@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 916d219d62
Fork 0

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] Use sccache to build images (#5419)

Signed-off-by: kevin <kevin@anyscale.com>

Kevin H. Luu committed Jun 12, 2024 at 17:58 UTC 916d219d62e9e4005e10be23f81d881afdb8d6d0
3 files changed +22 -4
.buildkite/test-template-aws.j2
+1 -1
@@ -7,7 +7,7 @@ steps:
7 queue: cpu_queue
8 commands:
9 - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
10 - - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
10 + - "docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --tag {{ docker_image }} --target test --progress plain ."
11 - "docker push {{ docker_image }}"
12 env:
13 DOCKER_BUILDKIT: "1"
Dockerfile
+20 -2
@@ -10,7 +10,7 @@
10 FROM nvidia/cuda:12.4.1-devel-ubuntu22.04 AS dev
11
12 RUN apt-get update -y \
13 - && apt-get install -y python3-pip git
13 + && apt-get install -y python3-pip git curl sudo
14
15 # Workaround for https://github.com/openai/triton/issues/2507 and
16 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -70,10 +70,28 @@ ENV NVCC_THREADS=$nvcc_threads
70 # make sure punica kernels are built (for LoRA)
71 ENV VLLM_INSTALL_PUNICA_KERNELS=1
72
73 +ARG USE_SCCACHE
74 +# if USE_SCCACHE is set, use sccache to speed up compilation
75 +RUN --mount=type=cache,target=/root/.cache/pip \
76 + if [ "$USE_SCCACHE" = "1" ]; then \
77 + echo "Installing sccache..." \
78 + && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
79 + && tar -xzf sccache.tar.gz \
80 + && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
81 + && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
82 + && export SCCACHE_BUCKET=vllm-build-sccache \
83 + && export SCCACHE_REGION=us-west-2 \
84 + && sccache --show-stats \
85 + && python3 setup.py bdist_wheel --dist-dir=dist \
86 + && sccache --show-stats; \
87 + fi
88 +
89 ENV CCACHE_DIR=/root/.cache/ccache
90 RUN --mount=type=cache,target=/root/.cache/ccache \
91 --mount=type=cache,target=/root/.cache/pip \
76 - python3 setup.py bdist_wheel --dist-dir=dist
92 + if [ "$USE_SCCACHE" != "1" ]; then \
93 + python3 setup.py bdist_wheel --dist-dir=dist; \
94 + fi
95
96 # check the size of the wheel, we cannot upload wheels larger than 100MB
97 COPY .buildkite/check-wheel-size.py check-wheel-size.py
setup.py
+1 -1
@@ -140,6 +140,7 @@ class cmake_build_ext(build_ext):
140 cmake_args += [
141 '-DCMAKE_CXX_COMPILER_LAUNCHER=sccache',
142 '-DCMAKE_CUDA_COMPILER_LAUNCHER=sccache',
143 + '-DCMAKE_C_COMPILER_LAUNCHER=sccache',
144 ]
145 elif is_ccache_available():
146 cmake_args += [
@@ -171,7 +172,6 @@ class cmake_build_ext(build_ext):
172 else:
173 # Default build tool to whatever cmake picks.
174 build_tool = []
174 -
175 subprocess.check_call(
176 ['cmake', ext.cmake_lists_dir, *build_tool, *cmake_args],
177 cwd=self.build_temp)