import random
import requests
from flask import Flask, request, jsonify, Response, stream_with_context, render_template_string
from mistral_common.protocol.instruct.messages import AssistantMessage, UserMessage, SystemMessage
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.protocol.instruct.request import ChatCompletionRequest
mt_v3 = MistralTokenizer.v3(is_tekken=True)
def calc_messages_tokens(json_data):
messages = json_data["messages"]
m_messages = []
for message in messages:
if message["role"] == "system":
m_messages.append(SystemMessage(content=message["content"]))
elif message["role"] == "user":
m_messages.append(UserMessage(content=message["content"]))
elif message["role"] == "assistant":
m_messages.append(AssistantMessage(content=message["content"]))
else:
continue
tokens = mt_v3.encode_chat_completion(ChatCompletionRequest(messages=m_messages)).tokens
return len(tokens) + len(m_messages)
app = Flask(__name__)
@app.route('/', methods=['GET'])
def index():
template = '''
Mistral-Nemo Chat API
Mistral-Nemo OpenAI Compatible API
Create your token(use as api key) [here] by selecting "serverless Inference API".
2. Set "https://tastypear-mistral-nemo-chat.hf.space/api" as the domain in the client configuration.
If you have multiple keys, you can concatenate them with a semicolon (`;`) to use them randomly, e.g., `hf_aaaa;hf_bbbb;hf_...`
'''
return render_template_string(template)
@app.route('/api/v1/chat/completions', methods=['POST'])
def proxy():
headers = dict(request.headers)
headers.pop('Host', None)
headers.pop('Content-Length', None)
keys = request.headers['Authorization'].split(' ')[1].split(';')
headers['Authorization'] = f'Bearer {random.choice(keys)}'
json_data = request.get_json()
# Avoid using cache
json_data["messages"][-1]['content'] = ' '*random.randint(1, 20)+json_data["messages"][-1]['content']
# Use the largest ctx
json_data['max_tokens'] = 32768 - calc_messages_tokens(json_data)
json_data['json_mode'] = False
model = 'mistralai/Mistral-Nemo-Instruct-2407'
def generate():
with requests.post(f"https://api-inference.huggingface.co/models/{model}/v1/chat/completions", json=request.json, headers=headers, stream=True) as resp:
for chunk in resp.iter_content(chunk_size=1024):
if chunk:
yield chunk
return Response(stream_with_context(generate()), content_type='text/event-stream')
#import gevent.pywsgi
#from gevent import monkey;monkey.patch_all()
if __name__ == "__main__":
app.run(debug=True)
# gevent.pywsgi.WSGIServer((args.host, args.port), app).serve_forever()