What is the best way to implement multithreaded SSH logins using Netmiko with Python?

I have a list of network devices I need to SSH into in order to get the output of a command. I'm using Netmiko (Python). The SSH login credentials of each "login request" are JSON objects sent as an array through a single POST request to my Python server:

[{ IP: ...,   CMD: ... }, { IP: ...,   CMD: ... },  ..., { IP: ...,   CMD: ... }]

So far, my thought process for this has been:

1. Group the requests by host IP (many of them belong to the same host)
2. Process each request from the same host in a multithreaded fashion.
3. Group together the outputs from each request and return.

And this is what I've implemented:

from fastapi import FastAPI, Request, Bodyfrom netmiko import ConnectHandlerfrom pydantic import BaseModelimport loggingimport jsonfrom typing import Anyimport uvicornimport concurrent.futures  format = "%(asctime)s: %(message)s"logging.basicConfig(format=format, filename='CNM.log', encoding='utf-8', level=logging.INFO)app = FastAPI()def handle_request(payload):    try:        device = {'device_type': payload['dev_vendor'],'host': payload['IP'],'username': 'user','password': 'password'        }        conn = ConnectHandler(**device)        conn.read_timeout_override = 20        output = conn.send_command(payload['CMD'])        conn.disconnect()        return output    except Exception as e:        logging.error(e)@app.post("/ssh_login/")def ssh_login(reqs: Any = Body(None)):    hosts_list = []    for req in reqs:        hosts_list.append(req['IP'])    hosts_set = set(hosts_list)    responses = []    with concurrent.futures.ThreadPoolExecutor() as executor:        for host in hosts_set:            host_reqs = []            # Get the host's requests            for req in reqs:                if req['IP'] == host:                    host_reqs.append(req)            for output in executor.map(handle_request, host_reqs):                responses.append(output)    return responses

Although faster than a conventional synchronous approach, it turns out to be slow IMO, ~1s per request.

I thought that opening a single SSH connection for a host and running all commands in parallel with the executor could solve the problem, but I researched and it seems not viable (looks like sharing the same SSH connection across threads would be problematic).

Is there a way to optimize this and gain performance?