minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[Fix] Try to fix error caused by latest cutedsl packages (#25690)

Browse Source 
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
This commit is contained in:
Baizhou Zhang
2026-05-18 16:51:32 -07:00
committed by GitHub
co-authored by hnyls2002
parent 314dedf7c6
commit b79e4b1e68
2 changed files with 21 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/ci/cuda/ci_install_dependency.sh
+19 -2
View File
@@ -186,8 +186,8 @@ uninstall_stale_flashinfer() {
# Keep flashinfer packages if version matches to avoid re-downloading:
# - flashinfer-cubin: 150+ MB
# - flashinfer-jit-cache: 1.2+ GB
FLASHINFER_PYTHON_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_python==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
FLASHINFER_CUBIN_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_cubin==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
FLASHINFER_PYTHON_REQUIRED=$(grep -Po -m1 'flashinfer_python(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
FLASHINFER_CUBIN_REQUIRED=$(grep -Po -m1 'flashinfer_cubin(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
FLASHINFER_CUBIN_INSTALLED=$(pip show flashinfer-cubin 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
FLASHINFER_JIT_INSTALLED=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//' || echo "")
FLASHINFER_JIT_CU_VERSION=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed -n 's/.*+//p' || echo "")
@@ -335,6 +335,22 @@ download_flashinfer_cache() {
mark_step_done "${FUNCNAME[0]}"
}
purge_cutlass_libs_base() {
# nvidia-cutlass-dsl[cu13] extras are additive on PyPI: requires_dist always
# pulls -libs-base AND -libs-cu13 when [cu13] is requested. Both wheels write
# to the same site-packages paths with different content, leaving the wrapper
# (cutlass.py, cu13 style) mismatched with the binding (_gpu_ops_gen.py, base
# style) -> GPUModuleOp signature TypeError. See vllm-project/vllm#40082.
# Uninstall -libs-base, then force-reinstall -libs-cu13 so its files win.
$PIP_UNINSTALL_CMD nvidia-cutlass-dsl-libs-base $PIP_UNINSTALL_SUFFIX || true
CUTLASS_DSL_VERSION=$(grep -Po -m1 'nvidia-cutlass-dsl(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
if [ -n "$CUTLASS_DSL_VERSION" ]; then
$PIP_CMD install --force-reinstall --no-deps "nvidia-cutlass-dsl-libs-cu13==${CUTLASS_DSL_VERSION}" $PIP_INSTALL_SUFFIX
fi
mark_step_done "${FUNCNAME[0]}"
}
stabilize_flashinfer_jit_paths() {
# In venv mode, FlashInfer JIT writes build.ninja with hardcoded -isystem
# paths. Per-job venvs get unique paths, but the JIT cache is shared on the
@@ -488,6 +504,7 @@ main() {
install_sglang_kernel
install_sglang_router
download_flashinfer_cache
purge_cutlass_libs_base
stabilize_flashinfer_jit_paths
install_extra_deps
install_test_tools