43 lines
2.7 KiB
Python
43 lines
2.7 KiB
Python
import time
|
|
|
|
from llm_server import opts
|
|
from llm_server.cluster.backend import get_a_cluster_backend, get_backends
|
|
from llm_server.cluster.cluster_config import cluster_config
|
|
from llm_server.custom_redis import redis
|
|
from llm_server.database.database import weighted_average_column_for_model
|
|
from llm_server.llm.info import get_info, get_running_model
|
|
|
|
|
|
def main_background_thread():
|
|
while True:
|
|
online, offline = get_backends()
|
|
for backend_url in online:
|
|
backend_info = cluster_config.get_backend(backend_url)
|
|
backend_mode = backend_info['mode']
|
|
backend_info = get_info(backend_url, backend_mode)
|
|
running_model = backend_info.get('model')
|
|
if not running_model:
|
|
continue
|
|
|
|
average_generation_elapsed_sec, average_output_tokens, estimated_avg_tps = calc_stats_for_backend(backend_url, running_model, backend_mode)
|
|
if average_generation_elapsed_sec: # returns None on exception
|
|
cluster_config.set_backend_value(backend_url, 'average_generation_elapsed_sec', average_generation_elapsed_sec)
|
|
if average_output_tokens:
|
|
cluster_config.set_backend_value(backend_url, 'average_output_tokens', average_output_tokens)
|
|
if average_generation_elapsed_sec and average_output_tokens:
|
|
cluster_config.set_backend_value(backend_url, 'estimated_avg_tps', estimated_avg_tps)
|
|
time.sleep(30)
|
|
|
|
|
|
def calc_stats_for_backend(backend_url, running_model, backend_mode):
|
|
# exclude_zeros=True filters out rows where an error message was returned. Previously, if there was an error, 0
|
|
# was entered into the column. The new code enters null instead but we need to be backwards compatible for now.
|
|
average_generation_elapsed_sec = weighted_average_column_for_model('prompts', 'generation_time',
|
|
running_model, backend_mode, backend_url, exclude_zeros=True,
|
|
include_system_tokens=opts.include_system_tokens_in_stats) or 0
|
|
average_output_tokens = weighted_average_column_for_model('prompts', 'response_tokens',
|
|
running_model, backend_mode, backend_url, exclude_zeros=True,
|
|
include_system_tokens=opts.include_system_tokens_in_stats) or 0
|
|
estimated_avg_tps = round(average_output_tokens / average_generation_elapsed_sec, 2) if average_generation_elapsed_sec > 0 else 0 # Avoid division by zero
|
|
return average_generation_elapsed_sec, average_output_tokens, estimated_avg_tps
|