-
Notifications
You must be signed in to change notification settings - Fork 4
Expand file tree
/
Copy pathDockerfile-gpu
More file actions
38 lines (30 loc) · 1.03 KB
/
Dockerfile-gpu
File metadata and controls
38 lines (30 loc) · 1.03 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
ARG PYTHON_VERSION=310
ARG CUDA_VERSION=cu121
ARG BASE_IMAGE=mirror.gcr.io/rayproject/ray:latest-py${PYTHON_VERSION}-${CUDA_VERSION}
FROM $BASE_IMAGE
# AP 19/05/2025:
# CUDA_HOME is not set in the image and is required
ENV CUDA_HOME=/usr/local/cuda
# Set LD_LIBRARY_PATH to avoid runtime errors (so far only experienced when running vLLM with the vllm_performance actuator)
ENV LD_LIBRARY_PATH=/home/ray/anaconda3/lib:${LD_LIBRARY_PATH}
# Pass SOURCE_DATE_EPOCH to guarantee that all the wheels built will use
# the same timestamp.
ARG SOURCE_DATE_EPOCH=0
USER 0
RUN chgrp -R 0 /home/ray && \
chmod -R g=u /home/ray
USER 1000
# Copy source code
COPY --chown=1000:0 ./ ado
## Install ado
RUN export SOURCE_DATE_EPOCH=${SOURCE_DATE_EPOCH} && \
cd ado && \
uv pip install --system -r requirements.txt && \
uv pip install --system --no-deps --strict \
'.'
## Install ray-tune operator
RUN export SOURCE_DATE_EPOCH=${SOURCE_DATE_EPOCH} && \
cd ado && \
uv pip install --system \
'plugins/operators/ray_tune/'
CMD ado -h