reorganize nvidia stats

llm_server/netdata.py

@@ -1,4 +1,5 @@
 import json
+from datetime import datetime, timedelta
 
 import requests
 
@@ -26,4 +27,26 @@ def get_power_states():
             print('Failed to fetch Netdata metrics:', e)
             return output
         gpu_num += 1
-    return {'power_states': output}
+    return output
+
+
+def get_gpu_wh(gpu_id: int):
+    chart_name = f"nvidia_smi.gpu{gpu_id}_power"
+    now = datetime.now()
+    one_hour_ago = now - timedelta(hours=1)
+    num_seconds = int((now - one_hour_ago).total_seconds())
+    params = {
+        "chart": chart_name,
+        "after": int(one_hour_ago.timestamp()),
+        "before": int(now.timestamp()),
+        "points": num_seconds,
+        "group": "second",
+        "format": "json",
+        "options": "absolute|jsonwrap"
+    }
+    response = requests.get(f'{opts.netdata_root}/api/v1/data', params=params)
+    data = json.loads(response.text)
+    total_power_usage_watts = sum(point[1] for point in data['result']['data'])
+    # total_power_usage_watt_hours = round(total_power_usage_watts / 3600, 1)
+    total_power_usage_kwh = round(total_power_usage_watts / 1000 / 3600, 3)
+    return total_power_usage_kwh

llm_server/routes/v1/generate_stats.py

@@ -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_power_states
+from llm_server.netdata import get_gpu_wh, 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
@@ -42,7 +42,14 @@ def generate_stats():
         raise Exception
 
     if opts.netdata_root:
-        netdata_stats = get_power_states()
+        netdata_stats = {}
+        power_states = get_power_states()
+        for gpu, power_state in power_states.items():
+            netdata_stats[gpu] = {
+                'power_state': power_state,
+                # 'wh_wasted_1_hr': get_gpu_wh(int(gpu.strip('gpu')))
+            }
+
     else:
         netdata_stats = {}