diff --git a/.circleci/config.yml b/.circleci/config.yml
index ebfd4302..b1778fca 100644
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -195,10 +195,13 @@ workflows:
                 - tf28-gpu
                 - pt-gpu
                 - pt2-gpu
-                - pytorch10-tf27-rocm50
+                - pytorch13-tf210-rocm56
+                - pytorch20-tf210-rocm56
             exclude:
               - with-mpi: 1
-                image-type: pytorch10-tf27-rocm50
+                image-type: 
+                  - pytorch13-tf210-rocm56
+                  - pytorch20-tf210-rocm56
       - build-and-publish-docker:
           name: build-and-publish-docker-<<matrix.image-type>>-<<matrix.with-mpi>>
           context: determined-production
@@ -251,17 +254,20 @@ workflows:
               image-type:
                 - tf2-cpu
                 - tf28-cpu
-                - pt-cpu
-                - pt2-cpu
+                - pytorch-cpu
+                - pytorch2-cpu
                 - tf2-gpu
                 - tf28-gpu
-                - pt-gpu
-                - pt2-gpu
-                - pytorch10-tf27-rocm50
+                - pytorch-gpu
+                - pytorch2-gpu
+                - pytorch13-tf210-rocm56
+                - pytorch20-tf210-rocm56
             exclude:
               - dev-mode: true
                 with-mpi: 1
-                image-type: pytorch10-tf27-rocm50
+                image-type: 
+                  - pytorch13-tf210-rocm56
+                  - pytorch20-tf210-rocm56
 
       - build-and-publish-docker:
           name: build-and-publish-docker-<<matrix.image-type>>-<<matrix.with-mpi>>-dev
@@ -281,8 +287,8 @@ workflows:
                 - gpu.nvidia.small.multi
               with-mpi: [0]
               image-type:
-                - deepspeed-gpu
-                - gpt-neox-deepspeed-gpu
+                - deepspeed
+                - gpt-neox-deepspeed
 
       - publish-cloud-images:
           name: publish-cloud-images-dev
diff --git a/Dockerfile-default-rocm b/Dockerfile-default-rocm
index 2e4071b8..3e6a0a94 100644
--- a/Dockerfile-default-rocm
+++ b/Dockerfile-default-rocm
@@ -1,40 +1,117 @@
 ARG BASE_IMAGE
 FROM ${BASE_IMAGE}
 
+RUN apt remove -y openmpi ucx
+#Let's remove existing /opt/ompi; and, link to our version.
+RUN rm -rf /opt/ompi 
+RUN ln -s /container/ompi /opt 
+
 RUN mkdir -p /var/run/sshd
 RUN rm /etc/apt/sources.list.d/rocm.list
 RUN apt-get update \
-	&& DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
-		autoconf \
-		automake \
-		autotools-dev \
-		build-essential \
-		ca-certificates \
-		curl \
-		daemontools \
-		libkrb5-dev \
-		libssl-dev \
-		libtool \
-		git \
-		krb5-user \
-		cmake \
-		g++-4.8 \
-		make \
-		openssh-client \
-		openssh-server \
-		pkg-config \
-		wget \
-		nfs-common \
-		unattended-upgrades \
-	&& unattended-upgrade \
-	&& rm -rf /var/lib/apt/lists/* \
-	&& rm /etc/ssh/ssh_host_ecdsa_key \
-	&& rm /etc/ssh/ssh_host_ed25519_key \
-	&& rm /etc/ssh/ssh_host_rsa_key
+        && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+                autoconf \
+                automake \
+                autotools-dev \
+                build-essential \
+                ca-certificates \
+                curl \
+                daemontools \
+                glibc-source \
+                ibverbs-providers \
+                libibverbs1 \
+                libkrb5-dev \
+                librdmacm1 \
+                libssl-dev \
+                libtool \
+                git \
+                krb5-user \
+                cmake \
+                g++ \
+                make \
+                openssh-client \
+                openssh-server \
+                pkg-config \
+                wget \
+                nfs-common \
+                libnuma1 \
+                libnuma-dev \
+                libpmi2-0-dev \
+                unattended-upgrades \
+        && unattended-upgrade \
+        && rm -rf /var/lib/apt/lists/* \
+        && rm /etc/ssh/ssh_host_ecdsa_key \
+        && rm /etc/ssh/ssh_host_ed25519_key \
+        && rm /etc/ssh/ssh_host_rsa_key
+RUN pip install pip install --upgrade pip
 
 COPY dockerfile_scripts /tmp/det_dockerfile_scripts
 
+ENV PATH="/opt/conda/envs/py_3.8/bin:${PATH}"
+
+ARG CONDA="${PATH}"
+
+ENV PYTHONUNBUFFERED=1 PYTHONFAULTHANDLER=1 PYTHONHASHSEED=0
+
+# Install fixed version of FFI package for Ubuntu 20.04.
+# This is done after above stuff to make sure we get right version.
+RUN /tmp/det_dockerfile_scripts/install_package_fixes.sh
+
+RUN apt install rocm-libs 
+
+#USING OFI
+ARG WITH_MPI=1
+ARG WITH_OFI=1
+ARG WITH_MPICH
+ARG UCX_INSTALL_DIR=/container/ucx
+ARG OMPI_INSTALL_DIR=/container/ompi
+ARG MPICH_INSTALL_DIR=/container/mpich
+ARG OFI_INSTALL_DIR=/container/ofi
+ARG OMPI_WITH_CUDA=0
+ARG OMPI_WITH_ROCM=1
+RUN if [ "$WITH_MPI" = "1" ]; then /tmp/det_dockerfile_scripts/ompi_rocm.sh "$UBUNTU_VERSION" "$WITH_OFI" "$OMPI_WITH_ROCM" "$WITH_MPICH"; fi
+
+# Make sure OMPI/UCX show up in the right paths
+ARG VERBS_LIB_DIR=/usr/lib/libibverbs
+ARG UCX_LIB_DIR=${UCX_INSTALL_DIR}/lib:${UCX_INSTALL_DIR}/lib64
+ARG UCX_PATH_DIR=${UCX_INSTALL_DIR}/bin
+ARG OFI_LIB_DIR=${OFI_INSTALL_DIR}/lib:${OFI_INSTALL_DIR}/lib64
+ARG OFI_PATH_DIR=${OFI_INSTALL_DIR}/bin
+ARG OMPI_LIB_DIR=${OMPI_INSTALL_DIR}/lib
+ARG OMPI_PATH_DIR=${OMPI_INSTALL_DIR}/bin
+ARG MPICH_LIB_DIR=${MPICH_INSTALL_DIR}/lib
+ARG MPICH_PATH_DIR=${MPICH_INSTALL_DIR}/bin
+
+# Set up UCX_LIBS and OFI_LIBS
+ENV UCX_LIBS="${VERBS_LIB_DIR}:${UCX_LIB_DIR}:${OMPI_LIB_DIR}:"
+ENV OFI_LIBS="${VERBS_LIB_DIR}:${OFI_LIB_DIR}:${MPICH_LIB_DIR}:"
+
+# If WITH_OFI is true, then set EXTRA_LIBS to OFI libs, else set to empty string
+ENV EXTRA_LIBS="${WITH_OFI:+${OFI_LIBS}}"
+
+# If EXTRA_LIBS is empty, set to UCX libs, else leave as OFI libs
+ENV EXTRA_LIBS="${EXTRA_LIBS:-${UCX_LIBS}}"
+
+# But, only add them if WITH_MPI
+ENV LD_LIBRARY_PATH=${WITH_MPI:+$EXTRA_LIBS}$LD_LIBRARY_PATH
+
+#USING OFI
+ENV PATH=${WITH_OFI:+$PATH:${WITH_MPI:+$OFI_PATH_DIR:$MPICH_PATH_DIR}}
+
+#USING UCX
+ENV PATH=${PATH:-$CONDA:${WITH_MPI:+$UCX_PATH_DIR:$OMPI_PATH_DIR}}
+
+# Enable running OMPI as root
+ENV OMPI_ALLOW_RUN_AS_ROOT ${WITH_MPI:+1}
+ENV OMPI_ALLOW_RUN_AS_ROOT_CONFIRM ${WITH_MPI:+1}
+
+
+
+RUN pip install cloudpickle
 RUN pip install determined && pip uninstall -y determined
+RUN pip install google-auth-oauthlib
+
+
 
 RUN pip install -r /tmp/det_dockerfile_scripts/notebook-requirements.txt
 ENV JUPYTER_CONFIG_DIR=/run/determined/jupyter/config
@@ -48,7 +125,6 @@ RUN /tmp/det_dockerfile_scripts/install_google_cloud_sdk.sh
 # google-api-python-client -> google-api-core -> googleapis-common-protos -> protobuf
 # Horovod cannot build with protobuf > 3.20.x
 # latest google-api-python-client requires protobuf >= 3.20.1
-RUN pip install protobuf==3.20.1
 
 ARG TENSORFLOW_PIP
 RUN if [ "$TENSORFLOW_PIP" ]; then pip install $TENSORFLOW_PIP; fi
@@ -69,13 +145,58 @@ ARG HOROVOD_WITH_TENSORFLOW=1
 ARG HOROVOD_WITH_PYTORCH=1
 ARG HOROVOD_WITHOUT_MXNET=1
 ARG HOROVOD_GPU_OPERATIONS=NCCL
-ARG HOROVOD_WITHOUT_MPI=1
+ARG HOROVOD_WITHOUT_MPI=0
+ARG HOROVOD_WITH_MPI=1
 ARG HOROVOD_GPU=ROCM
-ARG HOROVOD_WITHOUT_MPI=1
-ENV LD_LIBRARY_PATH=/opt/rocm/lib:/opt/rocm/hip/lib
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH;/opt/rocm/lib:/opt/rocm/hip/lib
+
+ENV HOROVOD_PIP $HOROVOD_PIP
+ENV HOROVOD_WITH_TENSORFLOW $HOROVOD_WITH_TENSORFLOW
+ENV HOROVOD_WITH_PYTORCH $HOROVOD_WITH_PYTORCH
+ENV HOROVOD_WITHOUT_MXNET $HOROVOD_WITHOUT_MXNET
+ENV HOROVOD_GPU_OPERATIONS $HOROVOD_GPU_OPERATIONS
+ENV HOROVOD_WITHOUT_MPI $HOROVOD_WITHOUT_MPI
+ENV HOROVOD_WITH_MPI $HOROVOD_WITH_MPI
+ENV HOROVOD_GPU $HOROVOD_GPU
+ENV HOROVOD_NCCL_HOME $HOROVOD_NCCL_HOME
+ENV NCCL_LIB_DIR=${HOROVOD_NCCL_HOME}/lib
+ENV HOROVOD_NCCL_LINK=${WITH_OFI:+SHARED}
+ENV LD_LIBRARY_PATH=${WITH_OFI:+$NCCL_LIB_DIR:}$LD_LIBRARY_PATH
+
 RUN if [ "$HOROVOD_PIP" != "0" ]; then pip install "${HOROVOD_PIP}" ; fi
 
-RUN rm -r /tmp/*
+RUN pip uninstall -y tb-nightly tensorboardX
+RUN pip install -r /tmp/det_dockerfile_scripts/additional-requirements-rocm.txt
+
+
+ENV HSA_FORCE_FINE_GRAIN_PCIE=1
+
+ARG AWS_PLUGIN_INSTALL_DIR=/container/aws
+ARG WITH_AWS_TRACE
+ARG INTERNAL_AWS_DS
+ARG INTERNAL_AWS_PATH
+ARG ROCM_DIR=/opt/rocm
+ENV ROCM_DIR $ROCM_DIR
+RUN if [ "$WITH_OFI" = "1" ]; then /tmp/det_dockerfile_scripts/build_aws_rocm.sh "$WITH_OFI" "$WITH_AWS_TRACE" "$WITH_MPICH"; fi
+ENV LD_LIBRARY_PATH=${WITH_OFI:+$AWS_PLUGIN_INSTALL_DIR:}$LD_LIBRARY_PATH
+RUN ldconfig
+
+ENV PATH=$OMPI_PATH_DIR:$OFI_INSTALL_DIR:$PATH
 # Reset entrypoint.
-ENTRYPOINT []
+
+# Set an entrypoint that can scrape up the host libfabric.so and then 
+# run the user command. This is intended to enable performant execution
+# on non-IB systems that have a proprietary libfabric.
+
+RUN mkdir -p /container/bin && cp /tmp/det_dockerfile_scripts/scrape_libs.sh /container/bin
+
+ARG WITH_RCCL=1
+ENV WITH_RCCL=$WITH_RCCL
+ARG WITH_NFS_WORKAROUND=1
+ENV WITH_NFS_WORKAROUND=$WITH_NFS_WORKAROUND
+
+ENTRYPOINT ["/container/bin/scrape_libs.sh"]
+CMD ["/bin/bash"]
 USER root
+
+RUN rm -r /tmp/*
diff --git a/Makefile b/Makefile
index 237869ba..a7ff5164 100644
--- a/Makefile
+++ b/Makefile
@@ -1,6 +1,4 @@
 SHELL := /bin/bash -o pipefail
-VERSION := $(shell cat VERSION)
-VERSION_DASHES := $(subst .,-,$(VERSION))
 SHORT_GIT_HASH := $(shell git rev-parse --short HEAD)
 
 NGC_REGISTRY := nvcr.io/isv-ngc-partner/determined
@@ -15,7 +13,7 @@ CUDA_111_PREFIX := $(REGISTRY_REPO):cuda-11.1-
 CUDA_112_PREFIX := $(REGISTRY_REPO):cuda-11.2-
 CUDA_113_PREFIX := $(REGISTRY_REPO):cuda-11.3-
 CUDA_118_PREFIX := $(REGISTRY_REPO):cuda-11.8-
-ROCM_50_PREFIX := $(REGISTRY_REPO):rocm-5.0-
+ROCM_56_PREFIX := $(REGISTRY_REPO):rocm-5.6-
 
 CPU_SUFFIX := -cpu
 GPU_SUFFIX := -gpu
@@ -91,7 +89,6 @@ build-cpu-py-38-base:
 		--build-arg "$(MPI_BUILD_ARG)" \
 		--build-arg "$(OFI_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME)-$(VERSION) \
 		--push \
 		.
 
@@ -108,7 +105,6 @@ build-cpu-py-39-base:
 		--build-arg "$(MPI_BUILD_ARG)" \
 		--build-arg "$(OFI_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(CPU_PY_39_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(CPU_PY_39_BASE_NAME)-$(VERSION) \
 		--push \
 		.
 
@@ -124,7 +120,6 @@ build-cpu-py-310-base:
 		--build-arg "$(MPI_BUILD_ARG)" \
 		--build-arg "$(OFI_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(CPU_PY_310_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(CPU_PY_310_BASE_NAME)-$(VERSION) \
 		--push \
 		.
 
@@ -136,7 +131,6 @@ build-gpu-cuda-111-base:
 		--build-arg UBUNTU_VERSION="$(UBUNTU_VERSION)" \
 		--build-arg "$(MPI_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_111_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_111_BASE_NAME)-$(VERSION) \
 		.
 
 .PHONY: build-gpu-cuda-112-base
@@ -147,7 +141,6 @@ build-gpu-cuda-112-base:
 		--build-arg UBUNTU_VERSION="$(UBUNTU_VERSION)" \
 		--build-arg "$(MPI_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_112_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_112_BASE_NAME)-$(VERSION) \
 		.
 
 .PHONY: build-gpu-cuda-113-base
@@ -161,7 +154,6 @@ build-gpu-cuda-113-base:
 		--build-arg "$(OFI_BUILD_ARG)" \
 		--build-arg "$(NCCL_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME)-$(VERSION) \
 		.
 
 .PHONY: build-gpu-cuda-118-base
@@ -175,33 +167,50 @@ build-gpu-cuda-118-base:
 		--build-arg "$(OFI_BUILD_ARG)" \
 		--build-arg "$(NCCL_BUILD_ARG)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_118_BASE_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_118_BASE_NAME)-$(VERSION) \
 		.
 
-export ROCM50_TORCH_TF_ENVIRONMENT_NAME := $(ROCM_50_PREFIX)pytorch-1.10-tf-2.7-rocm
-export TF_PROFILER_PIP := tensorboard-plugin-profile
-export TORCH_TB_PROFILER_PIP := torch-tb-profiler==0.4.1
+ifeq ($(WITH_MPICH),1)
+ROCM56_TORCH13_MPI :=pytorch-1.3-tf-2.10-rocm-mpich
+else
+ROCM56_TORCH13_MPI :=pytorch-1.3-tf-2.10-rocm-ompi
+endif
+export ROCM56_TORCH13_TF_ENVIRONMENT_NAME := $(ROCM_56_PREFIX)$(ROCM56_TORCH13_MPI)
+.PHONY: build-pytorch13-tf210-rocm56
+build-pytorch13-tf210-rocm56:
+	docker build -f Dockerfile-default-rocm \
+		--build-arg BASE_IMAGE="rocm/pytorch:rocm5.6_ubuntu20.04_py3.8_pytorch_1.13.1"\
+		--build-arg TENSORFLOW_PIP="tensorflow-rocm==2.10.1.540" \
+		--build-arg HOROVOD_PIP="horovod==0.28.1" \
+		--build-arg WITH_MPICH=$(WITH_MPICH) \
+		-t $(DOCKERHUB_REGISTRY)/$(ROCM56_TORCH13_TF_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
+		-t $(DOCKERHUB_REGISTRY)/$(ROCM56_TORCH13_TF_ENVIRONMENT_NAME)-$(VERSION) \
+		.
 
-.PHONY: build-pytorch10-tf27-rocm50
-build-pytorch10-tf27-rocm50:
+ifeq ($(WITH_MPICH),1)
+ROCM56_TORCH_MPI :=pytorch-2.0-tf-2.10-rocm-mpich
+else
+ROCM56_TORCH_MPI :=pytorch-2.0-tf-2.10-rocm-ompi
+endif
+export ROCM56_TORCH_TF_ENVIRONMENT_NAME := $(ROCM_56_PREFIX)$(ROCM56_TORCH_MPI)
+.PHONY: build-pytorch20-tf210-rocm56
+build-pytorch20-tf210-rocm56:
 	docker build -f Dockerfile-default-rocm \
-		--build-arg BASE_IMAGE="amdih/pytorch:rocm5.0_ubuntu18.04_py3.7_pytorch_1.10.0" \
-		--build-arg TORCH_TB_PROFILER_PIP="$(TORCH_TB_PROFILER_PIP)" \
-		--build-arg TENSORFLOW_PIP="tensorflow-rocm==2.7.1" \
-		--build-arg TF_PROFILER_PIP="$(TF_PROFILER_PIP)" \
-		--build-arg HOROVOD_PIP="horovod==0.25.0" \
-		-t $(DOCKERHUB_REGISTRY)/$(ROCM50_TORCH_TF_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(ROCM50_TORCH_TF_ENVIRONMENT_NAME)-$(VERSION) \
+		--build-arg BASE_IMAGE="rocm/pytorch:rocm5.6_ubuntu20.04_py3.8_pytorch_2.0.1" \
+		--build-arg TENSORFLOW_PIP="tensorflow-rocm==2.10.1.540" \
+		--build-arg HOROVOD_PIP="horovod==0.28.1" \
+                --build-arg WITH_MPICH=$(WITH_MPICH) \
+		-t $(DOCKERHUB_REGISTRY)/$(ROCM56_TORCH_TF_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
+		-t $(DOCKERHUB_REGISTRY)/$(ROCM56_TORCH_TF_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
 DEEPSPEED_VERSION := 0.8.3
-export GPU_DEEPSPEED_ENVIRONMENT_NAME := $(CUDA_113_PREFIX)pytorch-1.10-deepspeed-$(DEEPSPEED_VERSION)$(GPU_SUFFIX)
-export GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME := $(CUDA_113_PREFIX)pytorch-1.10-gpt-neox-deepspeed$(GPU_SUFFIX)
+export GPU_DEEPSPEED_ENVIRONMENT_NAME := deepspeed
+export GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME := gpt-neox-deepspeed
 export TORCH_PIP_DEEPSPEED_GPU := torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio==0.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
 
 # This builds deepspeed environment off of upstream microsoft/DeepSpeed.
-.PHONY: build-deepspeed-gpu
-build-deepspeed-gpu: build-gpu-cuda-113-base
+.PHONY: build-deepspeed
+build-deepspeed: build-gpu-cuda-113-base
 	docker build -f Dockerfile-default-gpu \
 		--build-arg BASE_IMAGE="$(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME)-$(SHORT_GIT_HASH)" \
 		--build-arg TORCH_PIP="$(TORCH_PIP_DEEPSPEED_GPU)" \
@@ -211,15 +220,13 @@ build-deepspeed-gpu: build-gpu-cuda-113-base
 		--build-arg "$(NCCL_BUILD_ARG)" \
 		--build-arg DEEPSPEED_PIP="deepspeed==$(DEEPSPEED_VERSION)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME)-$(VERSION) \
 		-t $(NGC_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(NGC_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
 # This builds deepspeed environment off of a patched version of EleutherAI's fork of DeepSpeed
 # that we need for gpt-neox support.
-.PHONY: build-gpt-neox-deepspeed-gpu
-build-gpt-neox-deepspeed-gpu: build-gpu-cuda-113-base
+.PHONY: build-gpt-neox-deepspeed
+build-gpt-neox-deepspeed: build-gpu-cuda-113-base
 	docker build -f Dockerfile-default-gpu \
 		--build-arg BASE_IMAGE="$(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME)-$(SHORT_GIT_HASH)" \
 		--build-arg TORCH_PIP="$(TORCH_PIP_DEEPSPEED_GPU)" \
@@ -229,27 +236,22 @@ build-gpt-neox-deepspeed-gpu: build-gpu-cuda-113-base
 		--build-arg "$(NCCL_BUILD_ARG)" \
 		--build-arg DEEPSPEED_PIP="git+https://github.com/determined-ai/deepspeed.git@eleuther_dai" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME)-$(VERSION) \
 		-t $(NGC_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(NGC_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
 ifeq ($(NGC_PUBLISH),)
 define CPU_TF28_TAGS
--t $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(VERSION)
+-t $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 else
 define CPU_TF28_TAGS
 -t $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(VERSION) \
--t $(NGC_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(NGC_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(VERSION)
+-t $(NGC_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 endif
 
-export CPU_TF28_ENVIRONMENT_NAME := $(CPU_PREFIX_38)tf-2.8$(CPU_SUFFIX)
-export GPU_TF28_ENVIRONMENT_NAME := $(CUDA_112_PREFIX)tf-2.8$(GPU_SUFFIX)
+export CPU_TF28_ENVIRONMENT_NAME := tf28$(CPU_SUFFIX)
+export GPU_TF28_ENVIRONMENT_NAME := tf28$(GPU_SUFFIX)
 
 .PHONY: build-tf28-cpu
 build-tf28-cpu: build-cpu-py-38-base
@@ -276,9 +278,7 @@ build-tf28-gpu: build-gpu-cuda-112-base
 		--build-arg HOROVOD_PIP="horovod==0.24.2" \
 		--build-arg HOROVOD_WITH_PYTORCH=0 \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME)-$(VERSION) \
 		-t $(NGC_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(NGC_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
 TORCH_VERSION := 1.12
@@ -290,32 +290,26 @@ TORCH_PIP_CPU := torch==1.12.0+cpu torchvision==0.13.0+cpu torchaudio==0.12.0+cp
 TORCH_PIP_GPU := torch==1.12.0+cu113 torchvision==0.13.0+cu113 torchaudio==0.12.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
 HOROVOD_PIP_COMMAND := horovod==0.28.1
 
-export CPU_TF2_ENVIRONMENT_NAME := $(CPU_PREFIX_39)pytorch-$(TORCH_VERSION)-tf-$(TF2_VERSION_SHORT)$(CPU_SUFFIX)
-export GPU_TF2_ENVIRONMENT_NAME := $(CUDA_113_PREFIX)pytorch-$(TORCH_VERSION)-tf-$(TF2_VERSION_SHORT)$(GPU_SUFFIX)
-export CPU_PT_ENVIRONMENT_NAME := $(CPU_PREFIX_39)pytorch-$(TORCH_VERSION)$(CPU_SUFFIX)
-export GPU_PT_ENVIRONMENT_NAME := $(CUDA_113_PREFIX)pytorch-$(TORCH_VERSION)$(GPU_SUFFIX)
+export CPU_TF2_ENVIRONMENT_NAME := tf2$(CPU_SUFFIX)
+export GPU_TF2_ENVIRONMENT_NAME := tf2$(GPU_SUFFIX)
+export CPU_PT_ENVIRONMENT_NAME := pytorch$(CPU_SUFFIX)
+export GPU_PT_ENVIRONMENT_NAME := pytorch$(GPU_SUFFIX)
 
 ifeq ($(NGC_PUBLISH),)
 define CPU_TF2_TAGS
--t $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(VERSION)
+-t $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 define CPU_PT_TAGS
--t $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(VERSION)
+-t $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 else
 define CPU_TF2_TAGS
 -t $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(VERSION) \
--t $(NGC_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(NGC_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(VERSION)
+-t $(NGC_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 define CPU_PT_TAGS
 -t $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(VERSION) \
--t $(NGC_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(NGC_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(VERSION)
+-t $(NGC_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 endif
 
@@ -336,7 +330,7 @@ build-tf2-cpu: build-cpu-py-39-base
 		--push \
 		.
 
-.PHONY: build-pt-cpu
+.PHONY: build-pytorch-cpu
 build-pt-cpu: build-cpu-py-39-base
 	docker buildx build -f Dockerfile-default-cpu \
 	    --platform "$(PLATFORMS)" \
@@ -373,13 +367,11 @@ build-tf2-gpu: build-gpu-cuda-113-base
 		--build-arg HOROVOD_GPU_OPERATIONS="$(HOROVOD_GPU_OPERATIONS)" \
 		--build-arg HOROVOD_GPU_ALLREDUCE="$(HOROVOD_GPU_ALLREDUCE)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME)-$(VERSION) \
 		-t $(NGC_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(NGC_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
-.PHONY: build-pt-gpu
-build-pt-gpu: build-gpu-cuda-113-base
+.PHONY: build-pytorch-gpu
+build-pytorch-gpu: build-gpu-cuda-113-base
 	docker build -f Dockerfile-default-gpu \
 		--build-arg BASE_IMAGE="$(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME)-$(SHORT_GIT_HASH)" \
 		--build-arg TORCH_PIP="$(TORCH_PIP_GPU)" \
@@ -392,9 +384,7 @@ build-pt-gpu: build-gpu-cuda-113-base
 		--build-arg HOROVOD_WITHOUT_MPI="$(HOROVOD_WITHOUT_MPI)" \
 		--build-arg HOROVOD_CPU_OPERATIONS="$(HOROVOD_CPU_OPERATIONS)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME)-$(VERSION) \
 		-t $(NGC_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(NGC_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
 # torch 2.0 recipes
@@ -402,25 +392,22 @@ TORCH2_VERSION := 2.0
 TORCH2_PIP_CPU := torch==2.0.1+cpu torchvision==0.15.2+cpu torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cpu
 TORCH2_PIP_GPU := torch==2.0.1+cu118 torchvision==0.15.2+cu118 torchaudio==2.0.2+cu118 --index-url https://download.pytorch.org/whl/cu118
 TORCH2_APEX_GIT_URL := https://github.com/determined-ai/apex.git@50ac8425403b98147cbb66aea9a2a27dd3fe7673
-export CPU_PT2_ENVIRONMENT_NAME := $(CPU_PREFIX_310)pytorch-$(TORCH2_VERSION)$(CPU_SUFFIX)
-export GPU_PT2_ENVIRONMENT_NAME := $(CUDA_118_PREFIX)pytorch-$(TORCH2_VERSION)$(GPU_SUFFIX)
+export CPU_PT2_ENVIRONMENT_NAME := pytorch2$(CPU_SUFFIX)
+export GPU_PT2_ENVIRONMENT_NAME := pytorch2$(GPU_SUFFIX)
 
 ifeq ($(NGC_PUBLISH),)
 define CPU_PT2_TAGS
--t $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(VERSION)
+-t $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 else
 define CPU_PT2_TAGS
 -t $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(VERSION) \
--t $(NGC_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
--t $(NGC_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(VERSION)
+-t $(NGC_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH)
 endef
 endif
 
-.PHONY: build-pt2-cpu
-build-pt2-cpu: build-cpu-py-310-base
+.PHONY: build-pytorch2-cpu
+build-pytorch2-cpu: build-cpu-py-310-base
 	docker buildx build -f Dockerfile-default-cpu \
 	    --platform "$(PLATFORMS)" \
 		--build-arg BASE_IMAGE="$(DOCKERHUB_REGISTRY)/$(CPU_PY_310_BASE_NAME)-$(SHORT_GIT_HASH)" \
@@ -434,8 +421,8 @@ build-pt2-cpu: build-cpu-py-310-base
 		--push \
 		.
 
-.PHONY: build-pt2-gpu
-build-pt2-gpu: build-gpu-cuda-118-base
+.PHONY: build-pytorch2-gpu
+build-pytorch2-gpu: build-gpu-cuda-118-base
 	docker build -f Dockerfile-default-gpu \
 		--build-arg BASE_IMAGE="$(DOCKERHUB_REGISTRY)/$(GPU_CUDA_118_BASE_NAME)-$(SHORT_GIT_HASH)" \
 		--build-arg TORCH_PIP="$(TORCH2_PIP_GPU)" \
@@ -448,81 +435,83 @@ build-pt2-gpu: build-gpu-cuda-118-base
 		--build-arg HOROVOD_WITHOUT_MPI="$(HOROVOD_WITHOUT_MPI)" \
 		--build-arg HOROVOD_CPU_OPERATIONS="$(HOROVOD_CPU_OPERATIONS)" \
 		-t $(DOCKERHUB_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(DOCKERHUB_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME)-$(VERSION) \
 		-t $(NGC_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME)-$(SHORT_GIT_HASH) \
-		-t $(NGC_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME)-$(VERSION) \
 		.
 
 # tf1 and tf2.4 images are not published to NGC due to vulnerabilities.
 .PHONY: publish-tf2-cpu
 publish-tf2-cpu:
-	scripts/publish-docker.sh tf2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
-	scripts/publish-docker.sh tf2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
+	scripts/publish-docker.sh tf2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
+	scripts/publish-docker.sh tf2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_TF2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
 
 .PHONY: publish-tf2-gpu
 publish-tf2-gpu:
-	scripts/publish-docker.sh tf2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
-	scripts/publish-docker.sh tf2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+	scripts/publish-docker.sh tf2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
+	scripts/publish-docker.sh tf2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
 ifneq ($(NGC_PUBLISH),)
-	scripts/publish-docker.sh tf2-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION)
+	scripts/publish-docker.sh tf2-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_TF2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) 
 endif
 
-.PHONY: publish-pt-cpu
-publish-pt-cpu:
-	scripts/publish-docker.sh pt-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
-	scripts/publish-docker.sh pt-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
+.PHONY: publish-pytorch-cpu
+publish-pytorch-cpu:
+	scripts/publish-docker.sh pytorch-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
+	scripts/publish-docker.sh pytorch-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PT_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
 
-.PHONY: publish-pt-gpu
-publish-pt-gpu:
-	scripts/publish-docker.sh pt-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
-	scripts/publish-docker.sh pt-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+.PHONY: publish-pytorch-gpu
+publish-pytorch-gpu:
+	scripts/publish-docker.sh pytorch-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_113_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
+	scripts/publish-docker.sh pytorch-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
 ifneq ($(NGC_PUBLISH),)
-	scripts/publish-docker.sh pt-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION)
+	scripts/publish-docker.sh pytorch-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_PT_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) 
 endif
 
-.PHONY: publish-pt2-cpu
-publish-pt2-cpu:
-	scripts/publish-docker.sh pt2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_310_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
-	scripts/publish-docker.sh pt2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
+.PHONY: publish-pytorch2-cpu
+publish-pytorch2-cpu:
+	scripts/publish-docker.sh pytorch2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_310_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
+	scripts/publish-docker.sh pytorch2-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PT2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
 
-.PHONY: publish-pt2-gpu
-publish-pt2-gpu:
-	scripts/publish-docker.sh pt2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_118_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
-	scripts/publish-docker.sh pt2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+.PHONY: publish-pytorch2-gpu
+publish-pytorch2-gpu:
+	scripts/publish-docker.sh pytorch2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_118_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
+	scripts/publish-docker.sh pytorch2-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
 ifneq ($(NGC_PUBLISH),)
-	scripts/publish-docker.sh pt2-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION)
+	scripts/publish-docker.sh pytorch2-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_PT2_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) 
 endif
 
-.PHONY: publish-deepspeed-gpu
-publish-deepspeed-gpu:
-	scripts/publish-docker.sh deepspeed-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+.PHONY: publish-deepspeed
+publish-deepspeed:
+	scripts/publish-docker.sh deepspeed-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
 ifneq ($(NGC_PUBLISH),)
-	scripts/publish-docker.sh deepspeed-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION)
+	scripts/publish-docker.sh deepspeed-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) 
 endif
 
-.PHONY: publish-gpt-neox-deepspeed-gpu
-publish-gpt-neox-deepspeed-gpu:
-	scripts/publish-docker.sh gpt-neox-deepspeed-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+.PHONY: publish-gpt-neox-deepspeed
+publish-gpt-neox-deepspeed:
+	scripts/publish-docker.sh gpt-neox-deepspeed-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
 ifneq ($(NGC_PUBLISH),)
-	scripts/publish-docker.sh gpt-neox-deepspeed-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION)
+	scripts/publish-docker.sh gpt-neox-deepspeed-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_GPT_NEOX_DEEPSPEED_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) 
 endif
 
 .PHONY: publish-tf28-cpu
 publish-tf28-cpu:
-	scripts/publish-docker.sh tf28-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
-	scripts/publish-docker.sh tf28-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR) --no-push
+	scripts/publish-docker.sh tf28-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_PY_38_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
+	scripts/publish-docker.sh tf28-cpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(CPU_TF28_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR) --no-push
 
 .PHONY: publish-tf28-gpu
 publish-tf28-gpu:
-	scripts/publish-docker.sh tf28-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_112_BASE_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
-	scripts/publish-docker.sh tf28-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+	scripts/publish-docker.sh tf28-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_CUDA_112_BASE_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
+	scripts/publish-docker.sh tf28-gpu-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(ARTIFACTS_DIR)
 ifneq ($(NGC_PUBLISH),)
-	scripts/publish-docker.sh tf28-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION)
+	scripts/publish-docker.sh tf28-gpu-$(WITH_MPI) $(NGC_REGISTRY)/$(GPU_TF28_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) 
 endif
 
-.PHONY: publish-pytorch10-tf27-rocm50
-publish-pytorch10-tf27-rocm50:
-	scripts/publish-docker.sh pytorch10-tf27-rocm50-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(ROCM50_TORCH_TF_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+.PHONY: publish-pytorch13-tf210-rocm56
+publish-pytorch13-tf210-rocm56:
+	scripts/publish-docker.sh pytorch13-tf210-rocm56-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(ROCM56_TORCH13_TF_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
+
+.PHONY: publish-pytorch20-tf210-rocm56
+publish-pytorch20-tf210-rocm56:
+	scripts/publish-docker.sh pytorch20-tf210-rocm56-$(WITH_MPI) $(DOCKERHUB_REGISTRY)/$(ROCM56_TORCH_TF_ENVIRONMENT_NAME) $(SHORT_GIT_HASH) $(VERSION) $(ARTIFACTS_DIR)
 
 .PHONY: publish-cloud-images
 publish-cloud-images:
diff --git a/VERSION b/VERSION
index 12a91df0..83b47304 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-0.26.4
+0.27.1
diff --git a/dockerfile_scripts/additional-requirements-rocm.txt b/dockerfile_scripts/additional-requirements-rocm.txt
new file mode 100644
index 00000000..27585b8e
--- /dev/null
+++ b/dockerfile_scripts/additional-requirements-rocm.txt
@@ -0,0 +1,16 @@
+attrdict3
+pandas
+matplotlib
+tensorflow-datasets==1.3.2
+Keras-Preprocessing[image]
+# TODO(DET-4259) Remove this when we fix the circular dependency with the main repo.
+petname
+azure-storage-blob
+Pillow>=8.3.2,<=9.5.0
+analytics-python
+nvidia-ml-py
+protobuf<=3.20.3
+tensorboard==2.10.1
+pynvml 
+tokenizers==0.13.0
+huggingface-hub==0.16.4 
diff --git a/dockerfile_scripts/build_aws_rocm.sh b/dockerfile_scripts/build_aws_rocm.sh
new file mode 100755
index 00000000..10528ef8
--- /dev/null
+++ b/dockerfile_scripts/build_aws_rocm.sh
@@ -0,0 +1,74 @@
+#!/bin/bash
+
+set -x
+
+WITH_AWS_TRACE=""
+if [ $# -gt 1 ] ; then
+    if [ "$2" = "1" ] ; then
+	# Tell AWS to build with trace messages enabled
+	WITH_AWS_TRACE="--enable-trace"
+    fi
+    if [ $# -gt 2 ] ; then
+        if [ "$3" = "1" ] ; then
+	    WITH_MPICH=1;
+        fi
+    fi
+fi
+
+OFI=$1
+
+if [ "$OFI" = "1" ]; then
+  apt-get update \
+        && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+                tcsh
+
+  # Install AWS_OFI_NCCL
+  AWS_VER=v1.4.0
+  AWS_VER_NUM=1.4.0
+  AWS_NAME=aws-ofi-rccl-cxi
+  AWS_FILE="${AWS_NAME}-${AWS_VER_NUM}"
+  # cuda install dir likely dependent on BaseOS (i.e. ubuntu 20.02)
+  # in case this changes in the future
+
+  cuda_ver_str=`echo $CUDA_VERSION | awk -F "." '{print $1"."$2}'`
+  CUDA_DIR="/usr/local/cuda-$cuda_ver_str/targets/x86_64-linux"
+  GDRCOPY_HOME="/usr"
+  OFI_INSTALL_DIR=/container/ofi
+  ls $OFI_INSTALL_DIR
+  [ "$WITH_MPICH" = "1" ]&&MPI_INSTALL_DIR=${MPICH_INSTALL_DIR}||MPI_INSTALL_DIR=${OMPI_INSTALL_DIR}
+  AWS_CONFIG_OPTIONS="--prefix ${AWS_PLUGIN_INSTALL_DIR} \
+	  --with-libfabric=${OFI_INSTALL_DIR}            \
+	  --with-rccl=${HOROVOD_NCCL_HOME}               \
+	  --with-mpi=${MPI_INSTALL_DIR}                  \
+	  --with-hip=${ROCM_DIR} ${WITH_AWS_TRACE}"
+
+  AWS_SRC_DIR=/tmp/aws-ofi-rccl
+  #AWS_BASE_URL="https://github.com/aws/aws-ofi-nccl/archive/refs/tags"
+  AWS_BASE_URL="https://github.com/ROCmSoftwarePlatform/aws-ofi-rccl/archive/refs/heads/"
+  #AWS_URL="${AWS_BASE_URL}/${AWS_VER}.tar.gz"
+  AWS_URL="${AWS_BASE_URL}/cxi.zip"
+  # The INTERNAL_AWS_DS variable must exist in the env
+  ## INTERNAL_AWS_DS="http://set.to.your.server.name"
+  ## INTERNAL_AWS_PATH="/set/to/aws/tarball/path"
+  #if [ -z "$INTERNAL_AWS_DS" ]
+  #then
+  #  echo "Using EXTERNAL AWS $AWS_URL" 
+    # aws-ofi-nccl-1.4.0
+  #  AWS_NAME="${AWS_NAME}-${AWS_VER_NUM}"
+  #else
+  #  AWS_BASE_URL="http://${INTERNAL_AWS_DS}${INTERNAL_AWS_PATH}"
+  #  AWS_URL="${AWS_BASE_URL}/${AWS_NAME}.tar.gz"
+  #  echo "Using INTERNAL AWS $AWS_URL" 
+  #fi
+
+  mkdir -p ${AWS_SRC_DIR}                         && \
+    cd ${AWS_SRC_DIR}                             && \
+    git clone https://github.com/ROCmSoftwarePlatform/aws-ofi-rccl && \
+    cd aws-ofi-rccl                               && \
+    ./autogen.sh                                  && \
+    CC=hipcc ./configure ${AWS_CONFIG_OPTIONS}    && \
+    make                                          && \
+    make install                                  && \
+    cd /tmp                                       && \
+    rm -rf ${AWS_SRC_DIR}
+fi
diff --git a/dockerfile_scripts/ompi_rocm.sh b/dockerfile_scripts/ompi_rocm.sh
new file mode 100755
index 00000000..6bff7005
--- /dev/null
+++ b/dockerfile_scripts/ompi_rocm.sh
@@ -0,0 +1,121 @@
+#!/bin/bash
+
+set -x
+
+# See if we should add ROCM to the OMPI build
+OMPI_WITH_ROCM=""
+if [ $# -gt 2 ] ; then
+    if [ "$3" = "1" ] ; then
+	# Tell OMPI to look for rocm in the default location
+	OMPI_WITH_ROCM="--with-rocm"
+    fi
+fi
+
+if [ "$4" = "1" ];then WITH_MPICH=1;fi
+
+OS_VER=$1
+OFI=$2
+if [ "$OFI" = "1" ]; then
+  # Install OFI
+  OFI_VER=1.19.0
+  OFI_CONFIG_OPTIONS="--prefix ${OFI_INSTALL_DIR}"
+  OFI_SRC_DIR=/tmp/ofi-src
+  OFI_BASE_URL="https://github.com/ofiwg/libfabric/releases/download"
+  OFI_URL="${OFI_BASE_URL}/v${OFI_VER}/libfabric-${OFI_VER}.tar.bz2"
+
+  mkdir -p ${OFI_SRC_DIR}                              && \
+  mkdir -p ${OFI_INSTALL_DIR}                          && \
+    cd ${OFI_SRC_DIR}                                  && \
+    wget ${OFI_URL}                                    && \
+    tar -xf libfabric-${OFI_VER}.tar.bz2              && \
+    cd libfabric-${OFI_VER}                            && \
+    ./configure ${OFI_CONFIG_OPTIONS}                  && \
+    make install                                       && \
+    cd /tmp                                            && \
+    rm -rf ${OFI_SRC_DIR}
+
+  #OMPI CONFIG ARGS FOR OFI
+  OMPI_CONFIG_OPTIONS_VAR="--prefix ${OMPI_INSTALL_DIR} --enable-orterun-prefix-by-default --enable-shared --with-cma --with-pic --enable-mpi-cxx --enable-mpi-thread-multiple --with-libfabric=${OFI_INSTALL_DIR} --without-ucx --with-pmi --with-pmix=internal ${OMPI_WITH_ROCM}"
+
+#echo OMPI_CONFIG_OPTIONS_VAR $OMPI_CONFIG_OPTIONS_VAR= 
+
+#exit 99
+
+else
+  # Install the Mellanox OFED stack.  Note that this is dependent on
+  # what the base OS is (ie, Ubuntu 20.04) so if that changes then
+  # this needs updated.  MOFED_VER=5.0-2.1.8.0 MOFED_VER=5.5-1.0.3.2
+  MOFED_VER=5.4-3.4.0.0
+  PLATFORM=x86_64
+  MOFED_TAR_URL="http://content.mellanox.com/ofed/MLNX_OFED-${MOFED_VER}"
+  MOFED_TAR="MLNX_OFED_LINUX-${MOFED_VER}-${OS_VER}-${PLATFORM}.tgz"
+  TMP_INSTALL_DIR=/tmp/ofed
+  
+  mkdir -p ${TMP_INSTALL_DIR}                                          && \
+     cd ${TMP_INSTALL_DIR}                                             && \
+     wget --quiet "${MOFED_TAR_URL}/${MOFED_TAR}"                      && \
+     tar -xvf ${MOFED_TAR}                                             && \
+     MLNX_OFED_LINUX-${MOFED_VER}-${OS_VER}-${PLATFORM}/mlnxofedinstall   \
+       --user-space-only --without-fw-update --all --force                \
+       --skip-unsupported-devices-check                                && \
+     rm -rf MLNX_OFED_LINUX-${MOFED_VER}-${OS_VER}-${PLATFORM}.tgz        \
+            MLNX_OFED_LINUX-${MOFED_VER}-${OS_VER}-${PLATFORM}            \
+            MLNX_OFED_LINUX.*.logs                                     && \
+     rm -rf ${TMP_INSTALL_DIR}
+
+  # Install UCX
+  UCX_VER=1.10.1
+  UCX_CONFIG_OPTIONS="--prefix ${UCX_INSTALL_DIR} --enable-mt"
+  UCX_SRC_DIR=/tmp/ucx-src
+  UCX_BASE_URL="https://github.com/openucx/ucx/releases/download"
+  UCX_URL="${UCX_BASE_URL}/v${UCX_VER}/ucx-${UCX_VER}.tar.gz"
+
+  mkdir -p ${UCX_SRC_DIR}                              && \
+    cd ${UCX_SRC_DIR}                                  && \
+    wget ${UCX_URL}                                    && \
+    tar -xzf ucx-${UCX_VER}.tar.gz                     && \
+    cd ucx-${UCX_VER}                                  && \
+    ./contrib/configure-release ${UCX_CONFIG_OPTIONS}  && \
+    make -j8 install                                   && \
+    cd /tmp                                            && \
+    rm -rf ${UCX_SRC_DIR}
+
+  #OMPI CONFIG ARGS FOR UCX
+  OMPI_CONFIG_OPTIONS_VAR="--prefix ${OMPI_INSTALL_DIR} --enable-shared --with-verbs --with-cma --with-pic --enable-mpi-cxx --enable-mpi-thread-multiple --with-pmi --with-pmix=internal --with-platform=contrib/platform/mellanox/optimized --with-ucx=/container/ucx ${OMPI_WITH_ROCM}"
+
+fi
+
+if [ "$WITH_MPICH" != "1" ]; then
+# Install OMPI
+OMPI_VER=v4.1
+ OMPI_VER_NUM=4.1.0
+ OMPI_CONFIG_OPTIONS=${OMPI_CONFIG_OPTIONS_VAR}
+ OMPI_SRC_DIR=/tmp/openmpi-src
+ OMPI_BASE_URL="https://download.open-mpi.org/release/open-mpi"
+ OMPI_URL="${OMPI_BASE_URL}/${OMPI_VER}/openmpi-${OMPI_VER_NUM}.tar.gz"
+
+ mkdir -p ${OMPI_SRC_DIR}                        && \
+   cd ${OMPI_SRC_DIR}                            && \
+   wget ${OMPI_URL}                              && \
+   tar -xzf openmpi-${OMPI_VER_NUM}.tar.gz       && \
+   cd openmpi-${OMPI_VER_NUM}                    && \
+  ./configure ${OMPI_CONFIG_OPTIONS}            && \
+   make                                          && \
+   make install                                  && \
+   cd /tmp                                       && \
+   rm -rf ${OMPI_SRC_DIR}
+else
+MPICH_VER=4.1.2
+MPICH_CONFIG_OPTIONS="--with-device=ch4:ofi --with-libfabric=/container/ofi --prefix=${MPICH_INSTALL_DIR} --disable-fortran"
+MPICH_URL="http://www.mpich.org/static/downloads/${MPICH_VER}/mpich-${MPICH_VER}.tar.gz"
+MPICH_SRC_DIR=/tmp/mpich-src
+
+mkdir -p ${MPICH_SRC_DIR}                                 &&\
+  cd ${MPICH_SRC_DIR}                                     &&\
+  curl -fSsL --retry 3 "${MPICH_URL}" | tar -xz --strip 1 &&\
+  ./configure ${MPICH_CONFIG_OPTIONS}                     &&\
+  make                                                    &&\
+  make install                                            &&\
+  rm -rf "${MPICH_SRC_DIR}"
+
+fi
diff --git a/dockerfile_scripts/scrape_libs.sh b/dockerfile_scripts/scrape_libs.sh
index 1c2c47ad..0ed1d2b6 100755
--- a/dockerfile_scripts/scrape_libs.sh
+++ b/dockerfile_scripts/scrape_libs.sh
@@ -1,12 +1,18 @@
 #!/bin/bash
 
+WHOAMI=$(whoami)
 # Only scrape host libs if AWS/NCCL/OFI was built for this image
 if [ -d /container/aws/lib ]
 then
-   # See if we mounted in host libs in the expected location
-   host_dir="/host"
+   host_dir="/det_libfabric"
+   if [ ! -d "$host_dir" ]; then
+     host_dir="/det_host"
+     if [ ! -d "$host_dir" ]; then
+       host_dir="/host"
+     fi
+   fi
    if [ -d "$host_dir" ]; then
-       libfabric=`find /host -name libfabric.so 2>/dev/null`
+       libfabric=`find $host_dir -name libfabric.so 2>/dev/null`
        libfabric_dir="$(dirname "$libfabric")"
        if [[ ! -z "$libfabric" ]] ; then
            # Need libfabric to be first in the LD_LIBRARY_PATH to
@@ -15,9 +21,9 @@ then
            # after the container ones so we append the rest of the
            # host libs.  tmp_dir="`pwd`/tmp"
            tmp_dir="/var/tmp"
-           tmp_lib_dir="$tmp_dir/$(whoami)/detAI/lib"
+           tmp_lib_dir="$tmp_dir/${WHOAMI}/detAI/lib"
            mkdir -p $tmp_lib_dir
-           for lib in `/bin/ls $libfabric_dir/ | grep libfabric` ; do
+           for lib in `/bin/ls $libfabric_dir/ | grep -elibfabric -ecxi` ; do
                ln -s $libfabric_dir/$lib $tmp_lib_dir 2>/dev/null
            done
            # Prepend the tmp dir for the host libfabric
@@ -27,13 +33,26 @@ then
            # to set these paths based on where we find the libs that
            # libfabric is dependent upon.
            export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$libfabric_dir
-           export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/host/usr/lib64
-           export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/host/usr/local/lib64
+       else 
+	       echo "libfabric not found within $host_dir." >&2
        fi # end if found libfabric.so
-   fi # end if /host exists
+   else echo "no suitable mounts available." >&2
+   fi # end if /det_libfabric exists
+   # See if we mounted in host libs in the expected location
+
+   host_dir="/det_host"
+   if [ ! -d "$host_dir" ]; then
+   	host_dir="/host"
+   fi
+   if [ -d "$host_dir" ]; then
+       # to set these paths based on where we find the libs that
+       # libfabric is dependent upon.
+       export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$host_dir/usr/lib64
+       export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$host_dir/usr/local/lib64
+   fi # end if /det_host exists
    export LD_LIBRARY_PATH=/container/aws/lib:$LD_LIBRARY_PATH
 
-   tmp_nvcache_dir=$(mktemp -d -p /var/tmp $(whoami)-nvcache-XXXXXXXX)
+   tmp_nvcache_dir=$(mktemp -d -p /var/tmp ${WHOAMI}-nvcache-XXXXXXXX)
 
    # The following env variables are only set if they are curretly unset,empty or null
    # To override one or more of these variales, simply set them prior to the enrtrypoint
@@ -55,7 +74,51 @@ then
    then
       export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libffi.so.7:$LD_PRELOAD
    fi
+# Execute what we were told to execute
+
+   if [ "$WITH_RCCL" = "1" ]; then
+      #export LD_LIBRARY_PATH=/scratch2/${WHOAMI}/lib/libfabric-1.21.1:/scratch2/${WHOAMI}/lib:/container/aws/lib:$LD_LIBRARY_PATH
+      #export LD_LIBRARY_PATH=/scratch2/${WHOAMI}/lib/libfabric-1.21.1:/scratch2/${WHOAMI}/lib:$LD_LIBRARY_PATH
 
+      export TF_FORCE_GPU_ALLOW_GROWTH=true
+      export CUDA_CACHE_PATH=/tmp/${WHOAMI}.nvcache
+      export FI_CXI_COMPAT=0
+      export NCCL_SOCKET_IFNAME="hsn0"
+      #export FI_CXI_DISABLE_HOST_REGISTER=1
+      export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+      export FI_CXI_DEFAULT_TX_SIZE=1024
+      export FI_CXI_DISABLE_CQ_HUGETLB=1
+      export FI_CXI_DEFAULT_CQ_SIZE=131072
+      export FI_CXI_DISABLE_HOST_REGISTER=1
+      export FI_CXI_RX_MATCH_MODE=software
+      export FI_CXI_RDZV_PROTO=alt_read
+      export FI_PROVIDER=^ofi_rxm,efa,ofi_rxd
+      export FI_HMEM_CUDA_USE_GDRCOPY=1
+      export FI_CXI_REQ_BUF_SIZE=8338608
+      export FI_LOG_PROV=cxi
+      export FI_LOG_LEVEL=info
+      export FI_LOG_SUBSYS=domain
+      export FI_CXI_TELEMENTRY=pct_no_mst_nacks,pct_mst_hit_on_som,pct_sct_timeouts,pct_spt_timeouts,pct_tct_timeouts
+
+      export CXI_FORK_SAFE=1
+      export FI_EFA_FORK_SAFE=${CXI_FORK_SAFE}
+      # Avoid hang with NCCL/RCCL and libfabric
+      export FI_MR_CACHE_MONITOR=userfaultfd
+
+      export NCCL_DEBUG=TRACE
+      export RCCL_DEBUG=${NCCL_DEBUG}
+
+      if [ "$WITH_NFS_WORKAROUND" = "1" ]; then
+         export MIOPEN_USER_DB_PATH="/tmp/${WHOAMI}_${SLURM_LOCALID}"
+         export MIOPEN_USER_CACHE_PATH="$MIOPEN_USER_DB_PATH/.cache"
+         export MIOPEN_CACHE_DIR=${MIOPEN_USER_CACHE_PATH}
+         echo "MIOPEN_USER_DB_PATH: $MIOPEN_USER_DB_PATH, cache dir: $MIOPEN_USER_CACHE_PATH"
+         mkdir -p $MIOPEN_USER_DB_PATH
+         mkdir -p $MIOPEN_USER_CACHE_PATH/miopen
+         export HOME=${MIOPEN_USER_DB_PATH}
+
+      fi
+   fi
 fi
-# Execute what we were told to execute
 exec "${@}"
diff --git a/scripts/publish-docker.sh b/scripts/publish-docker.sh
index cf91f1b0..a5b0423e 100755
--- a/scripts/publish-docker.sh
+++ b/scripts/publish-docker.sh
@@ -3,22 +3,20 @@
 set -e
 set -x
 
-if [ "$#" -lt 4 ] || [ "$#" -gt 6 ] || [ "$#" -eq 6 ] && [ "$6" != "--no-push" ]; then
-    echo "usage: $0 LOG_NAME BASE_TAG HASH VERSION ARTIFACTS_DIR [--no-push]" >&2
+if [ "$#" -gt 5 ] || [ "$#" -eq 5 ] && [ "$5" != "--no-push" ]; then
+    echo "usage: $0 LOG_NAME BASE_TAG HASH ARTIFACTS_DIR [--no-push]" >&2
     exit 1
 fi
 
 log_name="$1"
 base_tag="$2"
 hash="$3"
-version="$4"
-artifacts="$5"
+artifacts="$4"
 
 underscore_name="$(echo -n "$log_name" | tr - _)"
 
-if [ "$#" -lt 6 ] ; then
+if [ "$#" -lt 5 ] ; then
     docker push "$base_tag-$hash"
-    docker push "$base_tag-$version"
 fi
 
 if [ -n "$artifacts" ]; then
@@ -27,6 +25,5 @@ if [ -n "$artifacts" ]; then
     log_file="$artifacts/publish-$log_name"
     (
         echo "${underscore_name}_hashed: $base_tag-$hash"
-        echo "${underscore_name}_versioned: $base_tag-$version"
     ) > "$log_file"
 fi
diff --git a/version_matrix.yaml b/version_matrix.yaml
new file mode 100644
index 00000000..4b2f8102
--- /dev/null
+++ b/version_matrix.yaml
@@ -0,0 +1,57 @@
+git_hash: <git_hash_here>
+images:
+	- pytorch-deepspeed:
+	 	pytorch_version: 1.10.2
+	 	deepspeed_version: 0.8.3
+	 	python_version: 3.9.16
+	 	ubuntu_version: 20.04
+	- gpt-neox-deepspeed:
+	 	pytorch_version: 1.10.2
+	 	deepspeed_version: 0.8.3+https://github.com/determined-ai/deepspeed.git@eleuther_dai
+	 	python_version: 3.9.16
+	 	ubuntu_version: 20.04
+    - tf28-cpu:
+	 	tensorflow_version: 2.8.4
+	 	horovod_version: 0.24.2
+    - tf28-gpu:
+	 	tensorflow_version: 2.8.3
+	 	horovod_version: 0.24.2
+	 	ubuntu_version: 20.04
+    - tf2-cpu:
+	 	tensorflow_version: 2.11.1
+	 	horovod_version: 0.28.1
+	 	ubuntu_version: 20.04
+    - tf2-gpu:
+	 	tensorflow_version: 2.11.1
+	 	horovod_version: 0.28.1
+	 	ubuntu_version: 20.04
+    - pytorch-cpu:
+	 	pytorch_version: 1.12
+	 	horovod_version: 0.28.1
+	 	ubuntu_version: 20.04
+    - pytorch-gpu:
+	 	pytorch_version: 1.12
+	 	horovod_version: 0.28.1
+	 	ubuntu_version: 20.04
+    - pytorch2-cpu:
+	 	pytorch_version: 2.0
+	 	horovod_version: 0.28.1
+	 	ubuntu_version: 20.04
+    - pytorch2-gpu:
+	 	pytorch_version: 2.0
+	 	horovod_version: 0.28.1
+	 	ubuntu_version: 20.04
+	- pytorch13-tf210-rocm56:
+	 	pytorch_version: 1.13.1
+	 	tensorflow_version: 2.10.1.540
+	 	horovod_version: 0.28.1
+	 	python_version: 3.8
+	 	ubuntu_version: 20.04
+	 	rocm_version: 5.6
+	- pytorch20-tf210-rocm56:
+	 	pytorch_version: 2.0.1
+	 	tensorflow_version: 2.10.1.540
+	 	horovod_version: 0.28.1
+	 	python_version: 3.8
+	 	ubuntu_version: 20.04
+	 	rocm_version: 5.6