add pkgs

examples/vllm_test/run_stats_server.py

@@ -0,0 +1,90 @@
+import re
+import sys
+import numpy as np
+import pandas as pd
+
+# 用于记录每个度量的值
+first_token_times_values = []
+prompt_throughput_values = []
+generation_throughput_values = []
+running_values = []
+
+# 从文件中读取数据
+file_path = "server.log.valid"  # 替换成你的文件路径
+with open(file_path, 'r') as file:
+    # 遍历文件中的每一行进行统计
+    for line in file:
+        # 使用正则表达式提取Avg First Token times和Avg generation throughput以及Running的值
+        match_first_token = re.search(r"Avg First Token times:([0-9.]+)", line)
+        match_prompt_throughput = re.search(r"Avg prompt throughput: ([0-9.]+)", line)
+        match_generation_throughput = re.search(r"Avg generation throughput: ([0-9.]+)", line)
+        match_running = re.search(r"Running: (\d+)", line)
+
+        # 统计Avg First Token times
+        if match_first_token:
+            first_token_times = float(match_first_token.group(1))
+            if abs(first_token_times) > 1e-5:
+                first_token_times_values.append(first_token_times)
+
+        if match_prompt_throughput:
+            prompt_throughput = float(match_prompt_throughput.group(1))
+            if abs(prompt_throughput) > 1e-5:
+                prompt_throughput_values.append(prompt_throughput)
+
+        # 统计Avg generation throughput和Running
+        if match_generation_throughput and match_running:
+            generation_throughput = float(match_generation_throughput.group(1))
+            running = int(match_running.group(1))
+            if abs(generation_throughput) > 1e-5:
+                generation_throughput_values.append(generation_throughput)
+                running_values.append(running)
+
+# 计算平均值
+avg_first_token_times = np.mean(first_token_times_values) if len(first_token_times_values) > 0 else 0
+max_first_token_times = np.max(first_token_times_values) if len(first_token_times_values) > 0 else 0
+min_first_token_times = np.min(first_token_times_values) if len(first_token_times_values) > 0 else 0
+p10_first_token_times = np.percentile(first_token_times_values, 10) if len(first_token_times_values) > 0 else 0
+p90_first_token_times = np.percentile(first_token_times_values, 90) if len(first_token_times_values) > 0 else 0
+p99_first_token_times = np.percentile(first_token_times_values, 99) if len(first_token_times_values) > 0 else 0
+cnt_first_token_times = len(first_token_times_values) if len(first_token_times_values) > 0 else 0
+
+avg_prompt_throughput = np.mean(prompt_throughput_values) if len(prompt_throughput_values) > 0 else 0
+max_prompt_throughput = np.max(prompt_throughput_values) if len(prompt_throughput_values) > 0 else 0
+min_prompt_throughput = np.min(prompt_throughput_values) if len(prompt_throughput_values) > 0 else 0
+p10_prompt_throughput = np.percentile(prompt_throughput_values, 10) if len(prompt_throughput_values) > 0 else 0
+p90_prompt_throughput = np.percentile(prompt_throughput_values, 90) if len(prompt_throughput_values) > 0 else 0
+p99_prompt_throughput = np.percentile(prompt_throughput_values, 99) if len(prompt_throughput_values) > 0 else 0
+cnt_prompt_throughput = len(prompt_throughput_values) if len(prompt_throughput_values) > 0 else 0
+
+avg_generation_throughput = np.mean(generation_throughput_values) if len(generation_throughput_values) > 0 else 0
+max_generation_throughput = np.max(generation_throughput_values) if len(generation_throughput_values) > 0 else 0
+min_generation_throughput = np.min(generation_throughput_values) if len(generation_throughput_values) > 0 else 0
+p10_generation_throughput = np.percentile(generation_throughput_values, 10) if len(generation_throughput_values) > 0 else 0
+p90_generation_throughput = np.percentile(generation_throughput_values, 90) if len(generation_throughput_values) > 0 else 0
+p99_generation_throughput = np.percentile(generation_throughput_values, 99) if len(generation_throughput_values) > 0 else 0
+cnt_generation_throughput = len(generation_throughput_values) if len(generation_throughput_values) > 0 else 0
+
+avg_running = np.mean(running_values) if len(running_values) > 0 else 0
+max_running = np.max(running_values) if len(running_values) > 0 else 0
+min_running = np.min(running_values) if len(running_values) > 0 else 0
+p10_running = np.percentile(running_values, 10) if len(running_values) > 0 else 0
+p90_running = np.percentile(running_values, 90) if len(running_values) > 0 else 0
+p99_running = np.percentile(running_values, 99) if len(running_values) > 0 else 0
+cnt_running = len(running_values) if len(running_values) > 0 else 0
+
+# Create a DataFrame
+data = {
+    'avg': [avg_first_token_times, avg_prompt_throughput, avg_generation_throughput, avg_running],
+    'max': [max_first_token_times, max_prompt_throughput, max_generation_throughput, max_running],
+    'min': [min_first_token_times, min_prompt_throughput, min_generation_throughput, min_running],
+    'p10': [p10_first_token_times, p10_prompt_throughput, p10_generation_throughput, p10_running],
+    'p90': [p90_first_token_times, p90_prompt_throughput, p90_generation_throughput, p90_running],
+    'p99': [p99_first_token_times, p99_prompt_throughput, p99_generation_throughput, p99_running],
+    'num': [cnt_first_token_times, cnt_prompt_throughput, cnt_generation_throughput, cnt_running]
+}
+
+df = pd.DataFrame(data, index=['first_token_times', 'prompt_throughput', 'generation_throughput', 'running'])
+
+# Display the DataFrame
+print(df)
+