Update builder images to ROCm6.0 (#1647)

* Update ROCm versions for docker images

* Don't build MIOpen from source for ROCm6.0

* Temporarily use magma fork with ROCm6.0 patch

* Update ROCm versions for docker images

* Add gfx942

* Update MIOpen repo

* Magma PR 42 is merged, so use upstream repo master branch now

* gfx942 target only fully supported for ROCm6.0 and above

Author: jithunnair-amd

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

@@ -52,7 +52,7 @@ jobs:
     runs-on: linux.12xlarge
     strategy:
       matrix:
-        rocm_version: ["5.6", "5.7"]
+        rocm_version: ["5.7", "6.0"]
     env:
       GPU_ARCH_TYPE: rocm
       GPU_ARCH_VERSION: ${{ matrix.rocm_version }}

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

@@ -58,7 +58,7 @@ jobs:
     runs-on: linux.12xlarge
     strategy:
       matrix:
-        rocm_version: ["5.6", "5.7"]
+        rocm_version: ["5.7", "6.0"]
     env:
       GPU_ARCH_TYPE: rocm
       GPU_ARCH_VERSION: ${{ matrix.rocm_version }}

common/install_miopen.sh

@@ -56,7 +56,10 @@ MIOPEN_CMAKE_COMMON_FLAGS="
 -DMIOPEN_BUILD_DRIVER=OFF
 "
 # Pull MIOpen repo and set DMIOPEN_EMBED_DB based on ROCm version
-if [[ $ROCM_INT -ge 50700 ]] && [[ $ROCM_INT -lt 50800 ]]; then
+if [[ $ROCM_INT -ge 60000 ]] && [[ $ROCM_INT -lt 60100 ]]; then
+    echo "ROCm 6.0 MIOpen does not need any patches, do not build from source"
+    exit 0
+elif [[ $ROCM_INT -ge 50700 ]] && [[ $ROCM_INT -lt 60000 ]]; then
     echo "ROCm 5.7 MIOpen does not need any patches, do not build from source"
     exit 0
 elif [[ $ROCM_INT -ge 50600 ]] && [[ $ROCM_INT -lt 50700 ]]; then
@@ -85,7 +88,7 @@ fi
 
 yum remove -y miopen-hip
 
-git clone https://github.com/ROCmSoftwarePlatform/MIOpen -b ${MIOPEN_BRANCH}
+git clone https://github.com/ROCm/MIOpen -b ${MIOPEN_BRANCH}
 pushd MIOpen
 # remove .git to save disk space since CI runner was running out
 rm -rf .git

common/install_rocm_magma.sh

@@ -15,7 +15,7 @@ pushd magma
 if [[ $PYTORCH_BRANCH == "release/1.10.1" ]]; then
     git checkout magma_ctrl_launch_bounds
 else
-    git checkout 28592a7170e4b3707ed92644bf4a689ed600c27f
+    git checkout a1625ff4d9bc362906bd01f805dbbe12612953f6
 fi
 cp make.inc-examples/make.inc.hip-gcc-mkl make.inc
 echo 'LIBDIR += -L$(MKLROOT)/lib' >> make.inc

libtorch/build_all_docker.sh

@@ -8,6 +8,6 @@ for cuda_version in 12.1 11.8; do
     GPU_ARCH_TYPE=cuda GPU_ARCH_VERSION="${cuda_version}" "${TOPDIR}/libtorch/build_docker.sh"
 done
 
-for rocm_version in 5.6 5.7; do
+for rocm_version in 5.7 6.0; do
     GPU_ARCH_TYPE=rocm GPU_ARCH_VERSION="${rocm_version}" "${TOPDIR}/libtorch/build_docker.sh"
 done

libtorch/build_docker.sh

@@ -36,6 +36,9 @@ case ${GPU_ARCH_TYPE} in
             echo "ERROR: rocm regex failed"
             exit 1
         fi
+        if [[ $ROCM_VERSION_INT -ge 60000 ]]; then
+            PYTORCH_ROCM_ARCH+=";gfx942"
+        fi
         DOCKER_GPU_BUILD_ARG="--build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}"
         ;;
     *)

manywheel/build_all_docker.sh

@@ -16,7 +16,7 @@ for cuda_version in 12.1 11.8; do
     MANYLINUX_VERSION=2014 GPU_ARCH_TYPE=cuda GPU_ARCH_VERSION="${cuda_version}" "${TOPDIR}/manywheel/build_docker.sh"
 done
 
-for rocm_version in 5.6 5.7; do
+for rocm_version in 5.7 6.0; 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

manywheel/build_docker.sh

@@ -57,6 +57,9 @@ case ${GPU_ARCH_TYPE} in
             echo "ERROR: rocm regex failed"
             exit 1
         fi
+        if [[ $ROCM_VERSION_INT -ge 60000 ]]; then
+            PYTORCH_ROCM_ARCH+=";gfx942"
+        fi
         DOCKER_GPU_BUILD_ARG="--build-arg ROCM_VERSION=${GPU_ARCH_VERSION} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg DEVTOOLSET_VERSION=9"
         ;;
     *)