[CI] Install lmms-eval from PyPI, drop human-eval install, add clone token fallback (#37672)

scripts/ci/cuda/ci_install_dependency.sh

@@ -763,9 +763,8 @@ install_extra_deps() {
     fi
 
     if [ "$IS_BLACKWELL" != "1" ]; then
-        git_clone_with_retry https://github.com/EvolvingLMMs-Lab/lmms-eval.git lmms-eval "--branch v0.5"
-        $PIP_CMD install -e lmms-eval/ $PIP_INSTALL_SUFFIX
-        # lmms-eval v0.5 pulls antlr4-python3-runtime==4.7.2, clobbering the
+        $PIP_CMD install "lmms_eval==0.5.0" $PIP_INSTALL_SUFFIX
+        # lmms_eval 0.5.0 pulls antlr4-python3-runtime==4.7.2, clobbering the
         # 4.9.3 that sgl-eval's latex2sympy2_extended needs (4.7.2 ImportError
         # at sgl-eval import). Pin it back so the nightly sgl-eval path works.
         $PIP_CMD install "antlr4-python3-runtime==4.9.3" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
@@ -783,14 +782,6 @@ install_test_tools() {
     mkdir -p "${HOME}/.cache/sglang/"
     mv python/kernels.lock "${HOME}/.cache/sglang/" || true
 
-    # Install human-eval (subshell keeps cd local)
-    $PIP_CMD install "setuptools==70.0.0" $PIP_INSTALL_SUFFIX
-    [ -d human-eval ] || git_clone_with_retry https://github.com/merrymercy/human-eval.git human-eval
-    (
-        cd human-eval
-        $PIP_CMD install -e . --no-build-isolation $PIP_INSTALL_SUFFIX
-    )
-
     mark_step_done "${FUNCNAME[0]}"
 }