wheel_builder.Dockerfile 19.5 KB
Newer Older
1
2
{#
# SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
3
# SPDX-License-Identifier: Apache-2.0
4
#}
5
# === BEGIN templates/wheel_builder.Dockerfile ===
6
7
8
9
##################################
##### Wheel Build Image ##########
##################################

10
FROM ${WHEEL_BUILDER_IMAGE} AS wheel_builder
11
12
13
14
15

# Redeclare ARGs for this stage
ARG ARCH
ARG ARCH_ALT
ARG CARGO_BUILD_JOBS
16
ARG DEVICE
17
18
19

WORKDIR /workspace

20
{% if device == "cuda" %}
21
22
23
# Copy CUDA from base stage
COPY --from=dynamo_base /usr/local/cuda /usr/local/cuda
COPY --from=dynamo_base /etc/ld.so.conf.d/hpcx.conf /etc/ld.so.conf.d/hpcx.conf
24
{% endif %}
25
26
27
28
29
30
31
32
33
34
35

# Set environment variables first so they can be used in COPY commands
ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16} \
    RUSTUP_HOME=/usr/local/rustup \
    CARGO_HOME=/usr/local/cargo \
    CARGO_TARGET_DIR=/opt/dynamo/target \
    PATH=/usr/local/cargo/bin:$PATH

# Copy artifacts from base stage
COPY --from=dynamo_base $RUSTUP_HOME $RUSTUP_HOME
COPY --from=dynamo_base $CARGO_HOME $CARGO_HOME
36

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
{% if device == "xpu" %}
SHELL ["/bin/bash", "-o", "pipefail", "-c"]
RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
    add-apt-repository -y ppa:kobuk-team/intel-graphics

RUN wget --tries=3 --waitretry=5 https://raw.githubusercontent.com/intel/llm-scaler/35a14cbc08d714f460a29b7a7328df5620c8530f/vllm/patches/ai-dynamo-xpu/patches/ucx-v1.12.0.patch -O /tmp/ucx.patch

RUN apt clean && apt-get update -y && \
    apt-get install -y --no-install-recommends --fix-missing \
    curl \
    #ffmpeg \
    ca-certificates \
    zip \
    unzip \
    git \
    libsndfile1 \
    libsm6 \
    libxext6 \
    libgl1 \
    lsb-release \
    libaio-dev \
    numactl \
    wget \
    vim \
    linux-libc-dev && \
    # Install Intel GPU runtime packages
    apt update -y && apt upgrade -y && \
    apt-get install -y libze1 libze-dev libze-intel-gpu1 intel-opencl-icd  \
    libze-intel-gpu-raytracing intel-ocloc intel-oneapi-compiler-dpcpp-cpp-2025.3 && \
    apt-get clean && rm -rf /var/lib/apt/lists/*

RUN apt-get update -y \
    && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
        # NIXL build dependencies
        autoconf \
        automake \
        cmake \
        git-lfs \
        libtool \
        meson \
        net-tools \
        ninja-build \
        pybind11-dev \
        # Rust build dependencies
        clang \
        libclang-dev \
        protobuf-compiler \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*

RUN apt-get update -y \
    && DEBIAN_FRONTEND=noninteractive apt-get -y install --reinstall --no-install-recommends \
        libibverbs-dev \
        rdma-core \
        ibverbs-utils \
        libibumad-dev \
        libnuma-dev \
        librdmacm-dev \
        ibverbs-providers \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*
{% endif %}

{% if device == "cuda" %}
102
# Install system dependencies
103
104
105
# Cache dnf downloads; sharing=locked avoids dnf/rpm races with concurrent builds.
RUN --mount=type=cache,target=/var/cache/dnf,sharing=locked \
    dnf install -y almalinux-release-synergy && \
106
107
    dnf config-manager --set-enabled powertools && \
    dnf install -y \
108
109
110
111
112
        # Autotools (required for UCX, libfabric ./autogen.sh and ./configure)
        autoconf \
        automake \
        libtool \
        make \
113
        # RPM build tools (required for gdrcopy's build-rpm-packages.sh)
114
115
        rpm-build \
        rpm-sign \
116
117
118
119
        # Build tools
        cmake \
        ninja-build \
        clang-devel \
120
121
122
123
        # Install GCC toolset 14 (CUDA compatible, max version 14)
        gcc-toolset-14-gcc \
        gcc-toolset-14-gcc-c++ \
        gcc-toolset-14-binutils \
124
125
126
127
128
129
130
131
132
133
134
135
136
137
        flex \
        wget \
        # Kernel module build dependencies
        dkms \
        # Protobuf support
        protobuf-compiler \
        # RDMA/InfiniBand support (required for UCX build with --with-verbs)
        libibverbs \
        libibverbs-devel \
        rdma-core \
        rdma-core-devel \
        libibumad \
        libibumad-devel \
        librdmacm-devel \
Joe Chandler's avatar
Joe Chandler committed
138
139
140
        numactl-devel \
        # Libfabric support
        hwloc \
141
142
143
144
        hwloc-devel \
        libcurl-devel \
        openssl-devel \
        libuuid-devel \
145
        zlib-devel
146

147
148
149
150
151
# Set GCC toolset 14 as the default compiler (CUDA requires GCC <= 14)
ENV PATH="/opt/rh/gcc-toolset-14/root/usr/bin:${PATH}" \
    LD_LIBRARY_PATH="/opt/rh/gcc-toolset-14/root/usr/lib64:${LD_LIBRARY_PATH}" \
    CC="/opt/rh/gcc-toolset-14/root/usr/bin/gcc" \
    CXX="/opt/rh/gcc-toolset-14/root/usr/bin/g++"
152
{% endif %}
153

154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
# Ensure a modern protoc is available (required for --experimental_allow_proto3_optional)
RUN set -eux; \
    PROTOC_VERSION=25.3; \
    case "${ARCH_ALT}" in \
      x86_64) PROTOC_ZIP="protoc-${PROTOC_VERSION}-linux-x86_64.zip" ;; \
      aarch64) PROTOC_ZIP="protoc-${PROTOC_VERSION}-linux-aarch_64.zip" ;; \
      *) echo "Unsupported architecture: ${ARCH_ALT}" >&2; exit 1 ;; \
    esac; \
    wget --tries=3 --waitretry=5 -O /tmp/protoc.zip "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/${PROTOC_ZIP}"; \
    rm -f /usr/local/bin/protoc /usr/bin/protoc; \
    unzip -o /tmp/protoc.zip -d /usr/local bin/protoc include/*; \
    chmod +x /usr/local/bin/protoc; \
    ln -s /usr/local/bin/protoc /usr/bin/protoc; \
    protoc --version

# Point build tools explicitly at the modern protoc
ENV PROTOC=/usr/local/bin/protoc

172
173
174
175
176
{% if device == "xpu" %}
# Install uv package manager
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
ENV LD_LIBRARY_PATH=/usr/local/lib:/usr/local/lib64:$LD_LIBRARY_PATH
{% else %}
177
178
179
180
ENV CUDA_PATH=/usr/local/cuda \
    PATH=/usr/local/cuda/bin:$PATH \
    LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/lib:/usr/local/lib64:$LD_LIBRARY_PATH \
    NVIDIA_DRIVER_CAPABILITIES=video,compute,utility
181
{% endif %}
182
183
184
185

# Create virtual environment for building wheels
ARG PYTHON_VERSION
ENV VIRTUAL_ENV=/workspace/.venv
186
187
188
189
# Cache uv downloads; uv handles its own locking for this cache.
RUN --mount=type=cache,target=/root/.cache/uv \
    export UV_CACHE_DIR=/root/.cache/uv UV_HTTP_TIMEOUT=300 UV_HTTP_RETRIES=5 && \
    uv venv ${VIRTUAL_ENV} --python $PYTHON_VERSION && \
190
    uv pip install --upgrade meson pybind11 patchelf maturin[patchelf] tomlkit
191
192
193

ARG NIXL_UCX_REF
ARG NIXL_REF
194
195

{% if device == "cuda" %}
196
197
198
199
200
201
202
203
204
ARG NIXL_GDRCOPY_REF

# Build and install gdrcopy
RUN git clone --depth 1 --branch ${NIXL_GDRCOPY_REF} https://github.com/NVIDIA/gdrcopy.git && \
    cd gdrcopy/packages && \
    CUDA=/usr/local/cuda ./build-rpm-packages.sh && \
    rpm -Uvh gdrcopy-kmod-*.el8.noarch.rpm && \
    rpm -Uvh gdrcopy-*.el8.${ARCH_ALT}.rpm && \
    rpm -Uvh gdrcopy-devel-*.el8.noarch.rpm
205
{% endif %}
206

207
208
209
210
211
# sccache binary is pre-installed in dynamo_base; stage it off-PATH so
# Meson doesn't auto-detect it as a CUDA compiler launcher
# (https://github.com/mesonbuild/meson/issues/11118).
# When USE_SCCACHE=true the RUN below symlinks it onto PATH before install.
COPY --from=dynamo_base /usr/local/bin/sccache /opt/sccache/sccache
212

213
214
215
216
217
ARG USE_SCCACHE
ARG SCCACHE_BUCKET
ARG SCCACHE_REGION
COPY container/use-sccache.sh /tmp/use-sccache.sh
RUN if [ "$USE_SCCACHE" = "true" ]; then \
218
        ln -s /opt/sccache/sccache /usr/local/bin/sccache && \
219
220
221
        /tmp/use-sccache.sh install; \
    fi

222
223
# Set SCCACHE environment variables (RUSTC_WRAPPER is set dynamically by
# setup-env only when the sccache server starts successfully)
224
ENV SCCACHE_BUCKET=${USE_SCCACHE:+${SCCACHE_BUCKET}} \
225
    SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION}}
226

227
# Always build FFmpeg so libs are available for Rust checks in CI
milesial's avatar
milesial committed
228
# Do not delete the source tarball for legal reasons
229
ARG FFMPEG_VERSION
230
231
232
233
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
    if [ "$USE_SCCACHE" = "true" ]; then \
234
        eval $(/tmp/use-sccache.sh setup-env); \
235
    fi && \
236
237
238
239
240
241
    if [ "$DEVICE" = "xpu" ]; then \
    apt-get update -y && apt-get install -y pkg-config; \
    apt-get clean && rm -rf /var/lib/apt/lists/*; \
    elif [ "$DEVICE" = "cuda" ]; then \
    dnf install -y pkg-config; \
    fi && \
242
243
244
245
246
247
    cd /tmp && \
    curl -LO https://ffmpeg.org/releases/ffmpeg-${FFMPEG_VERSION}.tar.xz && \
    tar xf ffmpeg-${FFMPEG_VERSION}.tar.xz && \
    cd ffmpeg-${FFMPEG_VERSION} && \
    ./configure \
        --prefix=/usr/local \
milesial's avatar
milesial committed
248
249
        --disable-gpl \
        --disable-nonfree \
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
        --disable-programs \
        --disable-doc \
        --disable-static \
        --disable-x86asm \
        --disable-postproc \
        --disable-network \
        --disable-encoders \
        --disable-muxers \
        --disable-bsfs \
        --disable-devices \
        --disable-libdrm \
        --enable-shared && \
    make -j$(nproc) && \
    make install && \
    /tmp/use-sccache.sh show-stats "FFMPEG" && \
    ldconfig && \
milesial's avatar
milesial committed
266
    mkdir -p /usr/local/src/ffmpeg && \
267
    mv /tmp/ffmpeg-${FFMPEG_VERSION}* /usr/local/src/ffmpeg/
268

269
270
271
272
# Build and install UCX
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
273
    if [ "$USE_SCCACHE" = "true" ]; then \
274
        eval $(/tmp/use-sccache.sh setup-env); \
275
    fi && \
276
    cd /usr/local/src && \
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
    git clone https://github.com/openucx/ucx.git && \
    cd ucx &&  \
    git checkout $NIXL_UCX_REF &&	 \
    if [ "$DEVICE" = "xpu" ]; then \
    git apply --ignore-whitespace /tmp/ucx.patch; \
    fi && \
    ./autogen.sh &&      \
    if [ "$DEVICE" = "xpu" ]; then \
     ./contrib/configure-release     \
        --prefix=/usr/local/ucx     \
        --with-ze                   \
        --enable-shared             \
        --disable-static            \
        --disable-doxygen-doc       \
        --enable-optimizations      \
        --enable-cma                \
        --enable-devel-headers      \
        --with-verbs                \
        --with-dm                   \
        --with-efa                  \
        --without-cuda              \
        --enable-mt;                 \
    elif [ "$DEVICE" = "cuda" ]; then \
     ./contrib/configure-release     \
301
302
303
304
305
306
307
308
309
310
311
312
        --prefix=/usr/local/ucx     \
        --enable-shared             \
        --disable-static            \
        --disable-doxygen-doc       \
        --enable-optimizations      \
        --enable-cma                \
        --enable-devel-headers      \
        --with-cuda=/usr/local/cuda \
        --with-verbs                \
        --with-dm                   \
        --with-gdrcopy=/usr/local   \
        --with-efa                  \
313
314
        --enable-mt;                 \
     fi && \
315
316
317
318
319
320
321
     make -j &&                      \
     make -j install-strip &&        \
     /tmp/use-sccache.sh show-stats "UCX" && \
     echo "/usr/local/ucx/lib" > /etc/ld.so.conf.d/ucx.conf && \
     echo "/usr/local/ucx/lib/ucx" >> /etc/ld.so.conf.d/ucx.conf && \
     ldconfig

322
{% if device == "cuda" %}
Joe Chandler's avatar
Joe Chandler committed
323
324
325
326
327
ARG NIXL_LIBFABRIC_REF
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
    if [ "$USE_SCCACHE" = "true" ]; then \
328
        eval $(/tmp/use-sccache.sh setup-env); \
Joe Chandler's avatar
Joe Chandler committed
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
    fi && \
    cd /usr/local/src && \
    git clone https://github.com/ofiwg/libfabric.git && \
    cd libfabric && \
    git checkout $NIXL_LIBFABRIC_REF && \
    ./autogen.sh && \
    ./configure --prefix="/usr/local/libfabric" \
                --disable-verbs \
                --disable-psm3 \
                --disable-opx \
                --disable-usnic \
                --disable-rstream \
                --enable-efa \
                --with-cuda=/usr/local/cuda \
                --enable-cuda-dlopen \
                --with-gdrcopy \
                --enable-gdrcopy-dlopen && \
    make -j$(nproc) && \
    make install && \
    /tmp/use-sccache.sh show-stats "LIBFABRIC" && \
    echo "/usr/local/libfabric/lib" > /etc/ld.so.conf.d/libfabric.conf && \
    ldconfig
351
{% endif %}
Joe Chandler's avatar
Joe Chandler committed
352

353
{% if framework == "vllm" and device == "cuda" %}
354
355
356
357
358
# Build and install AWS SDK C++ (required for NIXL OBJ backend / S3 support)
ARG AWS_SDK_CPP_VERSION=1.11.581
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
359
360
361
    if [ "$USE_SCCACHE" = "true" ]; then \
        eval $(/tmp/use-sccache.sh setup-env cmake); \
    fi && \
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
    git clone --recurse-submodules --depth 1 --branch ${AWS_SDK_CPP_VERSION} \
        https://github.com/aws/aws-sdk-cpp.git /tmp/aws-sdk-cpp && \
    mkdir -p /tmp/aws-sdk-cpp/build && \
    cd /tmp/aws-sdk-cpp/build && \
    cmake .. \
        -DCMAKE_BUILD_TYPE=Release \
        -DBUILD_ONLY="s3" \
        -DENABLE_TESTING=OFF \
        -DCMAKE_INSTALL_PREFIX=/usr/local \
        -DBUILD_SHARED_LIBS=ON && \
    make -j$(nproc) && \
    make install && \
    cd / && \
    rm -rf /tmp/aws-sdk-cpp && \
    ldconfig && \
    /tmp/use-sccache.sh show-stats "AWS SDK C++"
{% endif %}

380
# build and install nixl
381
{% if device == "cuda" %}
382
ARG CUDA_MAJOR
383
384
{% endif %}

385
386
387
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
388
    if [ "$USE_SCCACHE" = "true" ]; then \
389
        eval $(/tmp/use-sccache.sh setup-env); \
390
    fi && \
391
    source ${VIRTUAL_ENV}/bin/activate && \
392
    git clone "https://github.com/ai-dynamo/nixl.git" && \
393
    cd nixl && \
394
    git checkout ${NIXL_REF} && \
395
396
397
398
399
    if [ "$DEVICE" = "cuda" ]; then \
        PKG_NAME="nixl-cu${CUDA_MAJOR}"; \
    elif [ "$DEVICE" = "xpu" ]; then \
        PKG_NAME="nixl-xpu"; \
    fi && \
400
    ./contrib/tomlutil.py --wheel-name $PKG_NAME pyproject.toml && \
401
    mkdir build && \
402
403
404
405
406
407
408
409
410
411
    if [ "$DEVICE" = "cuda" ]; then \
        meson setup build/ --prefix=/opt/nvidia/nvda_nixl --buildtype=release \
            -Dcudapath_lib="/usr/local/cuda/lib64" \
            -Dcudapath_inc="/usr/local/cuda/include" \
            -Ducx_path="/usr/local/ucx" \
            -Dlibfabric_path="/usr/local/libfabric"; \
    elif [ "$DEVICE" = "xpu" ]; then \
        meson setup build/ --prefix=/opt/intel/intel_nixl --buildtype=release \
            -Ducx_path="/usr/local/ucx"; \
    fi && \
412
413
414
415
416
    cd build && \
    ninja && \
    ninja install && \
    /tmp/use-sccache.sh show-stats "NIXL"

417
418
419
420
421
{% if device == "xpu" %}
ENV NIXL_LIB_DIR=/opt/intel/intel_nixl/lib/${ARCH_ALT}-linux-gnu  \
    NIXL_PLUGIN_DIR=/opt/intel/intel_nixl/lib/${ARCH_ALT}-linux-gnu/plugins \
    NIXL_PREFIX=/opt/intel/intel_nixl
{% else %}
422
423
424
ENV NIXL_LIB_DIR=/opt/nvidia/nvda_nixl/lib64  \
    NIXL_PLUGIN_DIR=/opt/nvidia/nvda_nixl/lib64/plugins \
    NIXL_PREFIX=/opt/nvidia/nvda_nixl
425
426
{% endif %}

427
428
429
430
431
432
ENV LD_LIBRARY_PATH=${NIXL_LIB_DIR}:${NIXL_PLUGIN_DIR}:/usr/local/ucx/lib:/usr/local/ucx/lib/ucx:${LD_LIBRARY_PATH}

RUN echo "$NIXL_LIB_DIR" > /etc/ld.so.conf.d/nixl.conf && \
    echo "$NIXL_PLUGIN_DIR" >> /etc/ld.so.conf.d/nixl.conf && \
    ldconfig

433
# Build NIXL wheel → /opt/dynamo/dist/nixl/nixl*.whl (C++ transport library, all targets)
434
435
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
436
437
    --mount=type=cache,target=/root/.cache/uv \
    export UV_CACHE_DIR=/root/.cache/uv && \
438
    export SCCACHE_S3_KEY_PREFIX="${SCCACHE_S3_KEY_PREFIX:-${ARCH}}" && \
439
    if [ "$USE_SCCACHE" = "true" ]; then \
440
        eval $(/tmp/use-sccache.sh setup-env); \
441
    fi && \
442
    cd /workspace/nixl && \
443
    uv build . --wheel --out-dir /opt/dynamo/dist/nixl --python $PYTHON_VERSION
444

445
{% if target not in ("dev", "local-dev") %}
446
447
448
449
450
# Copy source code (order matters for layer caching)
COPY pyproject.toml README.md LICENSE Cargo.toml Cargo.lock rust-toolchain.toml hatch_build.py /opt/dynamo/
COPY lib/ /opt/dynamo/lib/
COPY components/ /opt/dynamo/components/

451
452
453
454
455
# Build dynamo wheels → /opt/dynamo/dist/:
#   uv build        → ai_dynamo*any.whl          (main Python package)
#   maturin build   → ai_dynamo_runtime*.whl     (Rust bindings)
#   maturin build   → kvbm*.whl                  (KV block manager, conditional on ENABLE_KVBM)
# The caches do not need the "shared" lock because Cargo has its own locking mechanism.
456
ARG ENABLE_KVBM
457
ARG ENABLE_MEDIA_FFMPEG
458
459
RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
    --mount=type=secret,id=aws-secret-id,env=AWS_SECRET_ACCESS_KEY \
460
461
462
463
    --mount=type=cache,target=/root/.cargo/registry \
    --mount=type=cache,target=/root/.cargo/git \
    --mount=type=cache,target=/root/.cache/uv \
    export UV_CACHE_DIR=/root/.cache/uv && \
464
    export SCCACHE_S3_KEY_PREFIX=${SCCACHE_S3_KEY_PREFIX:-${ARCH}} && \
465
    if [ "$USE_SCCACHE" = "true" ]; then \
466
        eval $(/tmp/use-sccache.sh setup-env cmake); \
467
    fi && \
468
    mkdir -p ${CARGO_TARGET_DIR} && \
469
470
471
472
    source ${VIRTUAL_ENV}/bin/activate && \
    cd /opt/dynamo && \
    uv build --wheel --out-dir /opt/dynamo/dist && \
    cd /opt/dynamo/lib/bindings/python && \
473
    if [ "$ENABLE_MEDIA_FFMPEG" = "true" ]; then \
474
        maturin build --release --features "media-ffmpeg" --out /opt/dynamo/dist; \
475
476
477
    else \
        maturin build --release --out /opt/dynamo/dist; \
    fi && \
478
    if [ "$ENABLE_KVBM" = "true" ]; then \
479
480
        cd /opt/dynamo/lib/bindings/kvbm && \
        maturin build --release --out target/wheels && \
481
482
483
484
485
486
487
488
489
490
491
492
        if [ "$DEVICE" = "cuda" ]; then \
            auditwheel repair \
                --exclude libnixl.so \
                --exclude libnixl_build.so \
                --exclude libnixl_common.so \
                --exclude 'lib*.so*' \
                --plat manylinux_2_28_${ARCH_ALT} \
                --wheel-dir /opt/dynamo/dist \
                target/wheels/*.whl; \
        elif [ "$DEVICE" = "xpu" ]; then \
            cp target/wheels/*.whl /opt/dynamo/dist/; \
        fi; \
493
494
    fi && \
    /tmp/use-sccache.sh show-stats "Dynamo"
495

496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
{% else %}
# Dev/local-dev targets do not have pre-built wheels or /workspace source code.
# After you start the local-dev/dev container, you will need to build from source:
#   cargo build --features dynamo-llm/block-manager
#   cd /workspace/lib/bindings/python && maturin develop --uv && cd /workspace
#   uv pip install --no-deps -e /workspace
# See container/launch_message/dev.txt for the full setup steps.

# Create dist dir with a placeholder so downstream COPY --from=wheel_builder /opt/dynamo/dist/*.whl always has a match.
RUN mkdir -p /opt/dynamo/dist ${CARGO_TARGET_DIR} && \
    touch /opt/dynamo/dist/.placeholder.whl

# Dev/local-dev skip the full COPY lib/ above, so copy gpu_memory_service source explicitly for the wheel build below
COPY lib/gpu_memory_service/ /opt/dynamo/lib/gpu_memory_service/
{% endif %}

# Build gpu-memory-service wheel → /opt/dynamo/dist/gpu_memory_service*.whl (small C++ extension, fast build -- all targets, all frameworks)
513
514
{% if device == "cuda" %}
# Build gpu_memory_service wheel (C++ extension only needs Python headers, no CUDA/torch)
515
ARG ENABLE_GPU_MEMORY_SERVICE
516
517
518
RUN --mount=type=cache,target=/root/.cache/uv \
    if [ "$ENABLE_GPU_MEMORY_SERVICE" = "true" ]; then \
        export UV_CACHE_DIR=/root/.cache/uv && \
519
520
521
        source ${VIRTUAL_ENV}/bin/activate && \
        uv build --wheel --out-dir /opt/dynamo/dist /opt/dynamo/lib/gpu_memory_service; \
    fi
522
{% endif %}