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

.claude/skills/cookbook-add-model/templates/config.jsx.tmpl

@@ -104,7 +104,7 @@ export const config = {
     // in _deployment.jsx + the per-cell/defaultAccuracy keys.
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,

.github/workflows/_pr-test-stage-cpu.yml

@@ -119,8 +119,6 @@ jobs:
           UV_SYSTEM_PYTHON: "1"
         run: |
           uv pip install -e "python[dev]" --index-strategy unsafe-best-match --prerelease allow
-          source scripts/ci/utils/sgl_eval_ref.sh
-          uv pip install "$SGL_EVAL_SPEC" --index-strategy unsafe-best-match
 
       # Hosted runners are ephemeral, so models are re-fetched every run and the
       # Hub occasionally returns 429s. Persist the HF cache in GitHub's cache

docker/xeon.Dockerfile

@@ -43,8 +43,7 @@ RUN source /opt/.venv/bin/activate && \
     cd sglang/kernels/aot && \
     cp pyproject_cpu.toml pyproject.toml && \
     uv pip install . && \
-    source /sgl-workspace/sglang/scripts/ci/utils/sgl_eval_ref.sh && \
+    uv pip install "sgl-eval==0.1.0"
-    uv pip install "$SGL_EVAL_SPEC"
 
 ENV SGLANG_USE_CPU_ENGINE=1
 ENV LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc.so.4:/usr/lib/x86_64-linux-gnu/libtbbmalloc.so:/opt/.venv/lib/libiomp5.so

docs/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2.mdx

@@ -849,7 +849,7 @@ Results on AIME 2025 (8×B200), evaluated with [sgl-eval](https://github.com/sgl
 **Reproduction.** Install [sgl-eval](https://github.com/sgl-project/sgl-eval), launch the server with the tool-call and reasoning parsers, then run `sgl-eval run`. The AIME 2025 dataset ships with sgl-eval, and thinking is on by default for it:
 
 ```bash Command
-pip install git+https://github.com/sgl-project/sgl-eval.git
+pip install sgl-eval
 
 sgl-eval run aime25 \
   --base-url http://localhost:30000/v1 \

docs/cookbook/autoregressive/Google/Gemma4.mdx

@@ -1455,7 +1455,7 @@ Median ITL (ms):                         15.11
 Reproduce against a running server (`--base-url` points at your endpoint):
 
 ```bash Command
-pip install git+https://github.com/sgl-project/sgl-eval
+pip install sgl-eval
 
 # Sanity-check the endpoint
 sgl-eval ping --base-url http://localhost:30000/v1

docs/cookbook/autoregressive/MiniMax/MiniMax-M2.7.mdx

@@ -497,7 +497,7 @@ This section uses **industry-standard configurations** for comparable benchmark
 - Prompt: `eval/aai/mcq-4choices` (4-choice multiple choice, matching [Artificial Analysis methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking)) -- sgl-eval's default for this benchmark
 - Evaluation command:
 ```bash Command
-pip install git+https://github.com/sgl-project/sgl-eval.git
+pip install sgl-eval
 
 sgl-eval run gpqa \
     --base-url http://localhost:30000/v1 \

docs/cookbook/autoregressive/Moonshotai/Kimi-K2.6.mdx

@@ -628,7 +628,7 @@ python tool_calls_eval.py tool-calls/samples.jsonl \
 **Evaluation Command:**
 
 ```shell Command
-pip install git+https://github.com/sgl-project/sgl-eval.git
+pip install sgl-eval
 
 # The dataset ships with sgl-eval; --n-repeats runs all 32 samples per problem.
 sgl-eval run aime25 \

docs/src/snippets/configs/LiquidAI/lfm2.5.jsx

@@ -71,26 +71,26 @@ export const config = {
   --num-prompts {{NUM_PROMPTS}} --max-concurrency {{MAX_CONCURRENCY}}`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\
   --num-threads 128`,
       gpqa_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 # GPQA's HF dataset (Idavidrein/gpqa) is gated — accept its terms with your HF account first.
 sgl-eval run gpqa \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\
   --num-threads 128`,
       mmlu_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run mmlu \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\
   --num-threads 128`,
       aime25_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime25 \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\

docs/src/snippets/configs/MiniMaxAI/minimax-m3.jsx

@@ -61,21 +61,21 @@ export const config = {
   --num-prompts {{NUM_PROMPTS}} --max-concurrency {{MAX_CONCURRENCY}}`,
     accuracy: {
       gsm8k_pct:
-`pip install git+https://github.com/sgl-project/sgl-eval
+`pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\
   --temperature 1.0 --top-p 0.95 \\
   --thinking`,
       gpqa_pct:
-`pip install git+https://github.com/sgl-project/sgl-eval
+`pip install sgl-eval
 sgl-eval run gpqa \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\
   --temperature 1.0 --top-p 0.95 \\
   --thinking --n-repeats 4 --max-tokens 40960`,
       mmmu_pro_pct:
-`pip install git+https://github.com/sgl-project/sgl-eval
+`pip install sgl-eval
 sgl-eval run mmmu_pro \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\

docs/src/snippets/configs/deepseek-ai/deepseek-v4.jsx

@@ -98,13 +98,13 @@ export const config = {
   --warmup-requests 64 --flush-cache`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,
       gpqa_pct: {
         flash:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gpqa \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 200000 \\
@@ -112,7 +112,7 @@ sgl-eval run gpqa \\
   --out-dir /sgl-workspace/logs \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1`,
         "flash-official":
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gpqa \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 200000 \\
@@ -120,7 +120,7 @@ sgl-eval run gpqa \\
   --out-dir /sgl-workspace/logs \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1`,
         pro:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gpqa \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 400000 \\
@@ -130,7 +130,7 @@ sgl-eval run gpqa \\
       },
       aime25_pct: {
         "flash-official":
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime25 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 200000 \\
@@ -138,7 +138,7 @@ sgl-eval run aime25 \\
   --out-dir /sgl-workspace/logs \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1`,
         flash:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime25 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 200000 \\
@@ -146,7 +146,7 @@ sgl-eval run aime25 \\
   --out-dir /sgl-workspace/logs \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1`,
         pro:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime25 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 400000 \\
@@ -156,7 +156,7 @@ sgl-eval run aime25 \\
       },
       mmmu_pro_pct: {
         "flash-vision":
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run mmmu_pro \\
   --reasoning-effort max \\
   --temperature 1.0 --top-p 0.95 \\

docs/src/snippets/configs/inclusionAI/ling-3.0-flash.jsx

@@ -63,7 +63,7 @@ export const config = {
   --num-prompts {{NUM_PROMPTS}} --max-concurrency {{MAX_CONCURRENCY}} \\
   --flush-cache`,
     accuracy: {
-      gsm8k_pct: `# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+      gsm8k_pct: `# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,

docs/src/snippets/configs/inclusionAI/ling-3.0-tiny.jsx

@@ -57,7 +57,7 @@ export const config = {
   --num-prompts {{NUM_PROMPTS}} --max-concurrency {{MAX_CONCURRENCY}} \\
   --flush-cache`,
     accuracy: {
-      gsm8k_pct: `# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+      gsm8k_pct: `# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32 \\

docs/src/snippets/configs/internlm/intern-s2-mobius.jsx

@@ -51,12 +51,12 @@ export const config = {
   --warmup-requests 8 --flush-cache`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,
       gpqa_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gpqa \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 16`,

docs/src/snippets/configs/meituan-longcat/longcat-2.0.jsx

@@ -57,7 +57,7 @@ export const config = {
   --warmup-requests 64 --flush-cache`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,

docs/src/snippets/configs/nvidia/nemotron-3.5-lightning.jsx

@@ -51,7 +51,7 @@ export const config = {
   --flush-cache`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,

docs/src/snippets/configs/poolside/laguna-m1.jsx

@@ -82,7 +82,7 @@ export const config = {
     // (AIME 25 to be added back once truncation-free numbers are measured.)
     accuracy: {
       gsm8k_pct:
-`# pip install git+https://github.com/sgl-project/sgl-eval
+`# pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 128`,

docs/src/snippets/configs/poolside/laguna-s21.jsx

@@ -72,7 +72,7 @@ export const config = {
   --num-prompts {{NUM_PROMPTS}} --max-concurrency {{MAX_CONCURRENCY}}`,
     accuracy: {
       gsm8k_pct:
-`# pip install git+https://github.com/sgl-project/sgl-eval
+`# pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 128`,
@@ -80,7 +80,7 @@ sgl-eval run gsm8k \\
       // Serve with a copy of the model's chat template whose enable_thinking default
       // is flipped to true. For BF16: use --max-tokens 131072 (see Configuration Tips).
       aime25_pct:
-`# pip install git+https://github.com/sgl-project/sgl-eval
+`# pip install sgl-eval
 # Serve with an enable_thinking=true chat template (see Configuration Tips: Thinking).
 # For BF16: use --max-tokens 131072 (see Configuration Tips: BF16 reasoning length).
 sgl-eval run aime25 \\

docs/src/snippets/configs/poolside/laguna-xs21.jsx

@@ -112,7 +112,7 @@ export const config = {
     // GSM8K is the required accuracy sanity on every verified cell (cookbook_guide §3), via sgl-eval.
     accuracy: {
       gsm8k_pct:
-`# pip install git+https://github.com/sgl-project/sgl-eval
+`# pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 128`,
@@ -121,7 +121,7 @@ sgl-eval run gsm8k \\
       // AIME25 with thinking, serve with a copy of the model's chat template whose
       // enable_thinking default is flipped to true, passed via --chat-template.
       aime25_pct:
-`# pip install git+https://github.com/sgl-project/sgl-eval
+`# pip install sgl-eval
 # Serve with an enable_thinking=true chat template (see Configuration Tips: Thinking).
 sgl-eval run aime25 \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\

docs/src/snippets/configs/tencent/hy3.jsx

@@ -62,12 +62,12 @@ export const config = {
   --warmup-requests 64`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,
       aime26_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime26 \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\

docs/src/snippets/configs/tencent/hy4-preview.jsx

@@ -72,7 +72,7 @@ export const config = {
     // GSM8K harness — keep these exact settings for comparability across runs.
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\

docs/src/snippets/configs/zai-org/glm-5.2.jsx

@@ -63,12 +63,12 @@ export const config = {
   --warmup-requests 64 --flush-cache`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,
       aime25_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime25 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 64000 \\

docs/src/snippets/configs/zai-org/glm-5.3-flash.jsx

@@ -158,7 +158,7 @@ python3 -m sglang.bench_serving \\
     numPromptsByConc: { 1: 16, 16: 80, 64: 320, 256: 1280, 1024: 5120 },
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --model {{MODEL_NAME}} \\

docs/src/snippets/configs/zai-org/glm-5.3.jsx

@@ -61,12 +61,12 @@ export const config = {
   --warmup-requests 64 --flush-cache`,
     accuracy: {
       gsm8k_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run gsm8k \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1 \\
   --num-threads 32`,
       aime26_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime26 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 64000 \\
@@ -74,7 +74,7 @@ sgl-eval run aime26 \\
   --out-dir /sgl-workspace/logs \\
   --base-url http://{{CURL_HOST}}:{{CURL_PORT}}/v1`,
       aime25_pct:
-`# To install sgl-eval: pip install git+https://github.com/sgl-project/sgl-eval
+`# To install sgl-eval: pip install sgl-eval
 sgl-eval run aime25 \\
   --model {{MODEL_NAME}} --api-key <api-key> \\
   --n-repeats 16 --max-tokens 64000 \\

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
+  # raises ImportError on 4.7.x); pinned here so other test deps cannot pull
-  # compatible when sgl-eval is installed on top of it.
+  # it down.
-  # 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",
   "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(
+        test_case.skipTest("sgl-eval not installed; pip install 'sglang[test]'")
-            "sgl-eval not installed; pip install "
-            "'sgl-eval @ git+https://github.com/sgl-project/sgl-eval'"
-        )
 
     base_url = test_case.base_url.rstrip("/")
     if not base_url.endswith("/v1"):

scripts/ci/amd/amd_ci_install_dependency.sh

@@ -187,10 +187,6 @@ 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

@@ -726,8 +726,6 @@ stabilize_flashinfer_jit_paths() {
 install_extra_deps() {
     MOONCAKE_VERSION="0.3.13"
     NIXL_VERSION="1.3.0"
-    # 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"
@@ -763,8 +761,6 @@ install_extra_deps() {
             --no-deps --force-reinstall $PIP_INSTALL_SUFFIX
     fi
 
-    $PIP_CMD install "$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
         $PIP_CMD install -e lmms-eval/ $PIP_INSTALL_SUFFIX

scripts/ci/npu/npu_ci_install_dependency.sh

@@ -79,8 +79,3 @@ 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

@@ -1,13 +0,0 @@
-# Single source of truth for the sgl-eval commit every CI variant installs.
-# Meant to be sourced, not executed -- each variant then installs
-# "$SGL_EVAL_SPEC" with its own pip invocation, since those differ (uv pip on
-# CUDA/CPU, `docker exec ... pip` on AMD, `python3 -m pip` on NPU).
-#
-# sgl-eval is git-only and cannot be declared in python/pyproject.toml (see the
-# note there). Every eval that shells out to the `sgl-eval` CLI fails without
-# it, and a bump moves scoring for all of them at once -- so re-baseline
-# MODEL_SCORE_THRESHOLDS in
-# test/registered/eval/test_text_models_gsm8k_eval.py, and the mmlu thresholds
-# of run_eval's other callers, before changing this.
-SGL_EVAL_REF="a231b7a439b235090ff7baa30778fa2b514309ae"
-SGL_EVAL_SPEC="sgl-eval@git+https://github.com/sgl-project/sgl-eval.git@${SGL_EVAL_REF}"