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 14, 22:35 UTC
[CI/Build] get rid of unused VLLM_FA_CMAKE_GPU_ARCHES (#21599)
Signed-off-by: Daniele Trifirò <dtrifiro@redhat.com>
Daniele committed
Jul 31, 2025 at 09:00 UTC
d2aab336ad7822efe7cfc345fa3ad67d6f5cbe39
5 files changed
+2
-11
.buildkite/scripts/hardware_ci/run-gh200-test.sh
+1
-2
@@ -16,8 +16,7 @@ DOCKER_BUILDKIT=1 docker build . \
16
--build-arg max_jobs=66 \
17
--build-arg nvcc_threads=2 \
18
--build-arg RUN_WHEEL_CHECK=false \
19
- --build-arg torch_cuda_arch_list="9.0+PTX" \
20
- --build-arg vllm_fa_cmake_gpu_arches="90-real"
19
+ --build-arg torch_cuda_arch_list="9.0+PTX"
20
21
# Setup cleanup
22
remove_docker_container() { docker rm -f gh200-test || true; }
.github/workflows/scripts/build.sh
-1
@@ -15,7 +15,6 @@ $python_executable -m pip install -r requirements/build.txt -r requirements/cuda
15
export MAX_JOBS=1
16
# Make sure release wheels are built for the following architectures
17
export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
18
-export VLLM_FA_CMAKE_GPU_ARCHES="80-real;90-real"
18
19
bash tools/check_repo.sh
20
docker/Dockerfile
-3
@@ -164,9 +164,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
164
# see https://github.com/pytorch/pytorch/pull/123243
165
ARG torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0 10.0 12.0'
166
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
167
-# Override the arch list for flash-attn to reduce the binary size
168
-ARG vllm_fa_cmake_gpu_arches='80-real;90-real'
169
-ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
167
#################### BASE BUILD IMAGE ####################
168
169
#################### WHEEL BUILD IMAGE ####################
docker/Dockerfile.nightly_torch
-3
@@ -114,9 +114,6 @@ RUN cat torch_build_versions.txt
114
# explicitly set the list to avoid issues with torch 2.2
115
# see https://github.com/pytorch/pytorch/pull/123243
116
117
-# Override the arch list for flash-attn to reduce the binary size
118
-ARG vllm_fa_cmake_gpu_arches='80-real;90-real'
119
-ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
117
#################### BASE BUILD IMAGE ####################
118
119
#################### WHEEL BUILD IMAGE ####################
docs/deployment/docker.md
+1
-2
@@ -106,8 +106,7 @@ of PyTorch Nightly and should be considered **experimental**. Using the flag `--
106
-t vllm/vllm-gh200-openai:latest \
107
--build-arg max_jobs=66 \
108
--build-arg nvcc_threads=2 \
109
- --build-arg torch_cuda_arch_list="9.0 10.0+PTX" \
110
- --build-arg vllm_fa_cmake_gpu_arches="90-real"
109
+ --build-arg torch_cuda_arch_list="9.0 10.0+PTX"
110
```
111
112
!!! note