[CI] Install sgl-eval from PyPI through the test extra (#37504)

python/pyproject.toml

@@ -172,11 +172,8 @@ test = [
   "lm-eval[api]>=0.4.9.2",
   "matplotlib",
   # 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.
+  # raises ImportError on 4.7.x); pinned here so other test deps cannot pull
+  # it down.
   "antlr4-python3-runtime==4.9.3",
   "pandas",
   "parameterized",
@@ -185,6 +182,10 @@ test = [
   "pytest",
   "pytest-cov",
   "sentence_transformers",
+  # 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",
   "sglang[fastokens]",
   "tabulate",
 ]

python/pyproject_cpu.toml

@@ -109,6 +109,7 @@ test = [
   "pandas",
   "peft>=0.18.0",
   "sentence_transformers",
+  "sgl-eval==0.1.0",
 ]
 all = []
 dev = ["sglang[test]"]

python/pyproject_npu.toml

@@ -106,6 +106,7 @@ test = [
   "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
+  "sgl-eval==0.1.0",
   "tabulate",
 ]
 

python/pyproject_other.toml

@@ -197,6 +197,7 @@ 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",
   "tabulate",
 ]
 

python/sglang/test/kits/eval_accuracy_kit.py

@@ -111,7 +111,7 @@ def _run_sgl_eval(
     asserts the score meets ``score_threshold``, and checks the speculative accept
     length. ``thinking=True`` sends per-request ``chat_template_kwargs={"thinking":
     True}`` so the server separates reasoning from the final answer. Skips the test
-    if sgl-eval (git-only) is not installed. Returns the RunResult.
+    if sgl-eval is not installed. Returns the RunResult.
     """
     assert (
         score_threshold == score_threshold
@@ -122,10 +122,7 @@ def _run_sgl_eval(
         from sgl_eval.sampler import ChatCompletionSampler
         from sgl_eval.types import GenConfig
     except ImportError:
-        test_case.skipTest(
-            "sgl-eval not installed; pip install "
-            "'sgl-eval @ git+https://github.com/sgl-project/sgl-eval'"
-        )
+        test_case.skipTest("sgl-eval not installed; pip install 'sglang[test]'")
 
     base_url = test_case.base_url.rstrip("/")
     if not base_url.endswith("/v1"):