File size: 12,584 Bytes
e3278e4 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 |
#### What this does ####
# picks based on response time (for streaming, this is time to first token)
from datetime import datetime, timedelta
from typing import Dict, List, Optional, Union
import litellm
from litellm import ModelResponse, token_counter, verbose_logger
from litellm._logging import verbose_router_logger
from litellm.caching.caching import DualCache
from litellm.integrations.custom_logger import CustomLogger
class LowestCostLoggingHandler(CustomLogger):
test_flag: bool = False
logged_success: int = 0
logged_failure: int = 0
def __init__(
self, router_cache: DualCache, model_list: list, routing_args: dict = {}
):
self.router_cache = router_cache
self.model_list = model_list
def log_success_event(self, kwargs, response_obj, start_time, end_time):
try:
"""
Update 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
elif isinstance(id, int):
id = str(id)
# ------------
# Setup values
# ------------
"""
{
{model_group}_map: {
id: {
f"{date:hour:minute}" : {"tpm": 34, "rpm": 3}
}
}
}
"""
current_date = datetime.now().strftime("%Y-%m-%d")
current_hour = datetime.now().strftime("%H")
current_minute = datetime.now().strftime("%M")
precise_minute = f"{current_date}-{current_hour}-{current_minute}"
cost_key = f"{model_group}_map"
response_ms: timedelta = end_time - start_time
total_tokens = 0
if isinstance(response_obj, ModelResponse):
_usage = getattr(response_obj, "usage", None)
if _usage is not None and isinstance(_usage, litellm.Usage):
completion_tokens = _usage.completion_tokens
total_tokens = _usage.total_tokens
float(response_ms.total_seconds() / completion_tokens)
# ------------
# Update usage
# ------------
request_count_dict = self.router_cache.get_cache(key=cost_key) or {}
# check local result first
if id not in request_count_dict:
request_count_dict[id] = {}
if precise_minute not in request_count_dict[id]:
request_count_dict[id][precise_minute] = {}
## TPM
request_count_dict[id][precise_minute]["tpm"] = (
request_count_dict[id][precise_minute].get("tpm", 0) + total_tokens
)
## RPM
request_count_dict[id][precise_minute]["rpm"] = (
request_count_dict[id][precise_minute].get("rpm", 0) + 1
)
self.router_cache.set_cache(key=cost_key, value=request_count_dict)
### TESTING ###
if self.test_flag:
self.logged_success += 1
except Exception as e:
verbose_logger.exception(
"litellm.router_strategy.lowest_cost.py::log_success_event(): Exception occured - {}".format(
str(e)
)
)
pass
async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
try:
"""
Update cost 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
elif isinstance(id, int):
id = str(id)
# ------------
# Setup values
# ------------
"""
{
{model_group}_map: {
id: {
"cost": [..]
f"{date:hour:minute}" : {"tpm": 34, "rpm": 3}
}
}
}
"""
cost_key = f"{model_group}_map"
current_date = datetime.now().strftime("%Y-%m-%d")
current_hour = datetime.now().strftime("%H")
current_minute = datetime.now().strftime("%M")
precise_minute = f"{current_date}-{current_hour}-{current_minute}"
response_ms: timedelta = end_time - start_time
total_tokens = 0
if isinstance(response_obj, ModelResponse):
_usage = getattr(response_obj, "usage", None)
if _usage is not None and isinstance(_usage, litellm.Usage):
completion_tokens = _usage.completion_tokens
total_tokens = _usage.total_tokens
float(response_ms.total_seconds() / completion_tokens)
# ------------
# Update usage
# ------------
request_count_dict = (
await self.router_cache.async_get_cache(key=cost_key) or {}
)
if id not in request_count_dict:
request_count_dict[id] = {}
if precise_minute not in request_count_dict[id]:
request_count_dict[id][precise_minute] = {}
## TPM
request_count_dict[id][precise_minute]["tpm"] = (
request_count_dict[id][precise_minute].get("tpm", 0) + total_tokens
)
## RPM
request_count_dict[id][precise_minute]["rpm"] = (
request_count_dict[id][precise_minute].get("rpm", 0) + 1
)
await self.router_cache.async_set_cache(
key=cost_key, value=request_count_dict
) # reset map within window
### TESTING ###
if self.test_flag:
self.logged_success += 1
except Exception as e:
verbose_logger.exception(
"litellm.proxy.hooks.prompt_injection_detection.py::async_pre_call_hook(): Exception occured - {}".format(
str(e)
)
)
pass
async def async_get_available_deployments( # noqa: PLR0915
self,
model_group: str,
healthy_deployments: list,
messages: Optional[List[Dict[str, str]]] = None,
input: Optional[Union[str, List]] = None,
request_kwargs: Optional[Dict] = None,
):
"""
Returns a deployment with the lowest cost
"""
cost_key = f"{model_group}_map"
request_count_dict = await self.router_cache.async_get_cache(key=cost_key) or {}
# -----------------------
# Find lowest used model
# ----------------------
float("inf")
current_date = datetime.now().strftime("%Y-%m-%d")
current_hour = datetime.now().strftime("%H")
current_minute = datetime.now().strftime("%M")
precise_minute = f"{current_date}-{current_hour}-{current_minute}"
if request_count_dict is None: # base case
return
all_deployments = request_count_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"]] = {
precise_minute: {"tpm": 0, "rpm": 0},
}
try:
input_tokens = token_counter(messages=messages, text=input)
except Exception:
input_tokens = 0
# randomly sample from all_deployments, incase all deployments have latency=0.0
_items = all_deployments.items()
### GET AVAILABLE DEPLOYMENTS ### filter out any deployments > tpm/rpm limits
potential_deployments = []
_cost_per_deployment = {}
for item, item_map 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")
)
item_litellm_model_name = _deployment.get("litellm_params", {}).get("model")
item_litellm_model_cost_map = litellm.model_cost.get(
item_litellm_model_name, {}
)
# check if user provided input_cost_per_token and output_cost_per_token in litellm_params
item_input_cost = None
item_output_cost = None
if _deployment.get("litellm_params", {}).get("input_cost_per_token", None):
item_input_cost = _deployment.get("litellm_params", {}).get(
"input_cost_per_token"
)
if _deployment.get("litellm_params", {}).get("output_cost_per_token", None):
item_output_cost = _deployment.get("litellm_params", {}).get(
"output_cost_per_token"
)
if item_input_cost is None:
item_input_cost = item_litellm_model_cost_map.get(
"input_cost_per_token", 5.0
)
if item_output_cost is None:
item_output_cost = item_litellm_model_cost_map.get(
"output_cost_per_token", 5.0
)
# if litellm["model"] is not in model_cost map -> use item_cost = $10
item_cost = item_input_cost + item_output_cost
item_rpm = item_map.get(precise_minute, {}).get("rpm", 0)
item_tpm = item_map.get(precise_minute, {}).get("tpm", 0)
verbose_router_logger.debug(
f"item_cost: {item_cost}, item_tpm: {item_tpm}, item_rpm: {item_rpm}, model_id: {_deployment.get('model_info', {}).get('id')}"
)
# -------------- #
# Debugging Logic
# -------------- #
# We use _cost_per_deployment to log to langfuse, slack - this is not used to make a decision on routing
# this helps a user to debug why the router picked a specfic deployment #
_deployment_api_base = _deployment.get("litellm_params", {}).get(
"api_base", ""
)
if _deployment_api_base is not None:
_cost_per_deployment[_deployment_api_base] = item_cost
# -------------- #
# End of Debugging Logic
# -------------- #
if (
item_tpm + input_tokens > _deployment_tpm
or item_rpm + 1 > _deployment_rpm
): # if user passed in tpm / rpm in the model_list
continue
else:
potential_deployments.append((_deployment, item_cost))
if len(potential_deployments) == 0:
return None
potential_deployments = sorted(potential_deployments, key=lambda x: x[1])
selected_deployment = potential_deployments[0][0]
return selected_deployment
|