char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Adrien Gallouëtdevops : add llama in all docker images (#25035)5d8ccdf9d

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.7 KiB133 linesraw
1ARG UBUNTU_VERSION=24.04
2# This needs to generally match the container host's environment.
3ARG CUDA_VERSION=12.8.1
4ARG GCC_VERSION=14
5# Target the CUDA build image
6ARG BASE_CUDA_DEV_CONTAINER=docker.io/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
7
8ARG BASE_CUDA_RUN_CONTAINER=docker.io/nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
9
10ARG BUILD_DATE=N/A
11ARG APP_VERSION=N/A
12ARG APP_REVISION=N/A
13
14ARG NODE_VERSION=24
15
16FROM docker.io/node:$NODE_VERSION AS web
17
18ARG APP_VERSION
19
20WORKDIR /app/tools/ui
21
22COPY tools/ui/package.json tools/ui/package-lock.json ./
23RUN npm ci
24
25COPY tools/ui/ ./
26RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
27
28FROM ${BASE_CUDA_DEV_CONTAINER} AS build
29
30ARG GCC_VERSION
31# CUDA architecture to build for (defaults to all supported archs)
32ARG CUDA_DOCKER_ARCH=default
33
34RUN apt-get update && \
35    apt-get install -y gcc-${GCC_VERSION} g++-${GCC_VERSION} build-essential cmake python3 python3-pip git libssl-dev libgomp1
36
37ENV CC=gcc-${GCC_VERSION} CXX=g++-${GCC_VERSION} CUDAHOSTCXX=g++-${GCC_VERSION}
38
39WORKDIR /app
40
41COPY . .
42
43COPY --from=web /app/tools/ui/dist tools/ui/dist
44
45RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
46    export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
47    fi && \
48    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
49    cmake --build build --config Release -j$(nproc)
50
51RUN mkdir -p /app/lib && \
52    find build -name "*.so*" -exec cp -P {} /app/lib \;
53
54RUN mkdir -p /app/full \
55    && cp build/bin/* /app/full \
56    && cp *.py /app/full \
57    && cp -r conversion /app/full \
58    && cp -r gguf-py /app/full \
59    && cp -r requirements /app/full \
60    && cp requirements.txt /app/full \
61    && cp .devops/tools.sh /app/full/tools.sh
62
63## Base image
64FROM ${BASE_CUDA_RUN_CONTAINER} AS base
65
66ARG BUILD_DATE=N/A
67ARG APP_VERSION=N/A
68ARG APP_REVISION=N/A
69ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
70ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
71LABEL org.opencontainers.image.created=$BUILD_DATE \
72      org.opencontainers.image.version=$APP_VERSION \
73      org.opencontainers.image.revision=$APP_REVISION \
74      org.opencontainers.image.title="llama.cpp" \
75      org.opencontainers.image.description="LLM inference in C/C++" \
76      org.opencontainers.image.url=$IMAGE_URL \
77      org.opencontainers.image.source=$IMAGE_SOURCE
78
79RUN apt-get update \
80    && apt-get install -y libgomp1 curl ffmpeg \
81    && apt autoremove -y \
82    && apt clean -y \
83    && rm -rf /tmp/* /var/tmp/* \
84    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
85    && find /var/cache -type f -delete
86
87COPY --from=build /app/lib/ /app
88
89### Full
90FROM base AS full
91
92COPY --from=build /app/full /app
93
94WORKDIR /app
95
96RUN apt-get update \
97    && apt-get install -y \
98    git \
99    python3 \
100    python3-pip \
101    python3-wheel \
102    && pip install --break-system-packages --upgrade setuptools \
103    && pip install --break-system-packages -r requirements.txt \
104    && apt autoremove -y \
105    && apt clean -y \
106    && rm -rf /tmp/* /var/tmp/* \
107    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
108    && find /var/cache -type f -delete
109
110
111ENTRYPOINT ["/app/tools.sh"]
112
113### Light, CLI only
114FROM base AS light
115
116COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
117
118WORKDIR /app
119
120ENTRYPOINT [ "/app/llama-cli" ]
121
122### Server, Server only
123FROM base AS server
124
125ENV LLAMA_ARG_HOST=0.0.0.0
126
127COPY --from=build /app/full/llama /app/full/llama-server /app
128
129WORKDIR /app
130
131HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
132
133ENTRYPOINT [ "/app/llama-server" ]