Create app.py

app.py

@@ -0,0 +1,5 @@
+from transformers import pipeline
+import gradio as gr
+
+pipeline = pipeline("text-generation", model="google/gemma-2-2b", torch_dtype=torch.bfloat16, device_map="auto", model_kwargs={"quantization_config": BitsAndBytesConfig(load_in_8bit=True)})
+gr.Interface.from_pipeline(pipeline).launch()