From f73cc39061032fa8ce8ce02876440f45d2da073d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Daniel=20Grie=C3=9Fhaber?= <griesshaber@hdm-stuttgart.de>
Date: Fri, 24 May 2024 11:33:51 +0200
Subject: [PATCH] add manual prompt evolution endpoints
---
api/{optimization.py => backend.py} | 36 ++++++++++++++++++----
api/main.py | 4 +--
api/routers/optimization.py | 46 +++++++++++------------------
frontend | 2 +-
4 files changed, 52 insertions(+), 36 deletions(-)
rename api/{optimization.py => backend.py} (80%)
diff --git a/api/optimization.py b/api/backend.py
similarity index 80%
rename from api/optimization.py
rename to api/backend.py
index 403c836..df18776 100644
--- a/api/optimization.py
+++ b/api/backend.py
@@ -1,8 +1,12 @@
+import asyncio
from argparse import Namespace
-from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
-from functools import partial
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import asynccontextmanager
from typing import ClassVar
+from fastapi import FastAPI
+
+from api import config
from evolution import GeneticAlgorithm
from models import Llama2, LLMModel
from task import SentimentAnalysis
@@ -11,7 +15,7 @@ from task import SentimentAnalysis
class MultiProcessOptimizer:
_instance: ClassVar["MultiProcessOptimizer"] = None
_running: bool = False
- model_exec: ProcessPoolExecutor = None
+ model_exec: ThreadPoolExecutor = None
_evolution_model: LLMModel | None = None
_evaluation_model: LLMModel | None = None
@@ -21,8 +25,10 @@ class MultiProcessOptimizer:
cls._instance = super(MultiProcessOptimizer, cls).__new__(cls)
return cls._instance
- def __init__(self, *, debug: bool = False) -> None:
- # a flag indicating whether optimizer is currently running
+ def __init__(
+ self, *, event_loop: asyncio.BaseEventLoop, debug: bool = False
+ ) -> None:
+ self.event_loop = event_loop
self.debug = debug
def __enter__(self):
@@ -103,3 +109,23 @@ class MultiProcessOptimizer:
def get_run_name(self):
if self.optimizer.run_directory is not None:
return self.optimizer.run_directory.name
+
+ async def evolve(self, prompt1: str, prompt2: str):
+ return await self.event_loop.run_in_executor(
+ self.model_exec, self.optimizer.evolve, prompt1, prompt2
+ )
+
+
+instance: MultiProcessOptimizer | None = None
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+ global instance
+ # Load the instance (which runs models in a separate thread)
+ instance = MultiProcessOptimizer(
+ event_loop=asyncio.get_running_loop(),
+ debug=config.DEBUG,
+ )
+ with instance:
+ yield
diff --git a/api/main.py b/api/main.py
index a453208..cdb57b3 100644
--- a/api/main.py
+++ b/api/main.py
@@ -2,13 +2,13 @@ from fastapi import FastAPI, Request, Response
from fastapi.staticfiles import StaticFiles
from requests import request as make_request
-from api import config
+from api import backend, config
from api.routers import optimization, runs
app = FastAPI(
debug=config.DEBUG,
title="Prompt Optimization Backend",
- lifespan=optimization.lifespan,
+ lifespan=backend.lifespan,
)
app.include_router(runs.router, prefix="/api/runs")
app.include_router(optimization.router, prefix="/api/optimization")
diff --git a/api/routers/optimization.py b/api/routers/optimization.py
index c706fac..8366dc9 100644
--- a/api/routers/optimization.py
+++ b/api/routers/optimization.py
@@ -1,42 +1,32 @@
-from contextlib import asynccontextmanager
+from fastapi import APIRouter, BackgroundTasks
-from fastapi import APIRouter, BackgroundTasks, FastAPI
-
-from api import config
-from api.optimization import MultiProcessOptimizer # start optimization
+from api import backend
router = APIRouter()
-backend: MultiProcessOptimizer | None = None
-
-
-@asynccontextmanager
-async def lifespan(app: FastAPI):
- global backend
- # Load the backend (which runs models in a separate process)
- backend = MultiProcessOptimizer(debug=config.DEBUG)
- with backend:
- actions = [("/api/action/evolve", backend.optimizer.evolve)]
- for path, target in actions:
- app.add_api_route(path, target)
- app.openapi_schema = None
- app.openapi()
-
- yield
- # TODO somehow not all ressources are released upon uvicorn reload, need to investigate further..
@router.post("/run")
def run(num_iterations: int, background_tasks: BackgroundTasks) -> str:
- assert backend is not None
- background_tasks.add_task(backend.run_optimization, num_iterations)
+ assert backend.instance is not None
+ background_tasks.add_task(backend.instance.run_optimization, num_iterations)
return f"Running optimization with {num_iterations} iterations."
+@router.get("/actions/evolve")
+async def evolve(prompt1: str, prompt2: str, background_tasks: BackgroundTasks) -> str:
+ assert backend.instance is not None
+ result, usage = await backend.instance.evolve(prompt1, prompt2)
+ return {
+ "result": result,
+ "usage": usage,
+ }
+
+
@router.get("/status")
def get_status():
- assert backend is not None
+ assert backend.instance is not None
return {
- "running": backend._running,
- "progress": backend.get_progress(),
- "name": backend.get_run_name(),
+ "running": backend.instance._running,
+ "progress": backend.instance.get_progress(),
+ "name": backend.instance.get_run_name(),
}
diff --git a/frontend b/frontend
index 6f83106..01e209a 160000
--- a/frontend
+++ b/frontend
@@ -1 +1 @@
-Subproject commit 6f8310669f705264bc136b7d15a551bbfc2ebf08
+Subproject commit 01e209a3b80adf90f17b0697522d2ad644c0ba15
--
GitLab