chore: upgrade v0.3.9.post2 sgl-kernel (#10297)

.github/workflows/pr-test-pd-router.yml

@@ -119,7 +119,7 @@ jobs:
         python3 -m pip --no-cache-dir install -e "python[all]" --break-system-packages
         python3 -m pip --no-cache-dir install mooncake-transfer-engine==0.3.5
         python3 -m pip --no-cache-dir install --user --force-reinstall genai-bench==0.0.2
-        python3 -m pip --no-cache-dir install sgl-kernel==0.3.9
+        python3 -m pip --no-cache-dir install sgl-kernel==0.3.9.post2
 
     - name: Build and install sgl-router
       run: |

docker/Dockerfile

@@ -84,11 +84,8 @@ RUN python3 -m pip install --no-cache-dir --upgrade pip setuptools wheel html5li
  && python3 -m pip install --no-cache-dir -e "python[${BUILD_TYPE}]" --extra-index-url https://download.pytorch.org/whl/cu${CUINDEX} \
  && python3 -m pip install --no-cache-dir nvidia-nccl-cu12==2.27.6 --force-reinstall --no-deps \
  && python3 -m flashinfer --download-cubin \
- && if [ "$CUDA_VERSION" = "12.8.1" ]; then \
-      python3 -m pip install --no-cache-dir https://github.com/sgl-project/whl/releases/download/v0.3.8/sgl_kernel-0.3.8+cu128-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall --no-deps ; \
-    fi \
- && if [ "$CUDA_VERSION" = "12.9.1" ]; then \
-      python3 -m pip install --no-cache-dir https://github.com/sgl-project/whl/releases/download/v0.3.8/sgl_kernel-0.3.8+cu129-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall --no-deps ; \
+ && if [ "$CUDA_VERSION" = "12.6.1" ]; then \
+      python3 -m pip install --no-cache-dir https://github.com/sgl-project/whl/releases/download/v0.3.9.post2/sgl_kernel-0.3.9.post2+cu124-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall --no-deps ; \
     fi
 
 # Download source files

docker/Dockerfile.gb200

@@ -4,7 +4,7 @@ FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu22.04
 ARG BUILD_TYPE=blackwell
 ARG DEEPEP_COMMIT=1b14ad661c7640137fcfe93cccb2694ede1220b0
 ARG CMAKE_BUILD_PARALLEL_LEVEL=2
-ARG SGL_KERNEL_VERSION=0.3.8
+ARG SGL_KERNEL_VERSION=0.3.9.post2
 ENV DEBIAN_FRONTEND=noninteractive \
     CUDA_HOME=/usr/local/cuda \
     GDRCOPY_HOME=/usr/src/gdrdrv-2.4.4/ \

python/pyproject.toml

@@ -58,7 +58,7 @@ runtime_common = [
 
 srt = [
     "sglang[runtime_common]",
-    "sgl-kernel==0.3.8",
+    "sgl-kernel==0.3.9.post2",
     "torch==2.8.0",
     "torchaudio==2.8.0",
     "torchvision",
@@ -68,7 +68,7 @@ srt = [
 
 blackwell = [
     "sglang[runtime_common]",
-    "sgl-kernel",
+    "sgl-kernel==0.3.9.post2",
     "torch==2.8.0",
     "torchaudio==2.8.0",
     "torchvision",

python/sglang/srt/entrypoints/engine.py

@@ -682,7 +682,7 @@ def _set_envs_and_config(server_args: ServerArgs):
     if _is_cuda and not get_bool_env_var("SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK"):
         assert_pkg_version(
             "sgl-kernel",
-            "0.3.8",
+            "0.3.9.post2",
             "Please reinstall the latest version with `pip install sgl-kernel --force-reinstall`",
         )
 

scripts/ci/ci_install_dependency.sh

@@ -49,10 +49,12 @@ $PIP_CMD install -e "python[dev]" --extra-index-url https://download.pytorch.org
 SGLANG_ROUTER_BUILD_NO_RUST=1 $PIP_CMD install -e "sgl-router" $PIP_INSTALL_SUFFIX
 
 
+SGL_KERNEL_VERSION=0.3.9.post2
 if [ "$IS_BLACKWELL" = "1" ]; then
     # TODO auto determine sgl-kernel version
-    SGL_KERNEL_VERSION=0.3.8
     $PIP_CMD install https://github.com/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sgl_kernel-${SGL_KERNEL_VERSION}+cu128-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall $PIP_INSTALL_SUFFIX
+else
+    $PIP_CMD install https://github.com/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sgl_kernel-${SGL_KERNEL_VERSION}+cu124-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall $PIP_INSTALL_SUFFIX
 fi
 
 # Show current packages