From c01ee848b06dc14d7a2b84d0c59bc1dd3f29e467 Mon Sep 17 00:00:00 2001
From: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
Date: Mon, 23 Mar 2026 22:14:54 -0700
Subject: [PATCH] Revert "fix: use consistent time denominator for throughput
 metrics in bench_one_batch_server" (#21276)

---
 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 1a5755b1d..4585340da 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 / latency
-    output_throughput = batch_size * output_len / latency
+    input_throughput = batch_size * input_len / last_ttft
+    output_throughput = batch_size * output_len / (latency - last_ttft)
     overall_throughput = batch_size * (input_len + output_len) / latency
 
     if backend == "vllm":