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 16, 23:15 UTC
[ci] Setup Release pipeline and build release wheels with cache (#5610)
Signed-off-by: kevin <kevin@anyscale.com>
Kevin H. Luu committed
Jun 18, 2024 at 11:00 UTC
19091efc44c6f9b1e008dc5469c63a1f01684745
2 files changed
+62
-17
.buildkite/release-pipeline.yaml
new
+21
@@ -0,0 +1,21 @@
1
+steps:
2
+ - block: "Build wheels"
3
+
4
+ - label: "Build wheel - Python {{matrix.python_version}}, CUDA {{matrix.cuda_version}}"
5
+ agents:
6
+ queue: cpu_queue
7
+ commands:
8
+ - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION={{matrix.cuda_version}} --build-arg PYTHON_VERSION={{matrix.python_version}} --tag vllm-ci:build-image --target build --progress plain ."
9
+ - "mkdir artifacts"
10
+ - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image cp -r dist /artifacts_host"
11
+ - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
12
+ matrix:
13
+ setup:
14
+ cuda_version:
15
+ - "11.8.0"
16
+ - "12.1.0"
17
+ python_version:
18
+ - "3.8"
19
+ - "3.9"
20
+ - "3.10"
21
+ - "3.11"
Dockerfile
+41
-17
@@ -5,9 +5,26 @@
5
# docs/source/dev/dockerfile/dockerfile.rst and
6
# docs/source/assets/dev/dockerfile-stages-dependency.png
7
8
+ARG CUDA_VERSION=12.4.1
9
#################### BASE BUILD IMAGE ####################
10
# prepare basic build environment
10
-FROM nvidia/cuda:12.4.1-devel-ubuntu22.04 AS dev
11
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS base
12
+
13
+ARG CUDA_VERSION=12.4.1
14
+ARG PYTHON_VERSION=3
15
+
16
+ENV DEBIAN_FRONTEND=noninteractive
17
+
18
+RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
19
+ && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
20
+ && apt-get update -y \
21
+ && apt-get install -y ccache software-properties-common \
22
+ && add-apt-repository ppa:deadsnakes/ppa \
23
+ && apt-get update -y \
24
+ && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv python3-pip \
25
+ && if [ "${PYTHON_VERSION}" != "3" ]; then update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1; fi \
26
+ && python3 --version \
27
+ && python3 -m pip --version
28
29
RUN apt-get update -y \
30
&& apt-get install -y python3-pip git curl sudo
@@ -16,7 +33,7 @@ RUN apt-get update -y \
33
# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
34
# this won't be needed for future versions of this docker image
35
# or future versions of triton.
19
-RUN ldconfig /usr/local/cuda-12.4/compat/
36
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
37
38
WORKDIR /workspace
39
@@ -24,14 +41,7 @@ WORKDIR /workspace
41
COPY requirements-common.txt requirements-common.txt
42
COPY requirements-cuda.txt requirements-cuda.txt
43
RUN --mount=type=cache,target=/root/.cache/pip \
27
- pip install -r requirements-cuda.txt
28
-
29
-# install development dependencies
30
-COPY requirements-lint.txt requirements-lint.txt
31
-COPY requirements-test.txt requirements-test.txt
32
-COPY requirements-dev.txt requirements-dev.txt
33
-RUN --mount=type=cache,target=/root/.cache/pip \
34
- pip install -r requirements-dev.txt
44
+ python3 -m pip install -r requirements-cuda.txt
45
46
# cuda arch list used by torch
47
# can be useful for both `dev` and `test`
@@ -41,14 +51,16 @@ ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
51
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
52
#################### BASE BUILD IMAGE ####################
53
44
-
54
#################### WHEEL BUILD IMAGE ####################
46
-FROM dev AS build
55
+FROM base AS build
56
+
57
+ARG PYTHON_VERSION=3
58
59
# install build dependencies
60
COPY requirements-build.txt requirements-build.txt
61
+
62
RUN --mount=type=cache,target=/root/.cache/pip \
51
- pip install -r requirements-build.txt
63
+ python3 -m pip install -r requirements-build.txt
64
65
# install compiler cache to speed up compilation leveraging local or remote caching
66
RUN apt-get update -y && apt-get install -y ccache
@@ -101,9 +113,21 @@ RUN python3 check-wheel-size.py dist
113
114
#################### EXTENSION Build IMAGE ####################
115
116
+#################### DEV IMAGE ####################
117
+FROM base as dev
118
+
119
+COPY requirements-lint.txt requirements-lint.txt
120
+COPY requirements-test.txt requirements-test.txt
121
+COPY requirements-dev.txt requirements-dev.txt
122
+RUN --mount=type=cache,target=/root/.cache/pip \
123
+ python3 -m pip install -r requirements-dev.txt
124
+
125
+#################### DEV IMAGE ####################
126
+
127
#################### vLLM installation IMAGE ####################
128
# image with vLLM installed
106
-FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS vllm-base
129
+FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
130
+ARG CUDA_VERSION=12.4.1
131
WORKDIR /vllm-workspace
132
133
RUN apt-get update -y \
@@ -113,12 +137,12 @@ RUN apt-get update -y \
137
# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
138
# this won't be needed for future versions of this docker image
139
# or future versions of triton.
116
-RUN ldconfig /usr/local/cuda-12.4/compat/
140
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
141
142
# install vllm wheel first, so that torch etc will be installed
143
RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
144
--mount=type=cache,target=/root/.cache/pip \
121
- pip install dist/*.whl --verbose
145
+ python3 -m pip install dist/*.whl --verbose
146
#################### vLLM installation IMAGE ####################
147
148
@@ -131,7 +155,7 @@ ADD . /vllm-workspace/
155
156
# install development dependencies (for testing)
157
RUN --mount=type=cache,target=/root/.cache/pip \
134
- pip install -r requirements-dev.txt
158
+ python3 -m pip install -r requirements-dev.txt
159
160
# doc requires source code
161
# we hide them inside `test_docs/` , so that this source code