Pull request/1064 (#1069)

* Allow building Merlin with older versions of `docker`. * Did not copy perf_analyzer. * Comments to explain what we do there. * Properly fix the perf_analyzer issue. * `python-libnvinfer` is finally delivered properly on ARM64. * Fix HugeCTR compilation issues in TF image. --------- Co-authored-by: Matthias Langer <[email protected]>
NVIDIA-Merlin · Sep 21, 2023 · 64966c5 · 64966c5
1 parent 58ceea4
commit 64966c5
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 14 deletions.
diff --git a/docker/dockerfile.merlin b/docker/dockerfile.merlin
@@ -3,10 +3,12 @@ ARG TRITON_VERSION=23.06
 ARG DLFW_VERSION=23.06
 
 ARG FULL_IMAGE=nvcr.io/nvidia/tritonserver:${TRITON_VERSION}-py3
+ARG SDK_IMAGE=nvcr.io/nvidia/tritonserver:${TRITON_VERSION}-py3-sdk
 ARG BASE_IMAGE=nvcr.io/nvidia/tritonserver:${TRITON_VERSION}-py3-min
 ARG DLFW_IMAGE=nvcr.io/nvidia/tensorflow:${TRITON_VERSION}-tf2-py3
 
 FROM ${FULL_IMAGE} as triton
+FROM ${SDK_IMAGE} as sdk
 FROM ${DLFW_IMAGE} as dlfw
 FROM ${BASE_IMAGE} as build
 
@@ -118,8 +120,9 @@ COPY --chown=1000:1000 --from=triton /opt/tritonserver/lib lib/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/include include/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/repoagents/ repoagents/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/backends/python backends/
-# NOTE 2023-07: fil-backend is not available on ARM.
-COPY --chown=1000:1000 --from=triton /opt/tritonserver/backends/fil* backends/fil/
+# NOTE 2023-09: fil-backend is not available on ARM. Some docker versions flag an error if there is
+#               not a single source file to copy. To avoid this, we als specify a small dummy file.
+COPY --chown=1000:1000 --from=triton /opt/tritonserver/LICENSE /opt/tritonserver/backends/fil/* backends/fil/
 COPY --chown=1000:1000 --from=triton /usr/bin/serve /usr/bin/.
 
 ENV PATH=/opt/tritonserver/bin:${PATH}:
@@ -187,11 +190,12 @@ RUN ARCH=$([ "${TARGETARCH}" = "arm64" ] && echo "sbsa" || echo "x86_64") && \
         python3 \
         python3-pip \
         python3-dev \
+        python3-libnvinfer \
         rapidjson-dev \
         tree \
         wget \
         zlib1g-dev \
-        # Required to build RocksDB and RdKafka..
+        # Required to build RocksDB and RdKafka.
         libgflags-dev \
         libbz2-dev \
         libsnappy-dev \
@@ -208,11 +212,6 @@ RUN ARCH=$([ "${TARGETARCH}" = "arm64" ] && echo "sbsa" || echo "x86_64") && \
         openssh-server \
         # [ HugeCTR ]
         libaio-dev && \
-        # NOTE: libnvinfer is installed anyway, just Python bindings are missing on ARM.
-    if [[ "$TARGETARCH" != "arm64" ]]; then \
-        # TensorRT dependencies
-        apt install -y --no-install-recommends python3-libnvinfer \
-    ; fi && \
     apt autoremove -y && \
     apt clean && \
     rm -rf /var/lib/apt/lists/*
@@ -225,7 +224,7 @@ ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${JAVA_HOME}/lib:${JAVA_HOME}/lib/server
 # Binaries
 COPY --chown=1000:1000 --from=build /usr/local/bin/cmake /usr/local/bin/
 COPY --chown=1000:1000 --from=build /usr/local/bin/pytest /usr/local/bin/
-COPY --chown=1000:1000 --from=build /usr/local/bin/perf_* /usr/local/bin/
+COPY --chown=1000:1000 --from=sdk /usr/local/bin/perf_* /usr/local/bin/
 
 # Triton Server
 WORKDIR /opt/tritonserver
@@ -237,8 +236,9 @@ COPY --chown=1000:1000 --from=triton /opt/tritonserver/lib lib/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/include include/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/repoagents/ repoagents/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/backends/python backends/python/
-# NOTE 2023-07: fil-backend is not available on ARM.
-COPY --chown=1000:1000 --from=triton /opt/tritonserver/backends/fil* backends/fil/
+# NOTE 2023-09: fil-backend is not available on ARM. Some docker versions flag an error if there is
+#               not a single source file to copy. To avoid this, we als specify a small dummy file.
+COPY --chown=1000:1000 --from=triton /opt/tritonserver/LICENSE /opt/tritonserver/backends/fil/* backends/fil/
 COPY --chown=1000:1000 --from=triton /opt/tritonserver/backends/tensorrt backends/tensorrt/
 COPY --chown=1000:1000 --from=triton /usr/bin/serve /usr/bin/.
 COPY --chown=1000:1000 --from=triton /usr/lib/*-linux-gnu/libdcgm.so.2 /tmp
@@ -362,7 +362,7 @@ ENV PATH=${PATH}:${HADOOP_HOME}/bin:${HADOOP_HOME}/sbin \
     YARN_NODEMANAGER_USER=root \
     # Tackles with ThreadReaper stack overflow issues: https://bugs.openjdk.java.net/browse/JDK-8153057
     LIBHDFS_OPTS='-Djdk.lang.processReaperUseDefaultStackSize=true' \
-    # Tackles with JVM setting error signals that UCX library will check (GitLab issue #425).
+    # Tackles with JVM setting error signals that the UCX library checks (GitLab issue #425).
     UCX_ERROR_SIGNALS='' \
     CLASSPATH=${CLASSPATH}:\
 ${HADOOP_HOME}/etc/hadoop/*:\
@@ -389,7 +389,7 @@ ENV PATH=$PATH:${HUGECTR_HOME}/bin \
     LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${HUGECTR_HOME}/lib
 
 RUN if [ "${HUGECTR_DEV_MODE}" == "false" ]; then \
-        # Install HugeCTR inference which is dependency for hps_backenc
+        # Install HugeCTR inference which is dependency for hps_backend
         git clone --branch ${HUGECTR_VER} --depth 1 https://${_CI_JOB_TOKEN}${_HUGECTR_REPO} /hugectr && \
         cd /hugectr && \
         git submodule update --init --recursive && \

diff --git a/docker/dockerfile.tf b/docker/dockerfile.tf
@@ -42,7 +42,6 @@ ARG _CI_JOB_TOKEN=""
 ARG HUGECTR_VER=main
 
 ENV LD_LIBRARY_PATH=/usr/local/lib/python${PYTHON_VERSION}/dist-packages/tensorflow:$LD_LIBRARY_PATH \
-    LIBRARY_PATH=${HUGECTR_HOME}/lib:$LIBRARY_PATH \
     SOK_COMPILE_UNIT_TEST=ON
 
 RUN mkdir -p /usr/local/nvidia/lib64 && \
@@ -55,6 +54,9 @@ ARG INSTALL_DISTRIBUTED_EMBEDDINGS=false
 ARG TFDE_VER=v23.03.00
 
 RUN if [ "$HUGECTR_DEV_MODE" == "false" ]; then \
+        export HUGECTR_HOME=/usr/local/hugectr && \
+        rm -rf ${HUGECTR_HOME}/lib/libgmock* ${HUGECTR_HOME}/lib/pkgconfig/gmock* ${HUGECTR_HOME}/include/gmock && \
+        rm -rf ${HUGECTR_HOME}/lib/libgtest* ${HUGECTR_HOME}/lib/pkgconfig/gtest* ${HUGECTR_HOME}/include/gtest && \
         git clone --branch ${HUGECTR_VER} --depth 1 --recurse-submodules --shallow-submodules https://${_CI_JOB_TOKEN}${_HUGECTR_REPO} /hugectr && \
         pushd /hugectr && \
         rm -rf .git/modules && \