From a5e26d4ac39bb30cc82fd8c89a7dada178fb3cf8 Mon Sep 17 00:00:00 2001
From: Zhang Yulong <35552275+ZhangYulongg@users.noreply.github.com>
Date: Fri, 21 Aug 2026 15:15:08 +0800
Subject: [PATCH] Update warmup logic for min_tokens and max_tokens

Adjust warmup parameters to ensure min_tokens converges with max_tokens.
---
 benchmarks/benchmark_serving.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/benchmarks/benchmark_serving.py b/benchmarks/benchmark_serving.py
index 8a13f1cdded..1658bb166b6 100644
--- a/benchmarks/benchmark_serving.py
+++ b/benchmarks/benchmark_serving.py
@@ -443,6 +443,9 @@ async def benchmark(
     # warmup短输出：取128和hyper_parameters中max_tokens的最小值
     warmup_output_len = min(128, hyper_parameters.get("max_tokens", 128))
     warmup_hyper = {k: v for k, v in hyper_parameters.items() if k != "max_tokens"}
+    # warmup 缩短了 max_tokens，min_tokens 必须一起收敛，否则服务端返回 400
+    if warmup_hyper.get("min_tokens") is not None:
+        warmup_hyper["min_tokens"] = min(int(warmup_hyper["min_tokens"]), warmup_output_len)
     warmup_request_id = f"warmup_{uuid.uuid4().hex[:8]}"
     test_input = RequestFuncInput(
         model=model_id,