[CI] Route mmlu and GB300 MMMU-Pro evals through sgl-eval (#34477)

scripts/ci/amd/amd_ci_install_dependency.sh

@@ -134,6 +134,10 @@ else
   install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]"
 fi
 
+# shellcheck source=scripts/ci/utils/sgl_eval_ref.sh
+source "$(dirname "${BASH_SOURCE[0]}")/../utils/sgl_eval_ref.sh"
+install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache "$SGL_EVAL_SPEC"
+
 if [[ -n "${SKIP_TT_DEPS}" ]]; then
   echo "Didn't build lmms_eval, human-eval, and others"
 else

scripts/ci/cuda/ci_install_dependency.sh

@@ -682,12 +682,8 @@ stabilize_flashinfer_jit_paths() {
 install_extra_deps() {
     MOONCAKE_VERSION="0.3.12.post1"
     NIXL_VERSION="1.3.0"
-    # sgl-eval is git-only and cannot be declared in python/pyproject.toml (see
-    # the note there). The nightly GSM8K eval shells out to the sgl-eval CLI and
-    # fails without it. Bumping the SHA can change zero-shot \boxed{} grading, so
-    # re-baseline MODEL_SCORE_THRESHOLDS in
-    # test/registered/eval/test_text_models_gsm8k_eval.py first.
-    SGL_EVAL_REF="b2a2703c42cae379bbcb8b7ff092df6601a61694"
+    # shellcheck source=scripts/ci/utils/sgl_eval_ref.sh
+    source "${SCRIPT_DIR}/../utils/sgl_eval_ref.sh"
     if [ "$CU_MAJOR" = "13" ]; then
         MOONCAKE_PKG="mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}"
         MOONCAKE_STALE_PKG="mooncake-transfer-engine"
@@ -723,7 +719,7 @@ install_extra_deps() {
             --no-deps --force-reinstall $PIP_INSTALL_SUFFIX
     fi
 
-    $PIP_CMD install "sgl-eval @ git+https://github.com/sgl-project/sgl-eval.git@${SGL_EVAL_REF}" $PIP_INSTALL_SUFFIX
+    $PIP_CMD install "$SGL_EVAL_SPEC" $PIP_INSTALL_SUFFIX
 
     if [ "$IS_BLACKWELL" != "1" ]; then
         git clone --branch v0.5 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git

scripts/ci/npu/npu_ci_install_dependency.sh

@@ -84,3 +84,8 @@ rm -rf cann-custom-ops
 ### Install SGLang
 rm -rf python/pyproject.toml && mv python/pyproject_npu.toml python/pyproject.toml
 ${UV_PIP_INSTALL} -v -e "python[dev_npu]"
+
+### Install sgl-eval
+# shellcheck source=scripts/ci/utils/sgl_eval_ref.sh
+source "${SCRIPT_DIR}/../utils/sgl_eval_ref.sh"
+${UV_PIP_INSTALL} "$SGL_EVAL_SPEC"

scripts/ci/utils/sgl_eval_ref.sh

@@ -0,0 +1,13 @@
+# Single source of truth for the sgl-eval commit every CI variant installs.
+# Meant to be sourced, not executed -- each variant then installs
+# "$SGL_EVAL_SPEC" with its own pip invocation, since those differ (uv pip on
+# CUDA/CPU, `docker exec ... pip` on AMD, `python3 -m pip` on NPU).
+#
+# sgl-eval is git-only and cannot be declared in python/pyproject.toml (see the
+# note there). Every eval that shells out to the `sgl-eval` CLI fails without
+# it, and a bump moves scoring for all of them at once -- so re-baseline
+# MODEL_SCORE_THRESHOLDS in
+# test/registered/eval/test_text_models_gsm8k_eval.py, and the mmlu thresholds
+# of run_eval's other callers, before changing this.
+SGL_EVAL_REF="6690895609dcbc5df1e7b00dd57c9502b868ec4d"
+SGL_EVAL_SPEC="sgl-eval @ git+https://github.com/sgl-project/sgl-eval.git@${SGL_EVAL_REF}"