[Build] Remove Ubuntu build-stage option from the CUDA dockerfile (#51060)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
Michael Goin committed
Aug 5, 2026 at 13:50 UTC
e8586b28788ee9ef2286bca9d6e98724041f185b
8 files changed
+42
-97
.buildkite/image_build/image_build_arm64.sh
+1
@@ -26,6 +26,7 @@ else
26
--platform linux/arm64 \
27
--build-arg max_jobs=16 \
28
--build-arg nvcc_threads=4 \
29
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 \
30
--build-arg torch_cuda_arch_list="9.0 10.0 11.0 12.0" \
31
--build-arg USE_SCCACHE=1 \
32
--build-arg buildkite_commit="$BUILDKITE_COMMIT" \
.buildkite/image_build/image_build_torch_nightly.sh
+1
-1
@@ -48,7 +48,7 @@ echo "Image not found, proceeding with build..."
48
# --- CUDA 13.0 for nightly builds ---
49
# Nightly CI uses CUDA 13.0 while regular CI stays on CUDA 12.9
50
NIGHTLY_CUDA_VERSION="13.0.2"
51
-NIGHTLY_BUILD_BASE_IMAGE="nvidia/cuda:${NIGHTLY_CUDA_VERSION}-devel-ubuntu22.04"
51
+NIGHTLY_BUILD_BASE_IMAGE="pytorch/manylinux2_28-builder:cuda13.0"
52
NIGHTLY_FINAL_BASE_IMAGE="nvidia/cuda:${NIGHTLY_CUDA_VERSION}-base-ubuntu22.04"
53
54
echo "--- :docker: Building torch nightly image (CUDA ${NIGHTLY_CUDA_VERSION})"
.buildkite/release-pipeline.yaml
+12
-8
@@ -40,7 +40,7 @@ steps:
40
agents:
41
queue: arm64_cpu_queue_release
42
commands:
43
- - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
43
+ - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
44
- "mkdir artifacts"
45
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
46
- "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -54,7 +54,7 @@ steps:
54
agents:
55
queue: cpu_queue_release
56
commands:
57
- - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
57
+ - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
58
- "mkdir artifacts"
59
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
60
- "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -78,7 +78,7 @@ steps:
78
agents:
79
queue: arm64_cpu_queue_release
80
commands:
81
- - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda12.9 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
81
+ - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda12.9 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
82
- "mkdir artifacts"
83
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
84
- "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -149,7 +149,7 @@ steps:
149
agents:
150
queue: cpu_queue_release
151
commands:
152
- - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda12.9 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
152
+ - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda12.9 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
153
- "mkdir artifacts"
154
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
155
- "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -221,7 +221,7 @@ steps:
221
--build-arg INSTALL_KV_CONNECTORS=true \
222
--build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
223
--build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
224
- --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
224
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.0 \
225
--target vllm-openai \
226
--progress plain \
227
-f docker/Dockerfile .
@@ -249,7 +249,7 @@ steps:
249
--build-arg INSTALL_KV_CONNECTORS=true \
250
--build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
251
--build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
252
- --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
252
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 \
253
--target vllm-openai \
254
--progress plain \
255
-f docker/Dockerfile .
@@ -270,6 +270,7 @@ steps:
270
--build-arg USE_SCCACHE=1 \
271
--build-arg GIT_REPO_CHECK=1 \
272
--build-arg CUDA_VERSION=12.9.1 \
273
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda12.9 \
274
--build-arg torch_cuda_arch_list="${CUDA_ARCH_X86_CU129}" \
275
--build-arg INSTALL_KV_CONNECTORS=true \
276
--build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
@@ -297,6 +298,7 @@ steps:
298
--build-arg USE_SCCACHE=1 \
299
--build-arg GIT_REPO_CHECK=1 \
300
--build-arg CUDA_VERSION=12.9.1 \
301
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda12.9 \
302
--build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64_CU129}" \
303
--build-arg INSTALL_KV_CONNECTORS=true \
304
--build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
@@ -327,7 +329,7 @@ steps:
329
--build-arg INSTALL_KV_CONNECTORS=true \
330
--build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_39}" \
331
--build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
330
- --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 \
332
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.0 \
333
--target vllm-openai \
334
--progress plain \
335
-f docker/Dockerfile .
@@ -356,7 +358,7 @@ steps:
358
--build-arg INSTALL_KV_CONNECTORS=true \
359
--build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_39}" \
360
--build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
359
- --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 \
361
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 \
362
--target vllm-openai \
363
--progress plain \
364
-f docker/Dockerfile .
@@ -377,6 +379,7 @@ steps:
379
--build-arg USE_SCCACHE=1 \
380
--build-arg GIT_REPO_CHECK=1 \
381
--build-arg CUDA_VERSION=12.9.1 \
382
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda12.9 \
383
--build-arg UBUNTU_VERSION=24.04 \
384
--build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
385
--build-arg torch_cuda_arch_list="${CUDA_ARCH_X86_CU129}" \
@@ -405,6 +408,7 @@ steps:
408
--build-arg USE_SCCACHE=1 \
409
--build-arg GIT_REPO_CHECK=1 \
410
--build-arg CUDA_VERSION=12.9.1 \
411
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda12.9 \
412
--build-arg UBUNTU_VERSION=24.04 \
413
--build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
414
--build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64_CU129}" \
.buildkite/scripts/hardware_ci/run-gh200-test.sh
+1
@@ -15,6 +15,7 @@ DOCKER_BUILDKIT=1 docker build . \
15
-t gh200-test \
16
--build-arg max_jobs=66 \
17
--build-arg nvcc_threads=2 \
18
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 \
19
--build-arg torch_cuda_arch_list="9.0+PTX"
20
21
# Setup cleanup
docker/Dockerfile
+24
-83
@@ -35,23 +35,13 @@ ARG NCCL_VERSION=2.30.7
35
# private registries that use a different repository naming conventions.
36
#
37
# Example:
38
-# docker build --build-arg BUILD_BASE_IMAGE=registry.acme.org/mirror/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
38
+# docker build --build-arg BUILD_BASE_IMAGE=registry.acme.org/mirror/pytorch/manylinux2_28-builder:cuda13.0
39
40
-# Important: We build with an old version of Ubuntu to maintain broad
41
-# compatibility with other Linux OSes. The main reason for this is that the
42
-# glibc version is baked into the distro, and binaries built with one glibc
43
-# version are not backwards compatible with OSes that use an earlier version.
44
-ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
40
+# Build wheels against the same glibc floor as PyTorch's published wheels.
41
+ARG BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.0
42
# Using cuda base image with minimal dependencies necessary for JIT compilation (FlashInfer, DeepGEMM, EP kernels)
43
ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu${UBUNTU_VERSION}
44
48
-# OS family of BUILD_BASE_IMAGE. Controls package manager (apt vs dnf) and
49
-# Python bootstrap. Set to "manylinux" alongside a manylinux build base such
50
-# as pytorch/manylinux2_28-builder:cuda13.0 to produce wheels with a glibc
51
-# 2.28 floor (matches PyTorch's own published wheels). Default stays on
52
-# Ubuntu for backwards compatibility.
53
-ARG BUILD_OS=ubuntu
54
-
45
# By parameterizing the Deadsnakes repository URL, we allow third-party to use
46
# their own mirror. When doing so, we don't benefit from the transparent
47
# installation of the GPG key of the PPA, as done by add-apt-repository, so we
@@ -106,7 +96,6 @@ FROM ${BUILD_BASE_IMAGE} AS base
96
ARG TARGETPLATFORM
97
ARG CUDA_VERSION
98
ARG PYTHON_VERSION
109
-ARG BUILD_OS
99
ARG USE_SCCACHE
100
ARG SCCACHE_DOWNLOAD_URL
101
ARG SCCACHE_ENDPOINT
@@ -114,8 +103,6 @@ ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
103
ARG SCCACHE_REGION_NAME=us-west-2
104
ARG SCCACHE_S3_NO_CREDENTIALS=0
105
117
-ENV DEBIAN_FRONTEND=noninteractive
118
-
106
# Environment for uv
107
# Declared BEFORE the installer + `uv venv` invocations below so the uv
108
# binary, managed Python, download cache, and /opt/venv all land under
@@ -133,45 +120,16 @@ ENV UV_INSTALL_DIR=/opt/uv/bin
120
ENV PATH="/opt/venv/bin:/opt/uv/bin:$PATH"
121
ENV VIRTUAL_ENV="/opt/venv"
122
136
-# Install system dependencies including build tools.
137
-# The Ubuntu path uses apt + deadsnakes-via-uv for Python; the manylinux path
138
-# (AlmaLinux 8, e.g. pytorch/manylinux2_28-builder) uses dnf and the Python
139
-# interpreters pre-installed at /opt/python/cpXY-cpXY/.
140
-RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
141
- # rdma-core-devel provides libibverbs headers; ccache lives in EPEL,
142
- # which the pytorch manylinux image already enables. git/curl/sudo
143
- # are typically pre-installed but listed defensively.
144
- dnf install -y --setopt=install_weak_deps=False \
145
- ccache \
146
- git \
147
- curl \
148
- sudo \
149
- rdma-core-devel \
150
- && dnf clean all \
151
- && rm -rf /var/cache/dnf; \
152
- else \
153
- apt-get update -y \
154
- && apt-get install -y --no-install-recommends \
155
- ccache \
156
- software-properties-common \
157
- git \
158
- curl \
159
- sudo \
160
- python3-pip \
161
- libibverbs-dev \
162
- # GCC 10 was previously pinned to suppress spurious -Wredundant-move warnings
163
- # from CUTLASS (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519). That bug
164
- # was fixed in GCC 11. GCC >= 11.3 is now required because PyTorch's C++20 headers
165
- # (pytorch/pytorch#167929) are not compatible with GCC < 11.3.
166
- gcc-11 \
167
- g++-11 \
168
- && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 110 --slave /usr/bin/g++ g++ /usr/bin/g++-11 \
169
- # Install python dev headers if available (needed for cmake FindPython on Ubuntu 24.04
170
- # which ships cmake 3.28 and requires Development.SABIModule; silently skipped on
171
- # Ubuntu 20.04/22.04 where python3.x-dev is not available without a PPA)
172
- && (apt-get install -y --no-install-recommends python${PYTHON_VERSION}-dev 2>/dev/null || true) \
173
- && rm -rf /var/lib/apt/lists/*; \
174
- fi
123
+# Install system dependencies including build tools. The PyTorch manylinux
124
+# image enables EPEL, which provides ccache.
125
+RUN dnf install -y --setopt=install_weak_deps=False \
126
+ ccache \
127
+ git \
128
+ curl \
129
+ sudo \
130
+ rdma-core-devel \
131
+ && dnf clean all \
132
+ && rm -rf /var/cache/dnf
133
134
# Install sccache once in base so Rust and CMake/CUDA build stages share the
135
# same binary and remote cache configuration.
@@ -194,8 +152,8 @@ ENV SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION_NAME}}
152
ENV SCCACHE_S3_NO_CREDENTIALS=${USE_SCCACHE:+${SCCACHE_S3_NO_CREDENTIALS}}
153
ENV SCCACHE_IDLE_TIMEOUT=${USE_SCCACHE:+0}
154
197
-# Install uv and bootstrap /opt/venv. Both paths converge on /opt/venv so all
198
-# downstream stages stay distro-agnostic.
155
+# Install uv and bootstrap /opt/venv from the Python interpreter included in
156
+# the manylinux image.
157
RUN mkdir -p "${UV_PYTHON_INSTALL_DIR}" "${UV_CACHE_DIR}" "${UV_INSTALL_DIR}" \
158
&& chmod -R a+rX /opt/uv \
159
&& curl -LsSf https://astral.sh/uv/install.sh | sh \
@@ -203,15 +161,9 @@ RUN mkdir -p "${UV_PYTHON_INSTALL_DIR}" "${UV_CACHE_DIR}" "${UV_INSTALL_DIR}" \
161
# pip` works regardless of how uv happens to link the venv back to the
162
# managed Python install (which, at a non-default UV_PYTHON_INSTALL_DIR,
163
# doesn't always expose ensurepip via the default venv layout).
206
- && if [ "${BUILD_OS}" = "manylinux" ]; then \
207
- # manylinux images ship Python at /opt/python/cpXY-cpXY/; point uv
208
- # at the matching interpreter rather than letting it fetch one.
209
- PYV_NODOT=$(echo ${PYTHON_VERSION} | tr -d '.') \
210
- && MANYLINUX_PY=/opt/python/cp${PYV_NODOT}-cp${PYV_NODOT}/bin/python${PYTHON_VERSION} \
211
- && uv venv --seed /opt/venv --python "$MANYLINUX_PY"; \
212
- else \
213
- uv venv --seed /opt/venv --python ${PYTHON_VERSION}; \
214
- fi \
164
+ && PYV_NODOT=$(echo ${PYTHON_VERSION} | tr -d '.') \
165
+ && MANYLINUX_PY=/opt/python/cp${PYV_NODOT}-cp${PYV_NODOT}/bin/python${PYTHON_VERSION} \
166
+ && uv venv --seed /opt/venv --python "$MANYLINUX_PY" \
167
&& rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
168
&& ln -sf /opt/venv/bin/python3 /usr/bin/python3 \
169
&& ln -sf /opt/venv/bin/python3-config /usr/bin/python3-config \
@@ -295,21 +247,13 @@ ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
247
# This stage reuses the Python environment from base and runs in parallel with
248
# csrc-build/extensions-build.
249
FROM base AS rust-build
298
-ARG BUILD_OS
250
ARG USE_SCCACHE
251
ARG SCCACHE_ENDPOINT
252
253
# Install native tools needed only for Rust/protoc builds.
303
-RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
304
- dnf install -y --setopt=install_weak_deps=False \
305
- make unzip \
306
- && dnf clean all && rm -rf /var/cache/dnf; \
307
- else \
308
- apt-get update -y \
309
- && apt-get install -y --no-install-recommends \
310
- make unzip \
311
- && rm -rf /var/lib/apt/lists/*; \
312
- fi
254
+RUN dnf install -y --setopt=install_weak_deps=False make unzip \
255
+ && dnf clean all \
256
+ && rm -rf /var/cache/dnf
257
258
COPY tools/install_protoc.sh /tmp/install_protoc.sh
259
RUN /tmp/install_protoc.sh && rm /tmp/install_protoc.sh
@@ -606,7 +550,6 @@ FROM base AS dev
550
ARG PIP_INDEX_URL UV_INDEX_URL
551
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
552
ARG PYTORCH_CUDA_INDEX_BASE_URL
609
-ARG BUILD_OS
553
554
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
555
# Reference: https://github.com/astral-sh/uv/pull/1694
@@ -616,11 +559,9 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
559
ENV UV_LINK_MODE=copy
560
561
# Install libnuma-dev, required by fastsafetensors (fixes #20384)
619
-RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
620
- dnf install -y numactl-devel && dnf clean all && rm -rf /var/cache/dnf; \
621
- else \
622
- apt-get update && apt-get install -y --no-install-recommends libnuma-dev && rm -rf /var/lib/apt/lists/*; \
623
- fi
562
+RUN dnf install -y numactl-devel \
563
+ && dnf clean all \
564
+ && rm -rf /var/cache/dnf
565
566
567
# We can specify the standard or nightly build of PyTorch
docker/versions.json
+1
-4
@@ -14,14 +14,11 @@
14
"default": "2.30.7"
15
},
16
"BUILD_BASE_IMAGE": {
17
- "default": "nvidia/cuda:13.0.3-devel-ubuntu22.04"
17
+ "default": "pytorch/manylinux2_28-builder:cuda13.0"
18
},
19
"FINAL_BASE_IMAGE": {
20
"default": "nvidia/cuda:13.0.3-base-ubuntu22.04"
21
},
22
- "BUILD_OS": {
23
- "default": "ubuntu"
24
- },
22
"GET_PIP_URL": {
23
"default": "https://bootstrap.pypa.io/get-pip.py"
24
},
docs/assets/contributing/dockerfile-stages-dependency.png
Binary files a/docs/assets/contributing/dockerfile-stages-dependency.png and b/docs/assets/contributing/dockerfile-stages-dependency.png differ
docs/getting_started/installation/gpu.cuda.inc.md
+2
-1
@@ -384,6 +384,7 @@ A docker container can be built for aarch64 systems such as the Nvidia Grace-Hop
384
-t vllm/vllm-gh200-openai:latest \
385
--build-arg max_jobs=66 \
386
--build-arg nvcc_threads=2 \
387
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 \
388
--build-arg torch_cuda_arch_list="9.0 10.0+PTX"
389
```
390
@@ -394,7 +395,7 @@ For (G)B300, we recommend using CUDA 13, as shown in the following command.
395
```bash
396
DOCKER_BUILDKIT=1 docker build \
397
--build-arg CUDA_VERSION=13.0.2 \
397
- --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
398
+ --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 \
399
--build-arg max_jobs=256 \
400
--build-arg nvcc_threads=2 \
401
--build-arg torch_cuda_arch_list='9.0 10.0+PTX' \