A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 19:00 UTC
[Bugfix] Install nvidia-cutlass-dsl[cu13] extra on CUDA 13 platforms (#42438)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
May 13, 2026 at 16:57 UTC
140dc2ec30bad7e3745e070c50b559331a7eb703
4 files changed
+14
-2
.github/workflows/scripts/build.sh
+3
@@ -9,6 +9,9 @@ PATH=${cuda_home}/bin:$PATH
9
LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
10
11
# Install requirements
12
+if [ "$(echo $2 | cut -d. -f1)" = "12" ]; then
13
+ sed -i 's/^nvidia-cutlass-dsl\[cu13\]>=/nvidia-cutlass-dsl>=/' requirements/cuda.txt
14
+fi
15
$python_executable -m pip install -r requirements/build/cuda.txt -r requirements/cuda.txt
16
17
# Limit the number of parallel jobs to avoid OOM
docker/Dockerfile
+7
-1
@@ -199,7 +199,10 @@ COPY requirements/cuda.txt requirements/cuda.txt
199
COPY use_existing_torch.py use_existing_torch.py
200
COPY pyproject.toml pyproject.toml
201
RUN --mount=type=cache,target=/root/.cache/uv \
202
- if [ "${PYTORCH_NIGHTLY}" = "1" ]; then \
202
+ if [ "$(echo $CUDA_VERSION | cut -d. -f1)" = "12" ]; then \
203
+ sed -i 's/^nvidia-cutlass-dsl\[cu13\]>=/nvidia-cutlass-dsl>=/' requirements/cuda.txt; \
204
+ fi \
205
+ && if [ "${PYTORCH_NIGHTLY}" = "1" ]; then \
206
echo "Installing torch nightly..." \
207
&& uv pip install --python /opt/venv/bin/python3 torch torchaudio torchvision --pre \
208
--index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
@@ -626,6 +629,9 @@ ARG PYTORCH_CUDA_INDEX_BASE_URL
629
COPY requirements/common.txt /tmp/common.txt
630
COPY requirements/cuda.txt /tmp/requirements-cuda.txt
631
RUN --mount=type=cache,target=/root/.cache/uv \
632
+ if [ "$(echo $CUDA_VERSION | cut -d. -f1)" = "12" ]; then \
633
+ sed -i 's/^nvidia-cutlass-dsl\[cu13\]>=/nvidia-cutlass-dsl>=/' /tmp/requirements-cuda.txt; \
634
+ fi && \
635
uv pip install --system -r /tmp/requirements-cuda.txt \
636
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') && \
637
rm /tmp/requirements-cuda.txt /tmp/common.txt
requirements/cuda.txt
+1
-1
@@ -21,5 +21,5 @@ nvidia-cudnn-frontend>=1.13.0,<1.19.0
21
fastsafetensors >= 0.2.2
22
23
# QuACK and Cutlass DSL for FA4 (cute-DSL implementation)
24
-nvidia-cutlass-dsl>=4.4.2
24
+nvidia-cutlass-dsl[cu13]>=4.4.2
25
quack-kernels>=0.3.3
setup.py
+3
@@ -970,6 +970,9 @@ def get_requirements() -> list[str]:
970
# vllm-flash-attn is built only for CUDA 12.x.
971
# Skip for other versions.
972
continue
973
+ if "nvidia-cutlass-dsl[cu13]" in req and cuda_major == "12":
974
+ # [cu13] extra is the default; strip it on CUDA 12 builds.
975
+ req = req.replace("nvidia-cutlass-dsl[cu13]", "nvidia-cutlass-dsl")
976
modified_requirements.append(req)
977
requirements = modified_requirements
978
elif _is_hip():