup)
+    try:
+        pipe.to(device)
+        images_pil = pipe(
+            image=pil_images if len(pil_images) > 0 else None,
+            prompt=prompt,
+            height=height,
+            width=width,
+            negative_prompt=negative_prompt,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            true_cfg_scale=true_guidance_scale,
+            num_images_per_prompt=num_images_per_prompt,
+        ).images
+    finally:
+        pipe.to("cpu")
 
     # Anonymous diagnostics — fire-and-forget, must not block or fail generation.
     try: