use self-hosted to build sgl-kernel (#3154)

sgl-kernel/build.sh

@@ -15,6 +15,7 @@ docker run --rm \
     pytorch/manylinux-builder:cuda${CUDA_VERSION} \
     bash -c "
     ${PYTHON_ROOT_PATH}/bin/pip install --no-cache-dir torch==2.5.1 --index-url https://download.pytorch.org/whl/cu${CUDA_VERSION//.} && \
+    ${PYTHON_ROOT_PATH}/bin/pip install --no-cache-dir ninja && \
     export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0+PTX' && \
     export CUDA_VERSION=${CUDA_VERSION} && \
     export SGL_KERNEL_ENABLE_BF16=1 && \

sgl-kernel/setup.py

@@ -1,3 +1,4 @@
+import multiprocessing
 import os
 from pathlib import Path
 
@@ -70,6 +71,8 @@ nvcc_flags = [
     "-std=c++17",
     "-use_fast_math",
     "-DFLASHINFER_ENABLE_F16",
+    "-Xcompiler",
+    "-w",
 ]
 nvcc_flags_fp8 = [
     "-DFLASHINFER_ENABLE_FP8",
@@ -151,7 +154,11 @@ setup(
     packages=find_packages(),
     package_dir={"": "src"},
     ext_modules=ext_modules,
-    cmdclass={"build_ext": BuildExtension},
+    cmdclass={
+        "build_ext": BuildExtension.with_options(
+            use_ninja=True, max_jobs=multiprocessing.cpu_count()
+        )
+    },
     options={"bdist_wheel": {"py_limited_api": "cp39"}},
 )