Reduce memory during AOT compilation by offloading text encoder

Move text encoder (~16GB) to CPU before torch.export to free memory
during the memory-intensive AOT compilation step. Move it back to
GPU after compilation completes.

This should help stay under the 96GB build memory limit.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

app.py

@@ -1,3 +1,4 @@
+import gc
 import gradio as gr
 import numpy as np
 import random
@@ -112,8 +113,16 @@ pipe.transformer.__class__ = QwenImageTransformer2DModel
 pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
 
 # --- Ahead-of-time compilation ---
+# Temporarily offload text encoder to CPU to free ~16GB during memory-intensive torch.export
+pipe.text_encoder.to('cpu')
+gc.collect()
+torch.cuda.empty_cache()
+
 optimize_pipeline_(pipe, image=[Image.new("RGB", (1024, 1024)), Image.new("RGB", (1024, 1024))], prompt="prompt")
 
+# Move text encoder back to GPU for inference
+pipe.text_encoder.to(device)
+
 # --- UI Constants and Helpers ---
 MAX_SEED = np.iinfo(np.int32).max