async def gen():
	env = request._run_ns.copy()
	f = partial(inference_generator, request, params_kw=params_kw)
	try:
		if params_kw.stream:
			async for l in f():
				yield f'data: {l}\n'
			yield 'data: [DONE]\n\n'
		else:
			async for l in f():
				yield l
	except Exception as e:
		luid = params_kw.get('_luid')
		if luid:
			try:
				await env.refund_balance(luid)
			except:
				pass
		import json as _json
		err = _json.dumps({"error": str(e), "status": "FAILED"}, ensure_ascii=False)
		yield f'data: {err}\n' if params_kw.stream else err

debug_params('params_kw', params_kw)
catelogid = params_kw.catelogid or 't2t'
if params_kw.off_peak:
	off_peak = params_kw.off_peak
	if off_peak in [True, "Y" "y", 1, "1"]:
		off_peak = True
	else:
		off_peak = False
	params_kw.off_peak = off_peak
userid = await get_user()
userorgid = await get_userorgid()
if userid is None:
	debug(f'need login')
	return openai_403()

if not params_kw.prompt and not params_kw.messages:
	debug(f'missing prompt and messages, model={params_kw.model}')
	d = return_error('Missing need data(prompt or messages)')
	return json_response(d, status=400)
env = request._run_ns
llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', catelogid)
if not llmid:
	debug(f'{params_kw.model=} not found for catelogid={catelogid}')
	return openai_400()
params_kw.llmid = llmid
params_kw.llmcatelogid = catelogid

debug(f'{params_kw.llmid=}')
luid = getID()
reserved = await env.reserve_balance(params_kw.llmid, userorgid, luid, userid=userid)
if not reserved.get('ok'):
	debug(f'{userid=} balance not enough: {reserved}')
	return openai_429()
params_kw._luid = luid
params_kw._reserved = reserved
f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
if not f and reserved.get('no_redis'):
	debug(f'{userid=} balance not enough (DB fallback)')
	return openai_429()
return await env.stream_response(request, gen, content_type='application/json')