| """ | |
| The Streamlit app for the project demo. | |
| In the demo, the user can write a prompt | |
| and the model will generate a response using the grouped sampling algorithm. | |
| """ | |
| import streamlit as st | |
| from torch.cuda import CudaError | |
| from hanlde_form_submit import on_form_submit | |
| from on_server_start import main as on_server_start_main | |
| on_server_start_main() | |
| st.title("Grouped Sampling Demo") | |
| with st.form("request_form"): | |
| selected_model_name: str = st.text_input( | |
| label="Model name", | |
| value="gpt2", | |
| help=f"The name of the model to use." | |
| f"Supported models are all the models in:" | |
| f" https://huggingface.co/models?pipeline_tag=text-generation&library=pytorch", | |
| ) | |
| output_length: int = st.number_input( | |
| label="Number of word pieces in the generated text, 1-4096 (default: 100)", | |
| min_value=1, | |
| max_value=4096, | |
| value=100, | |
| help="The length of the output text in tokens (word pieces)." | |
| ) | |
| submitted_prompt: str = st.text_area( | |
| label="Input for the model, It is highly recommended to write an English prompt.", | |
| help="Enter the prompt for the model. The model will generate a response based on this prompt.", | |
| value="Instruction: Answer in yes or no.\n" | |
| "Question: Is this a prompt?\n" | |
| "Answer: ", | |
| max_chars=2048, | |
| ) | |
| web_search: bool = st.checkbox( | |
| label="Web search", | |
| value=True, | |
| help="If checked, the model will get your prompt as well as some web search results." | |
| ) | |
| submitted: bool = st.form_submit_button( | |
| label="Generate", | |
| help="Generate the output text.", | |
| disabled=False, | |
| ) | |
| if submitted: | |
| try: | |
| output = on_form_submit( | |
| selected_model_name, | |
| output_length, | |
| submitted_prompt, | |
| web_search, | |
| ) | |
| except CudaError as e: | |
| st.error("Out of memory. Please try a smaller model, shorter prompt, or a smaller output length.") | |
| except (ValueError, TypeError, RuntimeError) as e: | |
| st.error(e) | |
| st.write(f"Generated text: {output}") | |