Added an example that can compile on A40 with this PR but cannot under main

Author: cehongwang

examples/apps/flux_demo.py

@@ -62,6 +62,10 @@ def compile_model(
         torch_dtype=torch.float16,
     ).to(torch.float16)
 
+    pipe.transformer = FluxTransformer2DModel(
+        num_layers=23, num_single_layers=10, guidance_embeds=True
+    ).to(torch.float16)
+
     if args.low_vram_mode:
         pipe.enable_model_cpu_offload()
     else: