From 13af7cbb025b6d948df29e8cba75cbaff33720b4 Mon Sep 17 00:00:00 2001
From: Ajay Anubolu <124525760+AjAnubolu@users.noreply.github.com>
Date: Thu, 5 Mar 2026 15:58:17 -0800
Subject: [PATCH] fix: use consistent time denominator for throughput metrics
 in bench_one_batch_server (#19223)

---
 python/sglang/test/bench_one_batch_server_internal.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/python/sglang/test/bench_one_batch_server_internal.py b/python/sglang/test/bench_one_batch_server_internal.py
index 4585340da..1a5755b1d 100644
--- a/python/sglang/test/bench_one_batch_server_internal.py
+++ b/python/sglang/test/bench_one_batch_server_internal.py
@@ -600,8 +600,8 @@ def run_one_case(
 
     # Compute metrics
     latency = time.perf_counter() - tic
-    input_throughput = batch_size * input_len / last_ttft
-    output_throughput = batch_size * output_len / (latency - last_ttft)
+    input_throughput = batch_size * input_len / latency
+    output_throughput = batch_size * output_len / latency
     overall_throughput = batch_size * (input_len + output_len) / latency
 
     if backend == "vllm":