cyberes
/
local-llm-server
1
1
Fork 0
Code Issues Pull Requests Packages Projects Releases 1 Wiki Activity

adjust estimated wait time calculations

This commit is contained in:
Cyberes 2023-08-27 22:17:21 -06:00
parent 441a870e85
commit f43336c92c
1 changed files with 22 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
llm_server/routes/v1/generate_stats.py
View File

@ -5,7 +5,7 @@ from llm_server import opts
from llm_server.database import get_distinct_ips_24h, sum_column
from llm_server.helpers import deep_sort
from llm_server.llm.info import get_running_model
from llm_server.netdata import get_gpu_wh, get_power_states
from llm_server.netdata import get_power_states
from llm_server.routes.cache import redis
from llm_server.routes.queue import priority_queue
from llm_server.routes.stats import SemaphoreCheckerThread, calculate_avg_gen_time, get_active_gen_workers, get_total_proompts, server_start_time
@ -28,16 +28,29 @@ def generate_stats():
# waits = [elapsed for end, elapsed in t]
# estimated_wait = int(sum(waits) / len(waits))
active_gen_workers = get_active_gen_workers()
proompters_in_queue = len(priority_queue)
average_tps = float(redis.get('average_tps'))
if opts.average_generation_time_mode == 'database':
average_generation_time = int(float(redis.get('average_generation_elapsed_sec')))
average_output_tokens = int(float(redis.get('average_output_tokens')))
estimated_wait_sec = int(((average_output_tokens / average_tps) * proompters_in_queue) / opts.concurrent_gens) if average_tps > 0 else 0
average_generation_time = float(redis.get('average_generation_elapsed_sec'))
average_output_tokens = float(redis.get('average_output_tokens'))
# average_generation_time_from_tps = (average_output_tokens / average_tps)
# What to use in our math that calculates the wait time.
# We could use the average TPS but we don't know the exact TPS value, only
# the backend knows that. So, let's just stick with the elapsed time.
gen_time_calc = average_generation_time
estimated_wait_sec = (
(gen_time_calc * proompters_in_queue) / opts.concurrent_gens # Calculate wait time for items in queue
) + (
active_gen_workers * gen_time_calc # Calculate wait time for in-process items
) if average_tps > 0 else 0
elif opts.average_generation_time_mode == 'minute':
average_generation_time = int(calculate_avg_gen_time())
estimated_wait_sec = int((average_generation_time * proompters_in_queue) / opts.concurrent_gens)
average_generation_time = calculate_avg_gen_time()
gen_time_calc = average_generation_time
estimated_wait_sec = ((gen_time_calc * proompters_in_queue) / opts.concurrent_gens) + (active_gen_workers * gen_time_calc)
else:
raise Exception
@ -60,12 +73,12 @@ def generate_stats():
'24_hrs': get_distinct_ips_24h(),
},
'proompts': {
'processing': get_active_gen_workers(),
'processing': active_gen_workers,
'queued': proompters_in_queue,
'total': get_total_proompts() if opts.show_num_prompts else None,
},
'uptime': int((datetime.now() - server_start_time).total_seconds()) if opts.show_uptime else None,
'average_generation_elapsed_sec': average_generation_time,
'average_generation_elapsed_sec': int(gen_time_calc),
'average_tps': average_tps,
'tokens_generated': sum_column('prompts', 'response_tokens') if opts.show_total_output_tokens else None,
'nvidia': netdata_stats
@ -74,7 +87,7 @@ def generate_stats():
'endpoints': {
'blocking': opts.full_client_api,
},
'estimated_wait_sec': estimated_wait_sec,
'estimated_wait_sec': int(estimated_wait_sec),
'timestamp': int(time.time()),
'config': {
'gatekeeper': 'none' if opts.auth_required is False else 'token',