move duckdb out of gpu function, add prompt lookup decoding

app.py

@@ -1,4 +1,5 @@
 from gradio_huggingfacehub_search import HuggingfaceHubSearch
+from llama_cpp.llama_speculative import LlamaPromptLookupDecoding
 from huggingface_hub import HfApi
 import matplotlib.pyplot as plt
 from typing import Tuple, Optional
@@ -72,11 +73,13 @@ CREATE TABLE {} (
 
 
 @spaces.GPU(duration=120)
-def generate_query(dataset_id: str, query: str) -> dict:
+def generate_query(ddl: str, query: str) -> dict:
     llama = llama_cpp.Llama(
         model_path="Hermes-2-Pro-Llama-3-8B-Q8_0.gguf",
         n_gpu_layers=50,
         chat_format="chatml",
+        draft_model=LlamaPromptLookupDecoding(num_pred_tokens=2),
+        logits_all=True,
         n_ctx=2048,
         verbose=True,
         temperature=0.1,
@@ -87,8 +90,6 @@ def generate_query(dataset_id: str, query: str) -> dict:
         mode=instructor.Mode.JSON_SCHEMA,
     )
 
-    ddl = get_dataset_ddl(dataset_id)
-
     system_prompt = f"""
     You are an expert SQL assistant with access to the following PostgreSQL Table:
     
@@ -122,7 +123,8 @@ def generate_query(dataset_id: str, query: str) -> dict:
 
 
 def query_dataset(dataset_id: str, query: str) -> Tuple[pd.DataFrame, str, plt.Figure]:
-    response = generate_query(dataset_id, query)
+    ddl = get_dataset_ddl(dataset_id)
+    response = generate_query(ddl, query)
 
     print("Querying Parquet...")
     df = conn.execute(response.get("sql")).fetchdf()