"""Does GPU memory outside torch's allocator grow with the number of distinct threads that run a call?"""
import json, subprocess, sys, threading, time, urllib.request
URL = "http://127.0.0.1:18033"
TOKEN = open("token").read().strip()
BODY = json.dumps({"id": "x", "state": "The deploy passed.", "question": "Did it pass?",
                   "options": [{"id": "yes", "description": "Yes"}, {"id": "no", "description": "No"}]}).encode()
def card():
    return int(subprocess.check_output(["nvidia-smi", "-i", "3", "--query-gpu=memory.used", "--format=csv,noheader,nounits"]).strip())
def reserved():
    return json.load(urllib.request.urlopen(URL + "/health"))["model"]["reserved_gib"] * 1024
def post():
    r = urllib.request.Request(URL + "/decide", data=BODY, headers={"Authorization": "Bearer " + TOKEN, "Content-Type": "application/json"})
    urllib.request.urlopen(r).read()
def snap(tag):
    time.sleep(1.5); c, r = card(), reserved(); print(f"{tag}: card {c} MiB, reserved {r:.0f} MiB, outside allocator {c - 2 - r:.0f} MiB", flush=True)
snap("start")
for i in range(20): post()
snap("after 20 sequential requests")
for burst in range(3):
    ts = [threading.Thread(target=post) for _ in range(40)]
    [t.start() for t in ts]; [t.join() for t in ts]
    snap(f"after concurrent burst {burst + 1} (40 requests)")