|
from fastapi import FastAPI, HTTPException |
|
from pydantic import BaseModel |
|
from llama_cpp import Llama |
|
from multiprocessing import Process, Queue |
|
import uvicorn |
|
from dotenv import load_dotenv |
|
from difflib import SequenceMatcher |
|
|
|
load_dotenv() |
|
|
|
app = FastAPI() |
|
|
|
models = [ |
|
{"repo_id": "Ffftdtd5dtft/gpt2-xl-Q2_K-GGUF", "filename": "gpt2-xl-q2_k.gguf"}, |
|
{"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-instruct-q2_k.gguf"}, |
|
{"repo_id": "Ffftdtd5dtft/gemma-2-9b-it-Q2_K-GGUF", "filename": "gemma-2-9b-it-q2_k.gguf"}, |
|
{"repo_id": "Ffftdtd5dtft/gemma-2-27b-Q2_K-GGUF", "filename": "gemma-2-27b-q2_k.gguf"}, |
|
] |
|
|
|
llms = [] |
|
for model in models: |
|
llm = Llama.from_pretrained(repo_id=model['repo_id'], filename=model['filename']) |
|
llms.append(llm) |
|
|
|
class ChatRequest(BaseModel): |
|
message: str |
|
top_k: int = 50 |
|
top_p: float = 0.95 |
|
temperature: float = 0.7 |
|
|
|
def generate_chat_response(request, queue): |
|
try: |
|
user_input = request.message |
|
responses = [] |
|
for llm in llms: |
|
response = llm.create_chat_completion( |
|
messages=[{"role": "user", "content": user_input}], |
|
top_k=request.top_k, |
|
top_p=request.top_p, |
|
temperature=request.temperature |
|
) |
|
reply = response['choices'][0]['message']['content'] |
|
responses.append(reply) |
|
best_response = select_best_response(responses, request) |
|
queue.put(best_response) |
|
except Exception as e: |
|
queue.put(f"Error: {str(e)}") |
|
|
|
def select_best_response(responses, request): |
|
coherent_responses = filter_by_coherence(responses, request) |
|
best_response = filter_by_similarity(coherent_responses) |
|
return best_response |
|
|
|
def filter_by_coherence(responses, request): |
|
return responses |
|
|
|
def filter_by_similarity(responses): |
|
responses.sort(key=len, reverse=True) |
|
best_response = responses[0] |
|
for i in range(1, len(responses)): |
|
ratio = SequenceMatcher(None, best_response, responses[i]).ratio() |
|
if ratio < 0.9: |
|
best_response = responses[i] |
|
break |
|
return best_response |
|
|
|
@app.post("/generate_chat") |
|
async def generate_chat(request: ChatRequest): |
|
queue = Queue() |
|
p = Process(target=generate_chat_response, args=(request, queue)) |
|
p.start() |
|
p.join() |
|
response = queue.get() |
|
if "Error" in response: |
|
raise HTTPException(status_code=500, detail=response) |
|
return {"response": response} |
|
|
|
if __name__ == "__main__": |
|
uvicorn.run(app, host="0.0.0.0", port=7860) |
|
|
