fix(zerogpu): sync internal device pointer to cuda for noise generation

app.py

@@ -280,10 +280,16 @@ def run_generation(text, font, author, num_steps, start_seed, num_images):
     logger.info("Models are already globally initialized and managed by ZeroGPU.")
     gen = generator
     # ZeroGPU automatically maps these to the acquired GPU during execution.
-    gen.model.to("cuda")
-    gen.clip.to("cuda")
-    gen.t5.to("cuda")
-    gen.vae.to("cuda")
+    # We must also correctly update internal Python attributes so runtime-generated latents go to GPU.
+    target_device = torch.device("cuda")
+    gen.device = target_device
+    if hasattr(gen, "sampler") and gen.sampler is not None:
+        gen.sampler.device = target_device
+        
+    gen.model.to(target_device)
+    gen.clip.to(target_device)
+    gen.t5.to(target_device)
+    gen.vae.to(target_device)
     
     # Step 2: Since we reverted to bf16 load to avoid PyTorch native dtype mix issues, skip wrapping
     logger.info("Model weights decompressed to bfloat16 upon load. Skipping dynamic quantization to ensure stability.")