diff --git a/python/pyproject.toml b/python/pyproject.toml
index d612751ce..55ab27ee2 100755
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -153,12 +153,13 @@ test = [
   "jsonlines",
   "lm-eval[api]>=0.4.9.2",
   "matplotlib",
-  # Pin sgl-eval to a git SHA: upgrading changes zero-shot \boxed{} grading, so
-  # re-baseline MODEL_SCORE_THRESHOLDS in test_text_models_gsm8k_eval.py first.
-  # antlr4 4.9.3 is forced because latex2sympy2_extended raises ImportError on
-  # 4.7.x, and an older transitive pin can win during install.
+  # 4.9.3 is what sgl-eval's math_verify grader needs (latex2sympy2_extended
+  # raises ImportError on 4.7.x), so a `sglang[test]` environment is already
+  # compatible when sgl-eval is installed on top of it.
+  # Do NOT declare sgl-eval itself here: it is git-only, and PyPI rejects any
+  # uploaded distribution whose metadata carries a direct URL requirement.
+  # CUDA CI installs it in scripts/ci/cuda/ci_install_dependency.sh.
   "antlr4-python3-runtime==4.9.3",
-  "sgl-eval @ git+https://github.com/sgl-project/sgl-eval.git@b2a2703c42cae379bbcb8b7ff092df6601a61694",
   "pandas",
   "parameterized",
   "peft>=0.18.0",
diff --git a/scripts/ci/cuda/ci_install_dependency.sh b/scripts/ci/cuda/ci_install_dependency.sh
index c819a8781..025d6b64f 100755
--- a/scripts/ci/cuda/ci_install_dependency.sh
+++ b/scripts/ci/cuda/ci_install_dependency.sh
@@ -471,6 +471,12 @@ stabilize_flashinfer_jit_paths() {
 install_extra_deps() {
     MOONCAKE_VERSION="0.3.11.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"
     if [ "$CU_MAJOR" = "13" ]; then
         MOONCAKE_PKG="mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}"
         MOONCAKE_STALE_PKG="mooncake-transfer-engine"
@@ -506,6 +512,8 @@ 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}" $PIP_INSTALL_SUFFIX
+
     if [ "$IS_BLACKWELL" != "1" ]; then
         git clone --branch v0.5 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
         $PIP_CMD install -e lmms-eval/ $PIP_INSTALL_SUFFIX