log gen time to DB, also keep generation_elapsed under 3 min

llm_server/database.py

@@ -23,6 +23,7 @@ def init_db():
                 response TEXT,
                 response_tokens INTEGER,
                 response_status INTEGER,
+                generation_time INTEGER,
                 parameters TEXT CHECK (parameters IS NULL OR json_valid(parameters)),
                 headers TEXT CHECK (headers IS NULL OR json_valid(headers)),
                 timestamp INTEGER
@@ -43,7 +44,7 @@ def init_db():
         conn.close()
 
 
-def log_prompt(ip, token, prompt, response, parameters, headers, backend_response_code):
+def log_prompt(ip, token, prompt, response, gen_time, parameters, headers, backend_response_code):
     prompt_tokens = len(tokenizer.encode(prompt))
     response_tokens = len(tokenizer.encode(response))
 
@@ -53,8 +54,8 @@ def log_prompt(ip, token, prompt, response, parameters, headers, backend_respons
     timestamp = int(time.time())
     conn = sqlite3.connect(opts.database_path)
     c = conn.cursor()
-    c.execute("INSERT INTO prompts VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
-              (ip, token, prompt, prompt_tokens, response, response_tokens, backend_response_code, json.dumps(parameters), json.dumps(headers), timestamp))
+    c.execute("INSERT INTO prompts VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+              (ip, token, prompt, prompt_tokens, response, response_tokens, backend_response_code, gen_time, json.dumps(parameters), json.dumps(headers), timestamp))
     conn.commit()
     conn.close()
 

llm_server/routes/stats.py

@@ -10,21 +10,22 @@ server_start_time = datetime.now()
 
 # TODO: have a background thread put the averages in a variable so we don't end up with massive arrays
 
-wait_in_queue_elapsed = []
-wait_in_queue_elapsed_lock = Lock()
+# wait_in_queue_elapsed = []
+# wait_in_queue_elapsed_lock = Lock()
 
 generation_elapsed = []
 generation_elapsed_lock = Lock()
 
 
-def elapsed_times_cleanup():
-    global wait_in_queue_elapsed
-    while True:
-        current_time = time.time()
-        with wait_in_queue_elapsed_lock:
-            global wait_in_queue_elapsed
-            wait_in_queue_elapsed = [(end_time, elapsed_time) for end_time, elapsed_time in wait_in_queue_elapsed if current_time - end_time <= 60]
-        time.sleep(1)
+# TODO: do I need this?
+# def elapsed_times_cleanup():
+#     global wait_in_queue_elapsed
+#     while True:
+#         current_time = time.time()
+#         with wait_in_queue_elapsed_lock:
+#             global wait_in_queue_elapsed
+#             wait_in_queue_elapsed = [(end_time, elapsed_time) for end_time, elapsed_time in wait_in_queue_elapsed if current_time - end_time <= 60]
+#         time.sleep(1)
 
 
 def calculate_avg_gen_time():
@@ -37,10 +38,18 @@ def calculate_avg_gen_time():
 
 
 def process_avg_gen_time():
+    global generation_elapsed
     while True:
         with generation_elapsed_lock:
+            # Get the current time
+            current_time = time.time()
+
+            # Remove data older than 3 minutes
+            three_minutes_ago = current_time - 180
+            generation_elapsed[:] = [(end, elapsed) for end, elapsed in generation_elapsed if end >= three_minutes_ago]
+
             # Get the data from the last minute
-            one_minute_ago = time.time() - 60
+            one_minute_ago = current_time - 60
             recent_data = [elapsed for end, elapsed in generation_elapsed if end >= one_minute_ago]
 
         # Calculate the average

llm_server/routes/v1/generate.py

@@ -76,7 +76,7 @@ def generate():
         else:
             raise Exception
 
-        log_prompt(client_ip, token, request_json_body['prompt'], backend_response, parameters, dict(request.headers), response.status_code)
+        log_prompt(client_ip, token, request_json_body['prompt'], backend_response, elapsed_time, parameters, dict(request.headers), response.status_code)
         return jsonify({
             'code': 500,
             'error': 'failed to reach backend',
@@ -95,7 +95,7 @@ def generate():
             else:
                 raise Exception
 
-        log_prompt(client_ip, token, request_json_body['prompt'], backend_response, parameters, dict(request.headers), response.status_code)
+        log_prompt(client_ip, token, request_json_body['prompt'], backend_response, elapsed_time, parameters, dict(request.headers), response.status_code)
         return jsonify({
             **response_json_body
         }), 200
@@ -111,7 +111,7 @@ def generate():
             }
         else:
             raise Exception
-        log_prompt(client_ip, token, request_json_body['prompt'], backend_response, parameters, dict(request.headers), response.status_code)
+        log_prompt(client_ip, token, request_json_body['prompt'], backend_response, elapsed_time, parameters, dict(request.headers), response.status_code)
         return jsonify({
             'code': 500,
             'error': 'the backend did not return valid JSON',