File size: 6,921 Bytes
7db0ae4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 |
#### What this does ####
# identifies lowest tpm deployment
import dotenv, os, requests, random
from typing import Optional, Union, List, Dict
from datetime import datetime
dotenv.load_dotenv() # Loading env variables using dotenv
import traceback
from litellm import token_counter
from litellm.caching import DualCache
from litellm.integrations.custom_logger import CustomLogger
class LowestTPMLoggingHandler(CustomLogger):
test_flag: bool = False
logged_success: int = 0
logged_failure: int = 0
default_cache_time_seconds: int = 1 * 60 * 60 # 1 hour
def __init__(self, router_cache: DualCache, model_list: list):
self.router_cache = router_cache
self.model_list = model_list
def log_success_event(self, kwargs, response_obj, start_time, end_time):
try:
"""
Update TPM/RPM usage on success
"""
if kwargs["litellm_params"].get("metadata") is None:
pass
else:
model_group = kwargs["litellm_params"]["metadata"].get(
"model_group", None
)
id = kwargs["litellm_params"].get("model_info", {}).get("id", None)
if model_group is None or id is None:
return
total_tokens = response_obj["usage"]["total_tokens"]
# ------------
# Setup values
# ------------
current_minute = datetime.now().strftime("%H-%M")
tpm_key = f"{model_group}:tpm:{current_minute}"
rpm_key = f"{model_group}:rpm:{current_minute}"
# ------------
# Update usage
# ------------
## TPM
request_count_dict = self.router_cache.get_cache(key=tpm_key) or {}
request_count_dict[id] = request_count_dict.get(id, 0) + total_tokens
self.router_cache.set_cache(key=tpm_key, value=request_count_dict)
## RPM
request_count_dict = self.router_cache.get_cache(key=rpm_key) or {}
request_count_dict[id] = request_count_dict.get(id, 0) + 1
self.router_cache.set_cache(key=rpm_key, value=request_count_dict)
### TESTING ###
if self.test_flag:
self.logged_success += 1
except Exception as e:
traceback.print_exc()
pass
async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
try:
"""
Update TPM/RPM usage on success
"""
if kwargs["litellm_params"].get("metadata") is None:
pass
else:
model_group = kwargs["litellm_params"]["metadata"].get(
"model_group", None
)
id = kwargs["litellm_params"].get("model_info", {}).get("id", None)
if model_group is None or id is None:
return
total_tokens = response_obj["usage"]["total_tokens"]
# ------------
# Setup values
# ------------
current_minute = datetime.now().strftime("%H-%M")
tpm_key = f"{model_group}:tpm:{current_minute}"
rpm_key = f"{model_group}:rpm:{current_minute}"
# ------------
# Update usage
# ------------
# update cache
## TPM
request_count_dict = self.router_cache.get_cache(key=tpm_key) or {}
request_count_dict[id] = request_count_dict.get(id, 0) + total_tokens
self.router_cache.set_cache(key=tpm_key, value=request_count_dict)
## RPM
request_count_dict = self.router_cache.get_cache(key=rpm_key) or {}
request_count_dict[id] = request_count_dict.get(id, 0) + 1
self.router_cache.set_cache(key=rpm_key, value=request_count_dict)
### TESTING ###
if self.test_flag:
self.logged_success += 1
except Exception as e:
traceback.print_exc()
pass
def get_available_deployments(
self,
model_group: str,
healthy_deployments: list,
messages: Optional[List[Dict[str, str]]] = None,
input: Optional[Union[str, List]] = None,
):
"""
Returns a deployment with the lowest TPM/RPM usage.
"""
# get list of potential deployments
current_minute = datetime.now().strftime("%H-%M")
tpm_key = f"{model_group}:tpm:{current_minute}"
rpm_key = f"{model_group}:rpm:{current_minute}"
tpm_dict = self.router_cache.get_cache(key=tpm_key)
rpm_dict = self.router_cache.get_cache(key=rpm_key)
# -----------------------
# Find lowest used model
# ----------------------
lowest_tpm = float("inf")
deployment = None
if tpm_dict is None: # base case
item = random.choice(healthy_deployments)
return item
all_deployments = tpm_dict
for d in healthy_deployments:
## if healthy deployment not yet used
if d["model_info"]["id"] not in all_deployments:
all_deployments[d["model_info"]["id"]] = 0
try:
input_tokens = token_counter(messages=messages, text=input)
except:
input_tokens = 0
for item, item_tpm in all_deployments.items():
## get the item from model list
_deployment = None
for m in healthy_deployments:
if item == m["model_info"]["id"]:
_deployment = m
if _deployment is None:
continue # skip to next one
_deployment_tpm = (
_deployment.get("tpm", None)
or _deployment.get("litellm_params", {}).get("tpm", None)
or _deployment.get("model_info", {}).get("tpm", None)
or float("inf")
)
_deployment_rpm = (
_deployment.get("rpm", None)
or _deployment.get("litellm_params", {}).get("rpm", None)
or _deployment.get("model_info", {}).get("rpm", None)
or float("inf")
)
if item_tpm == 0:
deployment = _deployment
break
elif (
item_tpm + input_tokens > _deployment_tpm
or rpm_dict[item] + 1 > _deployment_rpm
): # if user passed in tpm / rpm in the model_list
continue
elif item_tpm < lowest_tpm:
lowest_tpm = item_tpm
deployment = _deployment
return deployment
|