from fastapi import FastAPI, File, UploadFile, Form, Request from fastapi.responses import HTMLResponse, FileResponse from fastapi.staticfiles import StaticFiles from fastapi.templating import Jinja2Templates from huggingface_hub import InferenceClient import random API_URL = "https://api-inference.huggingface.co/models/" client = InferenceClient( "mistralai/Mistral-7B-Instruct-v0.1" ) app = FastAPI() def format_prompt(message, history): prompt = "" for user_prompt, bot_response in history: prompt += f"[INST] {user_prompt} [/INST]" prompt += f" {bot_response} " prompt += f"[INST] {message} [/INST]" return prompt def generate(prompt, history, temperature=0.9, max_new_tokens=512, top_p=0.95, repetition_penalty=1.0): temperature = float(temperature) if temperature < 1e-2: temperature = 1e-2 top_p = float(top_p) generate_kwargs = dict( temperature=temperature, max_new_tokens=max_new_tokens, top_p=top_p, repetition_penalty=repetition_penalty, do_sample=True, seed=random.randint(0, 10**7), ) formatted_prompt = format_prompt(prompt, history) stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False) output = "" word = "" for response in stream: token_text = response.token.text.strip() if token_text != "": # Decode the token text to handle encoded characters decoded_text = token_text.encode("utf-8", "backslashreplace").decode("utf-8") # Add the decoded letter to the current word word += decoded_text # If the token is a space or the end of the stream, add the word to the output and reset the word if token_text == " " or response.is_end_of_stream: output += word + " " word = "" return output @app.post("/generate/") async def generate_chat(request: Request, prompt: str = Form(...), history: str = Form(...), temperature: float = Form(0.9), max_new_tokens: int = Form(512), top_p: float = Form(0.95), repetition_penalty: float = Form(1.0)): history = eval(history) # Convert history string back to list response = generate(prompt, history, temperature, max_new_tokens, top_p, repetition_penalty) # Concatenate the generated response strings into a single coherent response coherent_response = " ".join(response) return {"response": coherent_response} app.mount("/", StaticFiles(directory="static", html=True), name="static") @app.get("/") def index() -> FileResponse: return FileResponse(path="/app/static/index.html", media_type="text/html")