dynamically fetch GPUs for netdata

llm_server/netdata.py

@@ -5,18 +5,26 @@ import requests
 from llm_server import opts
 
 
-def get_power_state():
-    url = f"{opts.netdata_root}/api/v1/data?chart={opts.netdata_metric}"
-    try:
-        response = requests.get(url, timeout=3)
-        data = json.loads(response.text)
-        power_state_data = data['data'][0]
-        power_state = None
-        for i in range(1, len(power_state_data)):
-            if power_state_data[i] == 1:
-                power_state = data['labels'][i]
+def get_power_states():
+    gpu_num = 0
+    output = {}
+    while True:
+        url = f"{opts.netdata_root}/api/v1/data?chart=nvidia_smi.gpu{gpu_num}_power_state"
+        try:
+            response = requests.get(url, timeout=3)
+            if response.status_code != 200:
                 break
-        return power_state
-    except Exception as e:
-        print('Failed to fetch Netdata metrics:', e)
-        return None
+            data = json.loads(response.text)
+            power_state_data = data['data'][0]
+            power_state = None
+            for i in range(1, len(power_state_data)):
+                if power_state_data[i] == 1:
+                    power_state = data['labels'][i]
+                    break
+            output[f'gpu{gpu_num}'] = int(power_state.lower().strip('p'))
+            print(output)
+        except Exception as e:
+            print('Failed to fetch Netdata metrics:', e)
+            return output
+        gpu_num += 1
+    return output

llm_server/opts.py

@@ -18,5 +18,4 @@ show_num_prompts = True
 show_uptime = True
 average_generation_time_mode = 'database'
 show_total_output_tokens = True
-netdata_root = None
-netdata_metric = None
+netdata_root = None

llm_server/routes/v1/generate_stats.py

@@ -5,7 +5,7 @@ from llm_server import opts
 from llm_server.database import 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_state
+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
@@ -42,11 +42,7 @@ def generate_stats():
         raise Exception
 
     if opts.netdata_root:
-        netdata_stats = {
-            'gpu0': {
-                'power_state': int(get_power_state().lower().strip('p'))
-            }
-        }
+        netdata_stats = get_power_states()
     else:
         netdata_stats = {}
 

server.py

@@ -53,7 +53,6 @@ opts.show_uptime = config['show_uptime']
 opts.backend_url = config['backend_url'].strip('/')
 opts.show_total_output_tokens = config['show_total_output_tokens']
 opts.netdata_root = config['netdata_root']
-opts.netdata_metric = config['netdata_metric']
 
 opts.verify_ssl = config['verify_ssl']
 if not opts.verify_ssl: