From 590bf3823c20726cfd8622cf2a402f214cbcd993 Mon Sep 17 00:00:00 2001
From: Michael <13900043+michaelzhang-ai@users.noreply.github.com>
Date: Tue, 27 Oct 2026 00:58:54 -0700
Subject: [PATCH] [CI] [AMD] deps: Install openai 2.8.0 in AMD CI for the
 Responses API

serving_responses.py now imports responses_tools/beam.py at module level
(#45254), which needs ResponseFunctionShellToolCall from openai>=2.8.0.
AMD CI installs pyproject_other.toml (openai==2.6.1), so the Responses
handler fails to import and TestOpenAIServerv1Responses.test_error gets
500 instead of 400. Pin openai 2.8.0 in the AMD CI installer only.
---
 scripts/ci/amd/amd_ci_install_dependency.sh | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/scripts/ci/amd/amd_ci_install_dependency.sh b/scripts/ci/amd/amd_ci_install_dependency.sh
index 2684370eda20..0fae0648d604 100755
--- a/scripts/ci/amd/amd_ci_install_dependency.sh
+++ b/scripts/ci/amd/amd_ci_install_dependency.sh
@@ -165,6 +165,13 @@ else
   install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]"
 fi
 
+# serving_responses.py imports responses_tools/beam.py, which needs
+# ResponseFunctionShellToolCall from openai>=2.8.0. pyproject_other.toml still
+# pins openai==2.6.1, so the Responses handler fails to import and every
+# /v1/responses request returns 500. Drop this once pyproject_other.toml moves
+# to 2.8.0.
+install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache "openai==2.8.0"
+
 if [[ -n "${SKIP_TT_DEPS}" ]]; then
   echo "Didn't build lmms_eval, human-eval, and others"
 else