char-fork/llama.cpp

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

Aaron Teoci: rm `GGML_HIP_ROCWMMA_FATTN` (#26760)61141f148

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
4.0 KiB146 linesraw
1ARG UBUNTU_VERSION=24.04
2
3# This needs to generally match the container host's environment.
4ARG ROCM_VERSION=7.2.1
5ARG AMDGPU_VERSION=7.2.1
6
7# Target the ROCm build image
8ARG BASE_ROCM_DEV_CONTAINER=docker.io/rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
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
28### Build image
29FROM ${BASE_ROCM_DEV_CONTAINER} AS build
30
31# Unless otherwise specified, we make a fat build.
32# This is mostly tied to rocBLAS supported archs.
33# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-7.2.1/reference/system-requirements.html
34# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/native_linux/native_linux_compatibility.html
35# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityryz/native_linux/native_linux_compatibility.html
36
37ARG ROCM_DOCKER_ARCH='gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201'
38
39# Set ROCm architectures
40ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
41
42RUN apt-get update \
43    && apt-get install -y \
44    build-essential \
45    cmake \
46    git \
47    libssl-dev \
48    curl \
49    libgomp1
50
51WORKDIR /app
52
53COPY . .
54
55COPY --from=web /app/tools/ui/dist tools/ui/dist
56
57RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
58    cmake -S . -B build \
59        -DGGML_HIP=ON \
60        -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
61        -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
62        -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
63    && cmake --build build --config Release -j$(nproc)
64
65RUN mkdir -p /app/lib \
66    && find build -name "*.so*" -exec cp -P {} /app/lib \;
67
68RUN mkdir -p /app/full \
69    && cp build/bin/* /app/full \
70    && cp *.py /app/full \
71    && cp -r conversion /app/full \
72    && cp -r gguf-py /app/full \
73    && cp -r requirements /app/full \
74    && cp requirements.txt /app/full \
75    && cp .devops/tools.sh /app/full/tools.sh
76
77## Base image
78FROM ${BASE_ROCM_DEV_CONTAINER} AS base
79
80ARG BUILD_DATE=N/A
81ARG APP_VERSION=N/A
82ARG APP_REVISION=N/A
83ARG IMAGE_URL=https://github.com/ggml-org/llama.cpp
84ARG IMAGE_SOURCE=https://github.com/ggml-org/llama.cpp
85LABEL org.opencontainers.image.created=$BUILD_DATE \
86      org.opencontainers.image.version=$APP_VERSION \
87      org.opencontainers.image.revision=$APP_REVISION \
88      org.opencontainers.image.title="llama.cpp" \
89      org.opencontainers.image.description="LLM inference in C/C++" \
90      org.opencontainers.image.url=$IMAGE_URL \
91      org.opencontainers.image.source=$IMAGE_SOURCE
92
93RUN apt-get update \
94    && apt-get install -y libgomp1 curl ffmpeg \
95    && apt autoremove -y \
96    && apt clean -y \
97    && rm -rf /tmp/* /var/tmp/* \
98    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
99    && find /var/cache -type f -delete
100
101COPY --from=build /app/lib/ /app
102
103### Full
104FROM base AS full
105
106COPY --from=build /app/full /app
107
108WORKDIR /app
109
110RUN apt-get update \
111    && apt-get install -y \
112    git \
113    python3-pip \
114    python3 \
115    python3-wheel \
116    && pip install --break-system-packages --upgrade setuptools \
117    && pip install --break-system-packages -r requirements.txt \
118    && apt autoremove -y \
119    && apt clean -y \
120    && rm -rf /tmp/* /var/tmp/* \
121    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
122    && find /var/cache -type f -delete
123
124ENTRYPOINT ["/app/tools.sh"]
125
126### Light, CLI only
127FROM base AS light
128
129COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app
130
131WORKDIR /app
132
133ENTRYPOINT [ "/app/llama-cli" ]
134
135### Server, Server only
136FROM base AS server
137
138ENV LLAMA_ARG_HOST=0.0.0.0
139
140COPY --from=build /app/full/llama /app/full/llama-server /app
141
142WORKDIR /app
143
144HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
145
146ENTRYPOINT [ "/app/llama-server" ]