ROCm5.3 nightly wheels (#1193)

* Enable ROCm5.3 nightly wheels

* Enable ROCm5.3 docker builds

* Update amdgpu repo url for ROCm5.3

* ROCm5.3 not supported on Ubuntu 18.04

* empty

* Another empty commit

* Try disabling MLIR build to shorten docker build time

* Clean up disk space

* MLIR project changed names from ROCm5.4

* Retrigger CI to get around flaky magma git access error

Author: jithunnair-amd

.github/workflows/build-libtorch-images.yml

@@ -47,7 +47,7 @@ jobs:
     runs-on: ubuntu-18.04
     strategy:
       matrix:
-        rocm_version: ["5.1.1", "5.2"]
+        rocm_version: ["5.2", "5.3"]
     env:
       GPU_ARCH_TYPE: rocm
       GPU_ARCH_VERSION: ${{ matrix.rocm_version }}

.github/workflows/build-manywheel-images.yml

@@ -49,7 +49,7 @@ jobs:
     runs-on: ubuntu-18.04
     strategy:
       matrix:
-        rocm_version: ["5.1.1", "5.2"]
+        rocm_version: ["5.2", "5.3"]
     env:
       GPU_ARCH_TYPE: rocm
       GPU_ARCH_VERSION: ${{ matrix.rocm_version }}

common/install_miopen.sh

@@ -88,6 +88,13 @@ fi
 
 git clone https://github.com/ROCmSoftwarePlatform/MIOpen -b ${MIOPEN_BRANCH}
 pushd MIOpen
+# Don't build MLIR to save docker build time 
+# since we are disabling MLIR backend for MIOpen anyway
+if [[ $ROCM_INT -ge 50400 ]] && [[ $ROCM_INT -lt 50500 ]]; then
+    sed -i '/rocMLIR/d' requirements.txt
+elif [[ $ROCM_INT -ge 50200 ]] && [[ $ROCM_INT -lt 50400 ]]; then
+    sed -i '/llvm-project-mlir/d' requirements.txt
+fi
 ## MIOpen minimum requirements
 cmake -P install_deps.cmake --minimum
 ## Build MIOpen
@@ -99,12 +106,15 @@ PKG_CONFIG_PATH=/usr/local/lib/pkgconfig CXX=${ROCM_INSTALL_PATH}/llvm/bin/clang
     -DCMAKE_PREFIX_PATH="${ROCM_INSTALL_PATH}/hip;${ROCM_INSTALL_PATH}"
 make MIOpen -j $(nproc)
 make -j $(nproc) package
-yum install -y miopen-*.rpm
-popd
-rm -rf MIOpen
 
-# Cleanup
+# clean up since CI runner was running out of disk space
+rm -rf /usr/local/cget
+rm -rf /tmp/*
 yum clean all
 rm -rf /var/cache/yum
 rm -rf /var/lib/yum/yumdb
 rm -rf /var/lib/yum/history
+
+yum install -y miopen-*.rpm
+popd
+rm -rf MIOpen

common/install_rocm.sh

@@ -33,7 +33,12 @@ install_ubuntu() {
     if [[ $(ver $ROCM_VERSION) -ge $(ver 4.5) ]]; then
         # Add amdgpu repository
         UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
-        local amdgpu_baseurl="https://repo.radeon.com/amdgpu/${AMDGPU_VERSIONS[$ROCM_VERSION]}/ubuntu"
+        local amdgpu_baseurl
+        if [[ $(ver $ROCM_VERSION) -ge $(ver 5.3) ]]; then
+          amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
+        else
+          amdgpu_baseurl="https://repo.radeon.com/amdgpu/${AMDGPU_VERSIONS[$ROCM_VERSION]}/ubuntu"
+        fi
         echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
     fi
 
@@ -83,7 +88,12 @@ install_centos() {
 
   if [[ $(ver $ROCM_VERSION) -ge $(ver 4.5) ]]; then
       # Add amdgpu repository
-      local amdgpu_baseurl="https://repo.radeon.com/amdgpu/${AMDGPU_VERSIONS[$ROCM_VERSION]}/rhel/7.9/main/x86_64"
+      local amdgpu_baseurl
+      if [[ $(ver $ROCM_VERSION) -ge $(ver 5.3) ]]; then
+        amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/rhel/7.9/main/x86_64"
+      else
+        amdgpu_baseurl="https://repo.radeon.com/amdgpu/${AMDGPU_VERSIONS[$ROCM_VERSION]}/rhel/7.9/main/x86_64"
+      fi
       echo "[AMDGPU]" > /etc/yum.repos.d/amdgpu.repo
       echo "name=AMDGPU" >> /etc/yum.repos.d/amdgpu.repo
       echo "baseurl=${amdgpu_baseurl}" >> /etc/yum.repos.d/amdgpu.repo

libtorch/Dockerfile

@@ -69,13 +69,13 @@ RUN apt-get update -y && \
     apt-get install python -y && \
     apt-get clean
 
-FROM rocm as rocm5.1.1
-RUN ROCM_VERSION=5.1.1 bash ./install_rocm.sh && rm install_rocm.sh
+FROM rocm as rocm5.2
+RUN ROCM_VERSION=5.2 bash ./install_rocm.sh && rm install_rocm.sh
 RUN bash ./install_rocm_drm.sh && rm install_rocm_drm.sh
 RUN bash ./install_rocm_magma.sh && rm install_rocm_magma.sh
 
-FROM rocm as rocm5.2
-RUN ROCM_VERSION=5.2 bash ./install_rocm.sh && rm install_rocm.sh
+FROM rocm as rocm5.3
+RUN ROCM_VERSION=5.3 bash ./install_rocm.sh && rm install_rocm.sh
 RUN bash ./install_rocm_drm.sh && rm install_rocm_drm.sh
 RUN bash ./install_rocm_magma.sh && rm install_rocm_magma.sh
 

libtorch/build_all_docker.sh

@@ -8,6 +8,6 @@ for cuda_version in 11.8 11.7 11.6; do
     GPU_ARCH_TYPE=cuda GPU_ARCH_VERSION="${cuda_version}" "${TOPDIR}/libtorch/build_docker.sh"
 done
 
-for rocm_version in 5.1.1 5.2; do
+for rocm_version in 5.2 5.3; do
     GPU_ARCH_TYPE=rocm GPU_ARCH_VERSION="${rocm_version}" "${TOPDIR}/libtorch/build_docker.sh"
 done

libtorch/build_docker.sh

@@ -27,7 +27,7 @@ case ${GPU_ARCH_TYPE} in
     rocm)
         BASE_TARGET=rocm${GPU_ARCH_VERSION}
         DOCKER_TAG=rocm${GPU_ARCH_VERSION}
-        GPU_IMAGE=rocm/dev-ubuntu-18.04:${GPU_ARCH_VERSION}
+        GPU_IMAGE=rocm/dev-ubuntu-20.04:${GPU_ARCH_VERSION}
         PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908"
         ROCM_REGEX="([0-9]+)\.([0-9]+)[\.]?([0-9]*)"
         if [[ $GPU_ARCH_VERSION =~ $ROCM_REGEX ]]; then

manywheel/build_all_docker.sh

@@ -14,7 +14,7 @@ for cuda_version in 11.7 11.6; do
     MANYLINUX_VERSION=2014 GPU_ARCH_TYPE=cuda GPU_ARCH_VERSION="${cuda_version}" "${TOPDIR}/manywheel/build_docker.sh"
 done
 
-for rocm_version in 5.1.1 5.2; do
+for rocm_version in 5.2 5.3; do
     GPU_ARCH_TYPE=rocm GPU_ARCH_VERSION="${rocm_version}" "${TOPDIR}/manywheel/build_docker.sh"
     MANYLINUX_VERSION=2014 GPU_ARCH_TYPE=rocm GPU_ARCH_VERSION="${rocm_version}" "${TOPDIR}/manywheel/build_docker.sh"
 done