up

app.py

@@ -14,8 +14,6 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 # Load the model pipeline
 pipe = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=dtype).to(device)
 pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
-# --- Ahead-of-time compilation ---
-compiled_transformer = compile_transformer(pipe, prompt="prompt")
 
 @spaces.GPU(duration=120)
 def push_to_hub(repo_id, filename, oauth_token: gr.OAuthToken):
@@ -25,6 +23,9 @@ def push_to_hub(repo_id, filename, oauth_token: gr.OAuthToken):
     # this will throw if token is invalid
     _ = whoami(oauth_token.token)
 
+    # --- Ahead-of-time compilation ---
+    compiled_transformer = compile_transformer(pipe, prompt="prompt")
+
     token = oauth_token.token 
     out = _push_compiled_graph_to_hub(
         compiled_transformer.archive_file,