fix: lazy-load inference engine inside predict_sql

app.py

@@ -1,21 +1,9 @@
 import gradio as gr
 import spaces
-from app.engine.inference import generate_sql
 
-# 1. Satisfies ZeroGPU AST scanner & runs inference on GPU
-@spaces.GPU(duration=60)
-def predict_sql(question: str, schema_text: str = ""):
-    prompt = f"[SCHEMA] {schema_text} [QUESTION] {question} [SQL] "
-    candidates = generate_sql(prompt, n_candidates=5)
-    
-    # generate_sql returns a list of candidate strings; Gradio Textbox requires a single string
-    if isinstance(candidates, list):
-        return candidates[0] if len(candidates) > 0 else "SELECT 1;"
-    return str(candidates)
-
-# 2. UI Layout
+# 1. UI Definition (instantiated first so Gradio and HF register routes)
 with gr.Blocks(title="Text-to-SQL API") as demo:
-    gr.Markdown("# Text-to-SQL Inference Microservice")
+    gr.Markdown("# Text-to-SQL Backend Microservice")
     
     with gr.Row():
         question_in = gr.Textbox(
@@ -29,7 +17,17 @@ with gr.Blocks(title="Text-to-SQL API") as demo:
     
     sql_out = gr.Textbox(label="Generated SQL Output")
     btn = gr.Button("Generate SQL")
-    
+
+    # 2. GPU function wired inside the block context
+    @spaces.GPU(duration=60)
+    def predict_sql(question: str, schema_text: str = ""):
+        from app.engine.inference import generate_sql
+        prompt = f"[SCHEMA] {schema_text} [QUESTION] {question} [SQL] "
+        candidates = generate_sql(prompt, n_candidates=5)
+        if isinstance(candidates, list):
+            return candidates[0] if len(candidates) > 0 else "SELECT 1;"
+        return str(candidates)
+
     btn.click(
         fn=predict_sql,
         inputs=[question_in, schema_in],
@@ -37,8 +35,5 @@ with gr.Blocks(title="Text-to-SQL API") as demo:
         api_name="generate"
     )
 
-if __name__ == "__main__":s
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860
-    )
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860)