-
-
Save python273/563177b3ad5b9f74c0f8f3299ec13850 to your computer and use it in GitHub Desktop.
| import os | |
| os.environ["OPENAI_API_KEY"] = "" | |
| from flask import Flask, Response, request | |
| import threading | |
| import queue | |
| from langchain.chat_models import ChatOpenAI | |
| from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler | |
| from langchain.schema import AIMessage, HumanMessage, SystemMessage | |
| app = Flask(__name__) | |
| @app.route('/') | |
| def index(): | |
| # just for the example, html is included directly, move to .html file | |
| return Response(''' | |
| <!DOCTYPE html> | |
| <html> | |
| <head><title>Flask Streaming Langchain Example</title></head> | |
| <body> | |
| <form id="form"> | |
| <input name="prompt" value="write a short koan story about seeing beyond"/> | |
| <input type="submit"/> | |
| </form> | |
| <div id="output"></div> | |
| <script> | |
| const formEl = document.getElementById('form'); | |
| const outputEl = document.getElementById('output'); | |
| let aborter = new AbortController(); | |
| async function run() { | |
| aborter.abort(); // cancel previous request | |
| outputEl.innerText = ''; | |
| aborter = new AbortController(); | |
| const prompt = new FormData(formEl).get('prompt'); | |
| try { | |
| const response = await fetch( | |
| '/chain', { | |
| signal: aborter.signal, | |
| method: 'POST', | |
| headers: {'Content-Type': 'application/json'}, | |
| body: JSON.stringify({ | |
| prompt | |
| }), | |
| } | |
| ); | |
| const reader = response.body.getReader(); | |
| const decoder = new TextDecoder(); | |
| while (true) { | |
| const { done, value } = await reader.read(); | |
| if (done) { break; } | |
| const decoded = decoder.decode(value, {stream: true}); | |
| outputEl.innerText += decoded; | |
| } | |
| } catch (err) { | |
| console.error(err); | |
| } | |
| } | |
| run(); // run on initial prompt | |
| formEl.addEventListener('submit', function(event) { | |
| event.preventDefault(); | |
| run(); | |
| }); | |
| </script> | |
| </body> | |
| </html> | |
| ''', mimetype='text/html') | |
| class ThreadedGenerator: | |
| def __init__(self): | |
| self.queue = queue.Queue() | |
| def __iter__(self): | |
| return self | |
| def __next__(self): | |
| item = self.queue.get() | |
| if item is StopIteration: raise item | |
| return item | |
| def send(self, data): | |
| self.queue.put(data) | |
| def close(self): | |
| self.queue.put(StopIteration) | |
| class ChainStreamHandler(StreamingStdOutCallbackHandler): | |
| def __init__(self, gen): | |
| super().__init__() | |
| self.gen = gen | |
| def on_llm_new_token(self, token: str, **kwargs): | |
| self.gen.send(token) | |
| def llm_thread(g, prompt): | |
| try: | |
| chat = ChatOpenAI( | |
| verbose=True, | |
| streaming=True, | |
| callbacks=[ChainStreamHandler(g)], | |
| temperature=0.7, | |
| ) | |
| chat([HumanMessage(content=prompt)]) | |
| finally: | |
| g.close() | |
| def chain(prompt): | |
| g = ThreadedGenerator() | |
| threading.Thread(target=llm_thread, args=(g, prompt)).start() | |
| return g | |
| @app.route('/chain', methods=['POST']) | |
| def _chain(): | |
| return Response(chain(request.json['prompt']), mimetype='text/plain') | |
| if __name__ == '__main__': | |
| app.run(threaded=True, debug=True) |
@VionaWang not quite sure what you mean. You can send the prompt in the request, then you use request.json
JS code to send json:
https://github.com/python273/eimi/blob/5fc4a1744191a5954955d77786e94bcaf8d6d5de/app/src/Session.svelte#L207-L218
@VionaWang not quite sure what you mean. You can send the prompt in the request, then you use
request.jsonJS code to send json: https://github.com/python273/eimi/blob/5fc4a1744191a5954955d77786e94bcaf8d6d5de/app/src/Session.svelte#L207-L218
Thanks for your reply!
From what I mean is, I have an html template that specifies the place for the answer to put in. But using "Response(chain("# A koan story about AGI\n\n"), mimetype='text/plain')" as your code above would give me a one line write in the webpage instead. I'm wondering how to make the streaming response to show up in a webpage with a spot devoted for it instead of one line write?
I'm new to frontend and all these web development stuff so any help would be appreciated. Thanks!
@VionaWang You should look at index function. You can put this html to a template, the output will be streamed to element with id output (<div id="output"></div>), you can place this element anywhere on the page
@VionaWang You should look at
indexfunction. You can put this html to a template, the output will be streamed to element with idoutput(<div id="output"></div>), you can place this element anywhere on the page
Ahhh sounds good, thank you so much for your help and your prompt response!
This is very helpful, thanks @python273 ! I also wanted to implement similar streaming using my local huggingface models in Langchain Pipeline - however, the llm chain can't be instantiated everytime in a thread (takes ~10 sec to load all shards). Any idea on how to go about it?
Thank you for posting this!
One minor thing: had to change mimetype='text/event-stream', can I ask why did you choose text/plain?
@noreff The mimetype shouldn't really affect anything. Also event stream is not correct mimetype, as the data is plain text, not in event stream format. My only guess why it might have changed something is there's a server in front of Flask than disables buffering if it detects text/event-stream mimetype.
@xerxes01 I would probably make a separate process with tcp server that keeps the model in memory and serves requests. Then connect to it from Flask.
I did something similar for stable diffusion back when it was released. Script for reference, though it's not that good:
https://gist.github.com/python273/ae9d085ce9f2968b50c6ab90f2017076
CallbackManager was renamed into BaseCallbackManager
from langchain.callbacks.base import BaseCallbackManager
Since the OpenAI library is deprecated, I have tried to replace it with ChatOpenAI without succeeding because it gives me these errors, do you know how to help me?
from flask import Flask, Response
import threading
import queue
from langchain.chat_models import ChatOpenAI
from langchain.callbacks.base import BaseCallbackManager
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
app = Flask(__name__)
@app.route('/')
def index():
return Response('''<!DOCTYPE html>
<html>
<head><title>Flask Streaming Langchain Example</title></head>
<body>
<div id="output"></div>
<script>
const outputEl = document.getElementById('output');
(async function() {
try {
const controller = new AbortController();
const signal = controller.signal;
const timeout = 120000; // Imposta il timeout su 120 secondi
setTimeout(() => controller.abort(), timeout);
const response = await fetch('/chain', {method: 'POST', signal});
const reader = response.body.getReader();
const decoder = new TextDecoder();
let buffer = '';
while (true) {
const { done, value } = await reader.read();
if (done) { break; }
const text = decoder.decode(value, {stream: true});
outputEl.innerHTML += text;
}
} catch (err) {
console.error(err);
}
})();
</script>
</body>
</html>''', mimetype='text/html')
class ThreadedGenerator:
def __init__(self):
self.queue = queue.Queue()
def __iter__(self):
return self
def __next__(self):
item = self.queue.get()
if item is StopIteration: raise item
return item
def send(self, data):
self.queue.put(data)
def close(self):
self.queue.put(StopIteration)
class ChainStreamHandler(StreamingStdOutCallbackHandler):
def __init__(self, gen):
super().__init__()
self.gen = gen
def on_llm_new_token(self, token: str, **kwargs):
self.gen.send(token)
def llm_thread(g, prompt):
try:
llm = ChatOpenAI(
model_name="gpt-4",
verbose=True,
streaming=True,
callback_manager=BaseCallbackManager([ChainStreamHandler(g)]),
temperature=0.7,
)
llm(prompt)
finally:
g.close()
def chain(prompt):
g = ThreadedGenerator()
threading.Thread(target=llm_thread, args=(g, prompt)).start()
return g
@app.route('/chain', methods=['POST'])
def _chain():
return Response(chain("Create a poem about the meaning of life \n\n"), mimetype='text/plain')
if __name__ == '__main__':
app.run(threaded=True, debug=True)the error can be found Here
if someone can suggest a solution I think it would be helpful for many developers who currently do not know how to do it. Thanks
Updated the example
thank you very much, it works great
That is really great code. Thank you so much. If possible can you update it so that with new user input the responses are not replaced by new responses and they are in a chat like manner with previous responses.
Anyhow, great work and really appreciate you shared all this.
what if we are using Tools, how can we get streamings for each tool being invoked?
Hi all !
I wanted to share with you a Custom Stream Response that I implemented in my FastAPI application recently.
I created this solution to manage streaming data.
You can use Stream, Event of Langchain but I'm doing special things with the Handlers that's why I need it.
Here examples:
Fast API
@router.get("/myExample")
async def mySpecialAPI(
session_id: UUID,
input="Hello",
) -> StreamResponse:
# Note: Don't write await we need a coroutine
invoke = chain.ainvoke(..)
callback = MyCallback(..)
return StreamResponse(invoke, callback)
Custom Stream Response
from __future__ import annotations
import asyncio
import typing
from typing import Any, AsyncIterable, Coroutine
from fastapi.responses import StreamingResponse as FastApiStreamingResponse
from starlette.background import BackgroundTask
class StreamResponse(FastApiStreamingResponse):
def __init__(
self,
invoke: Coroutine,
callback: MyCustomAsyncIteratorCallbackHandler,
status_code: int = 200,
headers: typing.Mapping[str, str] | None = None,
media_type: str | None = "text/event-stream",
background: BackgroundTask | None = None,
) -> None:
super().__init__(
content=StreamResponse.send_message(callback, invoke),
status_code=status_code,
headers=headers,
media_type=media_type,
background=background,
)
@staticmethod
async def send_message(
callback: AsyncIteratorCallbackHandler, invoke: Coroutine
) -> AsyncIterable[str]:
asyncio.create_task(invoke)
async for token in callback.aiter():
yield token
My Custom Callbackhandler
from __future__ import annotations
import asyncio
from typing import Any, AsyncIterator, List
class MyCustomAsyncIteratorCallbackHandler(AsyncCallbackHandler):
"""Callback handler that returns an async iterator."""
# Note: Can be a BaseModel than str
queue: asyncio.Queue[Optional[str]]
# Pass your params as you want
def __init__(self) -> None:
self.queue = asyncio.Queue()
async def on_llm_new_token(
self,
token: str,
tags: List[str] | None = None,
**kwargs: Any,
) -> None:
self.queue.put_nowait(token)
async def on_llm_end(
self,
response: LLMResult,
tags: List[str] | None = None,
**kwargs: Any,
) -> None:
self.queue.put_nowait(None)
# Note: Ect.. for error
async def aiter(self) -> AsyncIterator[str]:
while True:
token = await self.queue.get()
if isinstance(token, str):
yield token # Note: or a BaseModel.model_dump_json() etc..
elif token is None:
self.queue.task_done()
break
https://gist.github.com/YanSte/7be29bc93f21b010f64936fa334a185f
Thanks! This is really helpful :) I'm wondering if the Response(chain("# A koan story about AGI\n\n"), mimetype='text/plain') can be put inside a webpage instead of one line writing?