app.py

import gradio as gr


from transformers import pipeline
import torch

MAX_NEW_TOKENS = 250

# MODEL="HuggingFaceTB/SmolLM2-135M-Instruct"
# MODEL="HuggingFaceTB/SmolLM2-360M-Instruct"

MODEL="HuggingFaceTB/SmolLM2-1.7B-Instruct"
TEMPERATURE = 0.7
TOP_P = 0.95
REPETITION_PENALTY = 1.1


pipe = pipeline("text-generation", model="HuggingFaceTB/SmolLM2-1.7B-Instruct")


def message_fx(message, history):
    if len(history) == 0:
        send_to_api = [{'role':'user', 'content':message}]
        print(send_to_api)
        with torch.no_grad():
            response = pipe(send_to_api,
                    do_sample=True,
                    max_new_tokens=MAX_NEW_TOKENS,
                    temperature=TEMPERATURE, # 1.0 = lots of creativity, high odd of hallucination 0.1 very specific writing and low odds 
                    # top_k=50,
                    top_p=TOP_P,
                    repetition_penalty=REPETITION_PENALTY,   # Added to discourage repetition
                    # no_repeat_ngram_size=3
            )[0]['generated_text'][1]['content']
        return response
        
    else:
        send_to_api = history + [{'role':'user', 'content':message}]
        print(send_to_api)
        with torch.no_grad():
            response = pipe(send_to_api,
                    do_sample=True,
                    max_new_tokens=MAX_NEW_TOKENS,
                    temperature=TEMPERATURE, # 1.0 = lots of creativity, high odd of hallucination 0.1 very specific writing and low odds 
                    # top_k=50,
                    top_p=TOP_P,
                    repetition_penalty=REPETITION_PENALTY,   # Added to discourage repetition
                    # no_repeat_ngram_size=3        
            )[0]['generated_text'][-1]['content']
        return response


gr.ChatInterface(
    fn=message_fx, 
    type="messages"
).launch()