diff --git a/.github/workflows/_npu-pr-test-stage.yml b/.github/workflows/_npu-pr-test-stage.yml
index 81ec0903a..0bf8a37c9 100644
--- a/.github/workflows/_npu-pr-test-stage.yml
+++ b/.github/workflows/_npu-pr-test-stage.yml
@@ -269,8 +269,11 @@ jobs:
             install_pkg "${pkg}"
           done
 
-          # Install sgl-eval (CLI used by run_eval.py's `sgl-eval run ...`) if missing.
-          command -v sgl-eval >/dev/null 2>&1 || install_pkg "sgl-eval==0.1.0"
+          # Fallback for images that predate the pin in pyproject_npu.toml. Checking
+          # the version, not the command, so a stale wheel is replaced rather than kept.
+          SGL_EVAL_VERSION=0.1.2
+          ${PYTHON_FOR_SGLANG} -c "import sgl_eval, sys; sys.exit(sgl_eval.__version__ != '${SGL_EVAL_VERSION}')" 2>/dev/null \
+            || install_pkg "sgl-eval==${SGL_EVAL_VERSION}"
 
           cd test
           # Append `--nightly` and `--continue-on-error` when the job belongs to the nightly pipeline.
diff --git a/python/pyproject.toml b/python/pyproject.toml
index 05ee34aa6..1ac9846e1 100755
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -186,7 +186,7 @@ test = [
   # Pinned: a bump moves scoring for every eval test 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 it.
-  "sgl-eval==0.1.0",
+  "sgl-eval==0.1.2",
   "sglang[fastokens]",
   "tabulate",
 ]
diff --git a/python/pyproject_cpu.toml b/python/pyproject_cpu.toml
index 7ac9a0e1c..f8fbf1fe1 100644
--- a/python/pyproject_cpu.toml
+++ b/python/pyproject_cpu.toml
@@ -110,7 +110,7 @@ test = [
   "pandas",
   "peft>=0.18.0",
   "sentence_transformers",
-  "sgl-eval==0.1.0",
+  "sgl-eval==0.1.2",
 ]
 all = []
 dev = ["sglang[test]"]
diff --git a/python/pyproject_npu.toml b/python/pyproject_npu.toml
index b48b4eb3a..5e925d3b1 100644
--- a/python/pyproject_npu.toml
+++ b/python/pyproject_npu.toml
@@ -106,7 +106,7 @@ test = [
   "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
-  "sgl-eval==0.1.0",
+  "sgl-eval==0.1.2",
   "tabulate",
 ]
 
diff --git a/python/pyproject_other.toml b/python/pyproject_other.toml
index e26da61da..178d6aa2a 100755
--- a/python/pyproject_other.toml
+++ b/python/pyproject_other.toml
@@ -208,7 +208,7 @@ test = [
   "peft>=0.18.0,<0.19.0", # Pin to <0.19.0 due to torchao incompatibility
   "pytest",
   "sentence_transformers",
-  "sgl-eval==0.1.0",
+  "sgl-eval==0.1.2",
   "tabulate",
 ]