7
7
8
8
# ################### BASE BUILD IMAGE ####################
9
9
# prepare basic build environment
10
- FROM nvidia/cuda:12.1.0 -devel-ubuntu22.04 AS dev
10
+ FROM nvidia/cuda:12.4.1 -devel-ubuntu22.04 AS dev
11
11
12
12
RUN apt-get update -y \
13
13
&& apt-get install -y python3-pip git
@@ -16,7 +16,7 @@ RUN apt-get update -y \
16
16
# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
17
17
# this won't be needed for future versions of this docker image
18
18
# or future versions of triton.
19
- RUN ldconfig /usr/local/cuda-12.1 /compat/
19
+ RUN ldconfig /usr/local/cuda-12.4 /compat/
20
20
21
21
WORKDIR /workspace
22
22
@@ -75,6 +75,10 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
75
75
--mount=type=cache,target=/root/.cache/pip \
76
76
python3 setup.py bdist_wheel --dist-dir=dist
77
77
78
+ # check the size of the wheel, we cannot upload wheels larger than 100MB
79
+ COPY .buildkite/check-wheel-size.py check-wheel-size.py
80
+ RUN python3 check-wheel-size.py dist
81
+
78
82
# the `vllm_nccl` package must be installed from source distribution
79
83
# pip is too smart to store a wheel in the cache, and other CI jobs
80
84
# will directly use the wheel from the cache, which is not what we want.
@@ -102,7 +106,7 @@ RUN pip --verbose wheel flash-attn==${FLASH_ATTN_VERSION} \
102
106
103
107
# ################### vLLM installation IMAGE ####################
104
108
# image with vLLM installed
105
- FROM nvidia/cuda:12.1.0 -base-ubuntu22.04 AS vllm-base
109
+ FROM nvidia/cuda:12.4.1 -base-ubuntu22.04 AS vllm-base
106
110
WORKDIR /vllm-workspace
107
111
108
112
RUN apt-get update -y \
@@ -112,7 +116,7 @@ RUN apt-get update -y \
112
116
# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
113
117
# this won't be needed for future versions of this docker image
114
118
# or future versions of triton.
115
- RUN ldconfig /usr/local/cuda-12.1 /compat/
119
+ RUN ldconfig /usr/local/cuda-12.4 /compat/
116
120
117
121
# install vllm wheel first, so that torch etc will be installed
118
122
RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
0 commit comments