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 17, 05:30 UTC
Fix docker python version (#2845)
Nikola Borisov committed
Feb 14, 2024 at 10:17 UTC
87069ccf68c1bd74aec5ff58db360977f0d9d757
2 files changed
+6
-17
Dockerfile
+5
-16
@@ -4,21 +4,8 @@
4
#################### BASE BUILD IMAGE ####################
5
FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS dev
6
7
-# Set the DEBIAN_FRONTEND variable to noninteractive to avoid interactive prompts
8
-ENV DEBIAN_FRONTEND=noninteractive
9
-
10
-# Preconfigure tzdata for US Central Time (build running in us-central-1 but this really doesn't matter.)
11
-RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
12
- && echo 'tzdata tzdata/Zones/America select Chicago' | debconf-set-selections
13
-
14
-# We install an older version of python here for testing to make sure vllm works with older versions of Python.
15
-# For the actual openai compatible server, we will use the latest version of Python.
7
RUN apt-get update -y \
17
- && apt-get install -y software-properties-common \
18
- && add-apt-repository ppa:deadsnakes/ppa -y \
19
- && apt-get update -y \
20
- && apt-get install -y python3.8 python3.8-dev python3.8-venv python3-pip git \
21
- && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.8 1
8
+ && apt-get install -y python3-pip git
9
10
# Workaround for https://github.com/openai/triton/issues/2507 and
11
# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -88,8 +75,10 @@ RUN --mount=type=cache,target=/root/.cache/pip VLLM_USE_PRECOMPILED=1 pip instal
75
76
77
#################### RUNTIME BASE IMAGE ####################
91
-# use CUDA base as CUDA runtime dependencies are already installed via pip
92
-FROM nvidia/cuda:12.1.0-base-ubuntu22.04 AS vllm-base
78
+# We used base cuda image because pytorch installs its own cuda libraries.
79
+# However cupy depends on cuda libraries so we had to switch to the runtime image
80
+# In the future it would be nice to get a container with pytorch and cuda without duplicating cuda
81
+FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 AS vllm-base
82
83
# libnccl required for ray
84
RUN apt-get update -y \
requirements.txt
+1
-1
@@ -12,4 +12,4 @@ pydantic >= 2.0 # Required for OpenAI server.
12
aioprometheus[starlette]
13
pynvml == 11.5.0
14
triton >= 2.1.0
15
-cupy-cuda12x == 12.3.0 # Required for CUDA graphs. CUDA 11.8 users should install cupy-cuda11x instead.
15
+cupy-cuda12x == 12.1.0 # Required for CUDA graphs. CUDA 11.8 users should install cupy-cuda11x instead.