From 81c88da1ab06635dacbff8fa0287599b661ba001 Mon Sep 17 00:00:00 2001
From: Mick <mickjagger19@icloud.com>
Date: Wed, 12 Aug 2026 10:25:14 +0800
Subject: [PATCH] [diffusion] fix: nightly diffusion benchmark passes the
 retired --warmup flag (#34423)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
---
 .../utils/diffusion/comparison_configs.json   | 22 +++++++++----------
 scripts/ci/utils/diffusion/run_comparison.py  |  6 ++---
 2 files changed, 14 insertions(+), 14 deletions(-)

diff --git a/scripts/ci/utils/diffusion/comparison_configs.json b/scripts/ci/utils/diffusion/comparison_configs.json
index 7694d3ac8..95d4a4f6e 100644
--- a/scripts/ci/utils/diffusion/comparison_configs.json
+++ b/scripts/ci/utils/diffusion/comparison_configs.json
@@ -13,7 +13,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --dit-layerwise-offload false --tp-size 2",
+                    "serve_args": "--warmup-mode server --dit-layerwise-offload false --tp-size 2",
                     "extra_env": {}
                 }
             }
@@ -29,7 +29,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --dit-layerwise-offload false --tp-size 2",
+                    "serve_args": "--warmup-mode server --dit-layerwise-offload false --tp-size 2",
                     "extra_env": {}
                 }
             }
@@ -45,7 +45,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --tp-size 2",
+                    "serve_args": "--warmup-mode server --tp-size 2",
                     "extra_env": {}
                 }
             }
@@ -62,7 +62,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --tp-size 2",
+                    "serve_args": "--warmup-mode server --tp-size 2",
                     "extra_env": {}
                 }
             }
@@ -78,7 +78,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --tp-size 2",
+                    "serve_args": "--warmup-mode server --tp-size 2",
                     "extra_env": {}
                 }
             }
@@ -95,7 +95,7 @@
             "num_gpus": 4,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory",
+                    "serve_args": "--warmup-mode server --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory",
                     "extra_env": {}
                 }
             }
@@ -113,7 +113,7 @@
             "num_gpus": 1,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup",
+                    "serve_args": "--warmup-mode server",
                     "extra_env": {}
                 }
             }
@@ -131,7 +131,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --pipeline-class-name LTX2TwoStagePipeline --cfg-parallel-size 2",
+                    "serve_args": "--warmup-mode server --pipeline-class-name LTX2TwoStagePipeline --cfg-parallel-size 2",
                     "extra_env": {}
                 }
             }
@@ -147,7 +147,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --tp-size 2 --attention-backend fa",
+                    "serve_args": "--warmup-mode server --tp-size 2 --attention-backend fa",
                     "extra_env": {}
                 }
             }
@@ -164,7 +164,7 @@
             "num_gpus": 2,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --tp-size 2",
+                    "serve_args": "--warmup-mode server --tp-size 2",
                     "extra_env": {"SGLANG_DISABLE_COSMOS3_GUARDRAILS": "1"}
                 }
             }
@@ -182,7 +182,7 @@
             "num_gpus": 4,
             "frameworks": {
                 "sglang": {
-                    "serve_args": "--warmup --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory",
+                    "serve_args": "--warmup-mode server --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory",
                     "extra_env": {}
                 }
             }
diff --git a/scripts/ci/utils/diffusion/run_comparison.py b/scripts/ci/utils/diffusion/run_comparison.py
index c625bfde9..6bfe6dd0b 100644
--- a/scripts/ci/utils/diffusion/run_comparison.py
+++ b/scripts/ci/utils/diffusion/run_comparison.py
@@ -802,9 +802,9 @@ def run_single(
         wait_for_health(base_url, framework)
 
         # No client-side warmup: each framework relies on its own server-side
-        # warmup before traffic. sglang's serve_args pass --warmup, which `serve`
-        # resolves to server-based (synthetic) warmup that primes kernels at
-        # startup, before the health check passes. This goes through the internal
+        # warmup before traffic. sglang's serve_args pass --warmup-mode server,
+        # which primes kernels with a synthetic request at startup, before the
+        # health check passes. This goes through the internal
         # warmup path that bypasses sampling-param preset validation (e.g.
         # Ideogram-4's preset-locked num_inference_steps), so no per-case warmup
         # special-casing is needed here.