diff --git a/.github/workflows/docker-efa.yml b/.github/workflows/docker-efa.yml
deleted file mode 100644
index 5f6ac0064f..0000000000
--- a/.github/workflows/docker-efa.yml
+++ /dev/null
@@ -1,46 +0,0 @@
-name: Build EFA Docker image
-
-on:
-  workflow_dispatch:
-    inputs:
-      image_name:
-        description: "Docker image name"
-        required: true
-        default: "dstackai/efa"
-      dstack_revision:
-        description: "Docker image revision"
-        required: true
-        default: 0
-
-jobs:
-  build-efa:
-    defaults:
-      run:
-        working-directory: docker/efa
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-      - name: Login to DockerHub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_TOKEN }}
-      - name: Build and upload to DockerHub
-        run: |
-          IMAGE_NAME=${{ inputs.image_name }}
-          BUILD_DATE=$(date --utc --iso-8601=seconds)Z
-          docker buildx build . \
-            --load \
-            --provenance=false \
-            --platform linux/amd64 \
-            --build-arg IMAGE_NAME=${IMAGE_NAME} \
-            --build-arg DSTACK_REVISION=${{ inputs.dstack_revision }} \
-            --build-arg BUILD_DATE=${BUILD_DATE} \
-            --tag ${IMAGE_NAME}:latest
-          VERSION=$(docker inspect --format '{{ index .Config.Labels "org.opencontainers.image.version" }}' ${IMAGE_NAME})
-          docker tag ${IMAGE_NAME}:latest ${IMAGE_NAME}:${VERSION}
-          docker push ${IMAGE_NAME}:${VERSION}
-          docker push ${IMAGE_NAME}:latest
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 0beaf582ba..f4c97b5da1 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -51,8 +51,8 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-        flavor: ["base", "devel"]
+        flavor: ["base", "devel", "devel-efa"]
+        ubuntu_version: ["22"]
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -67,7 +67,21 @@ jobs:
         uses: docker/setup-qemu-action@v3
       - name: Build and upload to DockerHub
         run: |
-          docker buildx build --platform linux/amd64 --build-arg FLAVOR=${{ matrix.flavor }} --build-arg PYTHON=${{ matrix.python }} --push --provenance=false --tag dstackai/${{ env.BUILD_DOCKER_REPO }}:py${{ matrix.python }}-${{ inputs.image_version }}-cuda-12.1${{ matrix.flavor == 'devel' && '-devel' || '' }} -f base/Dockerfile .
+         if [ "${{ matrix.flavor }}" = "base" ]; then
+            FILE="base/Dockerfile"
+          elif [ "${{ matrix.flavor }}" = "devel" ]; then
+            FILE="base/Dockerfile"
+          else
+            FILE="base/efa/Dockerfile"
+          fi
+          docker buildx build \
+            --platform linux/amd64 \
+            --tag dstackai/${{ env.BUILD_DOCKER_REPO }}:${{ inputs.image_version }}-${{ matrix.flavor }}-ubuntu${{ matrix.ubuntu_version }}.04 \
+            --build-arg FLAVOR=${{ matrix.flavor }} \
+            --build-arg UBUNTU_VERSION=${{ matrix.ubuntu_version }} \
+            --provenance=false \
+            --push \
+            -f $FILE .
 
   build-aws-images:
     needs: build-docker
diff --git a/docker/base/Dockerfile b/docker/base/Dockerfile
index 2ed2fbd984..a42788984a 100644
--- a/docker/base/Dockerfile
+++ b/docker/base/Dockerfile
@@ -1,28 +1,79 @@
+# syntax = edrevo/dockerfile-plus
+ARG UBUNTU_VERSION
+
+# Build stage
+FROM nvidia/cuda:12.1.1-base-ubuntu${UBUNTU_VERSION}.04 AS builder
+
+ENV NCCL_HOME=/opt/nccl
+ENV CUDA_HOME=/usr/local/cuda
+ENV OPEN_MPI_PATH=/usr/lib/x86_64-linux-gnu/openmpi
+
+# Prerequisites
+
+RUN export DEBIAN_FRONTEND=noninteractive \
+    && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${UBUNTU_VERSION}04/x86_64/3bf863cc.pub \
+    && apt-get update --fix-missing \
+    && apt-get upgrade -y \
+    && ln -fs /usr/share/zoneinfo/America/New_York /etc/localtime \
+    && apt-get install -y tzdata \
+    && dpkg-reconfigure --frontend noninteractive tzdata \
+    && cuda_version=$(echo ${CUDA_VERSION} | awk -F . '{ print $1"-"$2 }') \
+    && apt-get install -y --no-install-recommends \
+        cuda-libraries-dev-${cuda_version} \
+        cuda-nvcc-${cuda_version} \
+        libhwloc-dev \
+        autoconf \
+        automake \
+        libtool \
+        libopenmpi-dev \
+        git \
+        curl \
+        python3 \
+        build-essential
+
+# NCCL
+
+ARG NCCL_VERSION=2.26.2-1
+
+RUN cd /tmp \
+    && git clone https://github.com/NVIDIA/nccl.git -b v${NCCL_VERSION} \
+    && cd nccl \
+    && make -j$(nproc) src.build BUILDDIR=${NCCL_HOME}
+
+# NCCL tests
+
+RUN cd /opt \
+    && git clone https://github.com/NVIDIA/nccl-tests \
+    && cd nccl-tests \
+    && make -j$(nproc) \
+        MPI=1 \
+        MPI_HOME=${OPEN_MPI_PATH} \
+        CUDA_HOME=${CUDA_HOME} \
+        NCCL_HOME=${NCCL_HOME}
+
+# Final stage
+
+INCLUDE+ base/Dockerfile.common
+
+ENV NCCL_HOME=/opt/nccl
+
+COPY --from=builder ${NCCL_HOME} ${NCCL_HOME}
+COPY --from=builder /opt/nccl-tests/build /opt/nccl-tests/build
+
 ARG FLAVOR
-FROM nvidia/cuda:12.1.1-${FLAVOR}-ubuntu20.04
-
-ARG PYTHON
-ARG _UV_HOME="/opt/uv"
-ENV UV_PYTHON="${PYTHON}"
-ENV UV_INSTALL_DIR="${_UV_HOME}/bin"
-ENV UV_PYTHON_INSTALL_DIR="${_UV_HOME}/python"
-ENV UV_PYTHON_BIN_DIR="${UV_PYTHON_INSTALL_DIR}/bin"
-ENV UV_MANAGED_PYTHON=1
-ENV LANG=C.UTF-8 LC_ALL=C.UTF-8
-
-ENV PATH="${UV_INSTALL_DIR}:${UV_PYTHON_BIN_DIR}:${PATH}"
-
-RUN export DEBIAN_FRONTEND=noninteractive && \
-    apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/3bf863cc.pub && \
-    apt-get update --fix-missing && \
-    apt-get upgrade -y && \
-    ln -fs /usr/share/zoneinfo/America/New_York /etc/localtime && \
-    apt-get install -y tzdata && \
-    dpkg-reconfigure --frontend noninteractive tzdata && \
-    apt-get install -y bzip2 ca-certificates curl build-essential git libglib2.0-0 libsm6 libxext6 libxrender1 mercurial openssh-server subversion wget \
-    libibverbs1 ibverbs-providers ibverbs-utils libibverbs-dev infiniband-diags && \
-    sed -i "s/.*PasswordAuthentication.*/PasswordAuthentication no/g" /etc/ssh/sshd_config && mkdir /run/sshd && \
-    mkdir ~/.ssh && chmod 700 ~/.ssh && touch ~/.ssh/authorized_keys && chmod 600 ~/.ssh/authorized_keys && rm /etc/ssh/ssh_host_*
-
-RUN curl -LsSf https://astral.sh/uv/install.sh | INSTALLER_NO_MODIFY_PATH=1 sh && \
-    uv python install --preview --default
+
+# MPI, NVCC, and /etc/ld.so.conf.d
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        openmpi-bin \
+    && if [ "$FLAVOR" = "devel" ]; then \
+        cuda_version=$(echo ${CUDA_VERSION} | awk -F . '{ print $1"-"$2 }') \
+        && apt-get install -y --no-install-recommends \
+            cuda-libraries-dev-${cuda_version} \
+            cuda-nvcc-${cuda_version} \
+            libhwloc-dev; \
+    fi \
+    && rm -rf /var/lib/apt/lists/* \
+    && echo "${NCCL_HOME}/lib" >> /etc/ld.so.conf.d/nccl.conf \
+    && ldconfig
diff --git a/docker/base/Dockerfile.common b/docker/base/Dockerfile.common
new file mode 100644
index 0000000000..ae76f30124
--- /dev/null
+++ b/docker/base/Dockerfile.common
@@ -0,0 +1,35 @@
+ARG UBUNTU_VERSION
+
+FROM nvidia/cuda:12.1.1-base-ubuntu${UBUNTU_VERSION}.04
+
+ARG _UV_HOME="/opt/uv"
+
+ENV UV_INSTALL_DIR="${_UV_HOME}/bin"
+ENV UV_MANAGED_PYTHON=1
+ENV LANG=C.UTF-8 LC_ALL=C.UTF-8
+
+ENV PATH="${UV_INSTALL_DIR}:${PATH}"
+
+ENV OMPI_MCA_pml=^cm,ucx
+ENV OMPI_MCA_btl=tcp,self
+ENV OMPI_MCA_btl_tcp_if_exclude=lo,docker0
+ENV NCCL_SOCKET_IFNAME=^docker,lo
+
+RUN export DEBIAN_FRONTEND=noninteractive \
+    && apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${UBUNTU_VERSION}04/x86_64/3bf863cc.pub \
+    && apt-get update --fix-missing \
+    && apt-get upgrade -y \
+    && ln -fs /usr/share/zoneinfo/America/New_York /etc/localtime \
+    && apt-get install -y tzdata \
+    && dpkg-reconfigure --frontend noninteractive tzdata \
+    && apt-get install -y bzip2 ca-certificates curl build-essential git libglib2.0-0 libsm6 libxext6 libxrender1 mercurial openssh-server subversion wget \
+        libibverbs1 ibverbs-providers ibverbs-utils libibverbs-dev infiniband-diags \
+    && rm -rf /var/lib/apt/lists/* \
+    && sed -i "s/.*PasswordAuthentication.*/PasswordAuthentication no/g" /etc/ssh/sshd_config \
+    && mkdir /run/sshd \
+    && mkdir ~/.ssh && chmod 700 ~/.ssh && touch ~/.ssh/authorized_keys \
+    && chmod 600 ~/.ssh/authorized_keys \
+    && rm /etc/ssh/ssh_host_*
+
+RUN curl -LsSf https://astral.sh/uv/install.sh | INSTALLER_NO_MODIFY_PATH=1 sh \
+    && uv python install --preview --default
diff --git a/docker/efa/Dockerfile b/docker/base/efa/Dockerfile
similarity index 61%
rename from docker/efa/Dockerfile
rename to docker/base/efa/Dockerfile
index 0f8d717b0d..50b6c1c5ef 100644
--- a/docker/efa/Dockerfile
+++ b/docker/base/efa/Dockerfile
@@ -1,15 +1,15 @@
-ARG BASE_IMAGE=dstackai/base:py3.12-0.7-cuda-12.1
+# syntax = edrevo/dockerfile-plus
 
-FROM ${BASE_IMAGE}
+INCLUDE+ base/Dockerfile.common
 
-ENV PREFIX=/usr/local
-ENV CUDA_PATH=/usr/local/cuda
+ENV NCCL_HOME=/usr/local
+ENV CUDA_HOME=/usr/local/cuda
 ENV LIBFABRIC_PATH=/opt/amazon/efa
 ENV OPEN_MPI_PATH=/opt/amazon/openmpi
 ENV PATH="${LIBFABRIC_PATH}/bin:${OPEN_MPI_PATH}/bin:${PATH}"
 ENV LD_LIBRARY_PATH="${OPEN_MPI_PATH}/lib:${LD_LIBRARY_PATH}"
 
-# prerequisites
+# Prerequisites
 
 RUN cuda_version=$(echo ${CUDA_VERSION} | awk -F . '{ print $1"-"$2 }') \
     && apt-get update \
@@ -19,61 +19,58 @@ RUN cuda_version=$(echo ${CUDA_VERSION} | awk -F . '{ print $1"-"$2 }') \
         libhwloc-dev \
         autoconf \
         automake \
-        libtool
+        libtool \
+    && rm -rf /var/lib/apt/lists/*
 
 # EFA
 
 ARG EFA_VERSION=1.38.1
 
-RUN cd $HOME \
+RUN cd /tmp \
+    && apt-get update \
     && curl -O https://s3-us-west-2.amazonaws.com/aws-efa-installer/aws-efa-installer-${EFA_VERSION}.tar.gz \
     && tar -xf aws-efa-installer-${EFA_VERSION}.tar.gz \
     && cd aws-efa-installer \
-    && ./efa_installer.sh -y --skip-kmod -g
+    && ./efa_installer.sh -y --skip-kmod -g \
+    && rm -rf /tmp/aws-efa-installer /var/lib/apt/lists/*
 
 # NCCL
 
 ARG NCCL_VERSION=2.26.2-1
 
-RUN cd $HOME \
+RUN cd /tmp \
     && git clone https://github.com/NVIDIA/nccl.git -b v${NCCL_VERSION} \
     && cd nccl \
-    && make -j$(nproc) src.build BUILDDIR=${PREFIX}
+    && make -j$(nproc) src.build BUILDDIR=${NCCL_HOME} \
+    && rm -rf /tmp/nccl
 
 # AWS OFI NCCL
 
 ARG OFI_VERSION=1.14.0
 
-RUN cd $HOME \
+RUN cd /tmp \
     && git clone https://github.com/aws/aws-ofi-nccl.git -b v${OFI_VERSION} \
     && cd aws-ofi-nccl \
     && ./autogen.sh \
     && ./configure \
-        --with-cuda=${CUDA_PATH} \
+        --with-cuda=${CUDA_HOME} \
         --with-libfabric=${LIBFABRIC_PATH} \
         --with-mpi=${OPEN_MPI_PATH} \
-        --with-cuda=${CUDA_PATH} \
-        --with-nccl=${PREFIX} \
+        --with-cuda=${CUDA_HOME} \
+        --with-nccl=${NCCL_HOME} \
         --disable-tests \
-        --prefix=${PREFIX} \
-    && make -j$(numproc) \
-    && make install
+        --prefix=${NCCL_HOME} \
+    && make -j$(nproc) \
+    && make install \
+    && rm -rf /tmp/aws-ofi-nccl /var/lib/apt/lists/*
 
 # NCCL Tests
 
-RUN cd $HOME \
+RUN cd /opt \
     && git clone https://github.com/NVIDIA/nccl-tests \
     && cd nccl-tests \
-    && make -j$(numproc) \
+    && make -j$(nproc) \
         MPI=1 \
         MPI_HOME=${OPEN_MPI_PATH} \
-        CUDA_HOME=${CUDA_PATH} \
-        NCCL_HOME=${PREFIX}
-
-ARG BUILD_DATE
-ARG IMAGE_NAME
-ARG DSTACK_REVISION
-
-LABEL org.opencontainers.image.title="${IMAGE_NAME}"
-LABEL org.opencontainers.image.version="${EFA_VERSION}-${DSTACK_REVISION}"
-LABEL org.opencontainers.image.created="${BUILD_DATE}"
+        CUDA_HOME=${CUDA_HOME} \
+        NCCL_HOME=${NCCL_HOME}
diff --git a/docker/efa/README.md b/docker/base/efa/README.md
similarity index 100%
rename from docker/efa/README.md
rename to docker/base/efa/README.md
diff --git a/examples/clusters/nccl-tests/.dstack.yml b/examples/clusters/nccl-tests/.dstack.yml
index 3870731e35..164148b3c7 100644
--- a/examples/clusters/nccl-tests/.dstack.yml
+++ b/examples/clusters/nccl-tests/.dstack.yml
@@ -5,12 +5,9 @@ nodes: 2
 startup_order: workers-first
 stop_criteria: master-done
 
-# This image comes with MPI and NCCL tests pre-built
-image: dstackai/efa
 env:
   - NCCL_DEBUG=INFO
 commands:
-  - cd /root/nccl-tests/build
   - |
     if [ $DSTACK_NODE_RANK -eq 0 ]; then
       mpirun \
@@ -18,13 +15,12 @@ commands:
         --hostfile $DSTACK_MPI_HOSTFILE \
         -n $DSTACK_GPUS_NUM \
         -N $DSTACK_GPUS_PER_NODE \
-        --mca btl_tcp_if_exclude lo,docker0 \
         --bind-to none \
-        ./all_reduce_perf -b 8 -e 8G -f 2 -g 1
+        /opt/nccl-tests/build/all_reduce_perf -b 8 -e 8G -f 2 -g 1
     else
       sleep infinity
     fi
 
 resources:
-  gpu: nvidia:4:16GB
+  gpu: nvidia:1..8
   shm_size: 16GB
diff --git a/examples/distributed-training/torchrun/.dstack.yml b/examples/distributed-training/torchrun/.dstack.yml
index 062051f365..4eccdb263e 100644
--- a/examples/distributed-training/torchrun/.dstack.yml
+++ b/examples/distributed-training/torchrun/.dstack.yml
@@ -1,7 +1,6 @@
 type: task
 name: train-distrib
 
-# The size of the cluster
 nodes: 2
 
 python: 3.12
@@ -21,6 +20,5 @@ commands:
       multinode.py 50 10
 
 resources:
-  gpu: 24GB:1..2
-  # Uncomment if using multiple GPUs
-  #shm_size: 24GB
+  gpu: 1..8
+  shm_size: 16GB
diff --git a/scripts/packer/aws-image-cuda.json b/scripts/packer/aws-image-cuda.json
index 64f4280bb3..f7719304ef 100644
--- a/scripts/packer/aws-image-cuda.json
+++ b/scripts/packer/aws-image-cuda.json
@@ -81,14 +81,6 @@
     {
       "type": "shell",
       "script": "provisioners/install-nvidia-container-toolkit.sh"
-    },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
     }
   ]
 }
diff --git a/scripts/packer/aws-image.json b/scripts/packer/aws-image.json
index c9e3cd5fb8..0327d6f9fb 100644
--- a/scripts/packer/aws-image.json
+++ b/scripts/packer/aws-image.json
@@ -71,14 +71,6 @@
         "cd /tmp",
         "chmod +x install-docker.sh",
         "./install-docker.sh --version {{user `docker_version`}}"]
-    },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
     }
   ]
 }
diff --git a/scripts/packer/azure-image-cuda.json b/scripts/packer/azure-image-cuda.json
index 4107c1ff65..c191282aec 100644
--- a/scripts/packer/azure-image-cuda.json
+++ b/scripts/packer/azure-image-cuda.json
@@ -73,14 +73,6 @@
       "type": "shell",
       "script": "provisioners/install-nvidia-container-toolkit.sh"
     },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
-    },
     {
       "type": "shell",
       "execute_command": "chmod +x {{ .Path }}; {{ .Vars }} sudo -E sh '{{ .Path }}'",
diff --git a/scripts/packer/azure-image-grid.json b/scripts/packer/azure-image-grid.json
index adef2d4ba9..58239b777a 100644
--- a/scripts/packer/azure-image-grid.json
+++ b/scripts/packer/azure-image-grid.json
@@ -71,14 +71,6 @@
       "type": "shell",
       "script": "provisioners/install-nvidia-container-toolkit.sh"
     },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
-    },
     {
       "type": "shell",
       "execute_command": "chmod +x {{ .Path }}; {{ .Vars }} sudo -E sh '{{ .Path }}'",
diff --git a/scripts/packer/azure-image.json b/scripts/packer/azure-image.json
index 2c5602040a..0b0d378335 100644
--- a/scripts/packer/azure-image.json
+++ b/scripts/packer/azure-image.json
@@ -63,14 +63,6 @@
         "./install-docker.sh --version {{user `docker_version`}}"
       ]
     },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
-    },
     {
       "type": "shell",
       "execute_command": "chmod +x {{ .Path }}; {{ .Vars }} sudo -E sh '{{ .Path }}'",
diff --git a/scripts/packer/build-cuda-image.pkr.hcl b/scripts/packer/build-cuda-image.pkr.hcl
index b50dbada9f..48b1c20024 100644
--- a/scripts/packer/build-cuda-image.pkr.hcl
+++ b/scripts/packer/build-cuda-image.pkr.hcl
@@ -31,9 +31,4 @@ build {
   provisioner "shell" {
     script = "provisioners/install-nvidia-container-toolkit.sh"
   }
-
-  provisioner "shell" {
-    environment_vars = ["IMAGE_VERSION=${var.image_version}"]
-    script           = "provisioners/pull-docker-images.sh"
-  }
 }
diff --git a/scripts/packer/build-image.pkr.hcl b/scripts/packer/build-image.pkr.hcl
index 4cb1517dc3..6033ee4b1f 100644
--- a/scripts/packer/build-image.pkr.hcl
+++ b/scripts/packer/build-image.pkr.hcl
@@ -22,9 +22,4 @@ build {
   provisioner "shell" {
     inline = ["cd /tmp", "chmod +x install-docker.sh", "./install-docker.sh --version ${local.docker_version}"]
   }
-
-  provisioner "shell" {
-    environment_vars = ["IMAGE_VERSION=${var.image_version}"]
-    script           = "provisioners/pull-docker-images.sh"
-  }
 }
diff --git a/scripts/packer/gcp-image-cuda.json b/scripts/packer/gcp-image-cuda.json
index 9ebb24e7c9..2d606a2b42 100644
--- a/scripts/packer/gcp-image-cuda.json
+++ b/scripts/packer/gcp-image-cuda.json
@@ -56,14 +56,6 @@
     {
       "type": "shell",
       "script": "provisioners/install-nvidia-container-toolkit.sh"
-    },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
     }
   ]
 }
diff --git a/scripts/packer/gcp-image.json b/scripts/packer/gcp-image.json
index 6e6ba64537..7b9fd2f950 100644
--- a/scripts/packer/gcp-image.json
+++ b/scripts/packer/gcp-image.json
@@ -46,14 +46,6 @@
         "cd /tmp",
         "chmod +x install-docker.sh",
         "./install-docker.sh --version {{user `docker_version`}}"]
-    },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
     }
   ]
 }
diff --git a/scripts/packer/oci-image-cuda.json b/scripts/packer/oci-image-cuda.json
index 7d7251bc0b..e0406d8738 100644
--- a/scripts/packer/oci-image-cuda.json
+++ b/scripts/packer/oci-image-cuda.json
@@ -65,14 +65,6 @@
     {
       "type": "shell",
       "script": "provisioners/install-nvidia-container-toolkit.sh"
-    },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
     }
   ]
 }
diff --git a/scripts/packer/oci-image.json b/scripts/packer/oci-image.json
index 742a8649a0..370cb4bf2d 100644
--- a/scripts/packer/oci-image.json
+++ b/scripts/packer/oci-image.json
@@ -55,14 +55,6 @@
         "cd /tmp",
         "chmod +x install-docker.sh",
         "./install-docker.sh --version {{user `docker_version`}}"]
-    },
-    {
-      "type": "shell",
-      "environment_vars": [
-        "IMAGE_REPO={{user `image_repo`}}",
-        "IMAGE_VERSION={{user `image_version`}}"
-      ],
-      "script": "provisioners/pull-docker-images.sh"
     }
   ]
 }
diff --git a/scripts/packer/provisioners/pull-docker-images.sh b/scripts/packer/provisioners/pull-docker-images.sh
deleted file mode 100644
index 4d9d826435..0000000000
--- a/scripts/packer/provisioners/pull-docker-images.sh
+++ /dev/null
@@ -1,18 +0,0 @@
-#!/bin/bash
-
-set -e
-
-IMAGES="
- dstackai/${IMAGE_REPO}:py3.13-${IMAGE_VERSION}-cuda-12.1
- dstackai/${IMAGE_REPO}:py3.12-${IMAGE_VERSION}-cuda-12.1
- dstackai/${IMAGE_REPO}:py3.11-${IMAGE_VERSION}-cuda-12.1
- dstackai/${IMAGE_REPO}:py3.10-${IMAGE_VERSION}-cuda-12.1
- dstackai/${IMAGE_REPO}:py3.9-${IMAGE_VERSION}-cuda-12.1
-"
-echo "START pull image"
-for img in $IMAGES; do
- docker pull --platform linux/amd64 $img
-done 
-echo "LIST installed images"
-docker image ls --all
-echo "END "
diff --git a/src/dstack/_internal/server/background/tasks/process_running_jobs.py b/src/dstack/_internal/server/background/tasks/process_running_jobs.py
index e05f98fd23..b834db39b9 100644
--- a/src/dstack/_internal/server/background/tasks/process_running_jobs.py
+++ b/src/dstack/_internal/server/background/tasks/process_running_jobs.py
@@ -1,4 +1,5 @@
 import asyncio
+import re
 from collections.abc import Iterable
 from datetime import timedelta, timezone
 from typing import Dict, List, Optional
@@ -7,6 +8,7 @@
 from sqlalchemy.ext.asyncio import AsyncSession
 from sqlalchemy.orm import joinedload
 
+from dstack._internal import settings
 from dstack._internal.core.consts import DSTACK_RUNNER_HTTP_PORT, DSTACK_SHIM_HTTP_PORT
 from dstack._internal.core.errors import GatewayError
 from dstack._internal.core.models.backends.base import BackendType
@@ -517,14 +519,14 @@ def _process_provisioning_with_shim(
         cpu = None
         memory = None
         network_mode = NetworkMode.HOST
-
+    image_name = _patch_base_image_for_aws_efa(job_spec, job_provisioning_data)
     if shim_client.is_api_v2_supported():
         shim_client.submit_task(
             task_id=job_model.id,
             name=job_model.job_name,
             registry_username=registry_username,
             registry_password=registry_password,
-            image_name=job_spec.image_name,
+            image_name=image_name,
             container_user=container_user,
             privileged=job_spec.privileged,
             gpu=gpu,
@@ -545,7 +547,7 @@ def _process_provisioning_with_shim(
         submitted = shim_client.submit(
             username=registry_username,
             password=registry_password,
-            image_name=job_spec.image_name,
+            image_name=image_name,
             privileged=job_spec.privileged,
             container_name=job_model.job_name,
             container_user=container_user,
@@ -969,3 +971,43 @@ def _get_instance_specific_gpu_devices(
             GPUDevice(path_on_host="/dev/nvidiactl", path_in_container="/dev/nvidiactl")
         )
     return gpu_devices
+
+
+def _patch_base_image_for_aws_efa(
+    job_spec: JobSpec, job_provisioning_data: JobProvisioningData
+) -> str:
+    image_name = job_spec.image_name
+
+    if job_provisioning_data.backend != BackendType.AWS:
+        return image_name
+
+    instance_type = job_provisioning_data.instance_type.name
+    efa_enabled_patterns = [
+        # TODO: p6-b200 isn't supported yet in gpuhunt
+        r"^p6-b200\.(48xlarge)$",
+        r"^p5\.(48xlarge)$",
+        r"^p5e\.(48xlarge)$",
+        r"^p5en\.(48xlarge)$",
+        r"^p4d\.(24xlarge)$",
+        r"^p4de\.(24xlarge)$",
+        r"^g6\.(8xlarge|12xlarge|16xlarge|24xlarge|48xlarge)$",
+        r"^g6e\.(8xlarge|12xlarge|16xlarge|24xlarge|48xlarge)$",
+        r"^gr6\.8xlarge$",
+        r"^g5\.(8xlarge|12xlarge|16xlarge|24xlarge|48xlarge)$",
+        r"^g4dn\.(8xlarge|12xlarge|16xlarge|metal)$",
+        r"^p3dn\.(24xlarge)$",
+    ]
+
+    is_efa_enabled = any(re.match(pattern, instance_type) for pattern in efa_enabled_patterns)
+    if not is_efa_enabled:
+        return image_name
+
+    if not image_name.startswith(f"{settings.DSTACK_BASE_IMAGE}:"):
+        return image_name
+
+    if image_name.endswith(f"-base-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"):
+        return image_name[:-17] + f"-devel-efa-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+    elif image_name.endswith(f"-devel-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"):
+        return image_name[:-18] + f"-devel-efa-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+
+    return image_name
diff --git a/src/dstack/_internal/server/services/jobs/configurators/base.py b/src/dstack/_internal/server/services/jobs/configurators/base.py
index e3c7b89ee4..6ef0ca7712 100644
--- a/src/dstack/_internal/server/services/jobs/configurators/base.py
+++ b/src/dstack/_internal/server/services/jobs/configurators/base.py
@@ -50,11 +50,15 @@ def get_default_python_verison() -> str:
         )
 
 
-def get_default_image(python_version: str, nvcc: bool = False) -> str:
-    suffix = ""
-    if nvcc:
-        suffix = "-devel"
-    return f"{settings.DSTACK_BASE_IMAGE}:py{python_version}-{settings.DSTACK_BASE_IMAGE_VERSION}-cuda-12.1{suffix}"
+def get_default_image(nvcc: bool = False) -> str:
+    """
+    Note: May be overridden by dstack (e.g., EFA-enabled version for AWS EFA-capable instances).
+    See `dstack._internal.server.background.tasks.process_running_jobs._patch_base_image_for_aws_efa` for details.
+
+    Args:
+        nvcc: If True, returns 'devel' variant, otherwise 'base'.
+    """
+    return f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}-{'devel' if nvcc else 'base'}-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
 
 
 class JobConfigurator(ABC):
@@ -173,7 +177,7 @@ def _dstack_image_commands(self) -> List[str]:
         ):
             return []
         return [
-            f"uv venv --prompt workflow --seed {DEFAULT_REPO_DIR}/.venv > /dev/null 2>&1",
+            f"uv venv --python {self._python()} --prompt workflow --seed {DEFAULT_REPO_DIR}/.venv > /dev/null 2>&1",
             f"echo 'source {DEFAULT_REPO_DIR}/.venv/bin/activate' >> ~/.bashrc",
             f"source {DEFAULT_REPO_DIR}/.venv/bin/activate",
         ]
@@ -199,7 +203,7 @@ def _home_dir(self) -> Optional[str]:
     def _image_name(self) -> str:
         if self.run_spec.configuration.image is not None:
             return self.run_spec.configuration.image
-        return get_default_image(self._python(), nvcc=bool(self.run_spec.configuration.nvcc))
+        return get_default_image(nvcc=bool(self.run_spec.configuration.nvcc))
 
     async def _user(self) -> Optional[UnixUser]:
         user = self.run_spec.configuration.user
diff --git a/src/dstack/_internal/settings.py b/src/dstack/_internal/settings.py
index 92a7326a0c..2636a3b362 100644
--- a/src/dstack/_internal/settings.py
+++ b/src/dstack/_internal/settings.py
@@ -14,6 +14,9 @@
 
 DSTACK_BASE_IMAGE = os.getenv("DSTACK_BASE_IMAGE", "dstackai/base")
 DSTACK_BASE_IMAGE_VERSION = os.getenv("DSTACK_BASE_IMAGE_VERSION", version.base_image)
+DSTACK_BASE_IMAGE_UBUNTU_VERSION = os.getenv(
+    "DSTACK_BASE_IMAGE_UBUNTU_VERSION", version.base_image_ubuntu_version
+)
 
 
 class FeatureFlags:
diff --git a/src/dstack/version.py b/src/dstack/version.py
index c71d1a7665..d4255a2301 100644
--- a/src/dstack/version.py
+++ b/src/dstack/version.py
@@ -1,3 +1,4 @@
 __version__ = "0.0.0"
 __is_release__ = False
-base_image = "0.9"
+base_image = "0.10"
+base_image_ubuntu_version = "22.04"
diff --git a/src/tests/_internal/server/background/tasks/test_process_running_jobs.py b/src/tests/_internal/server/background/tasks/test_process_running_jobs.py
index d842441fe9..59a08ddc4d 100644
--- a/src/tests/_internal/server/background/tasks/test_process_running_jobs.py
+++ b/src/tests/_internal/server/background/tasks/test_process_running_jobs.py
@@ -7,16 +7,21 @@
 from freezegun import freeze_time
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from dstack._internal import settings
 from dstack._internal.core.errors import SSHError
 from dstack._internal.core.models.backends.base import BackendType
 from dstack._internal.core.models.common import NetworkMode
 from dstack._internal.core.models.configurations import DevEnvironmentConfiguration
-from dstack._internal.core.models.instances import InstanceStatus
+from dstack._internal.core.models.instances import InstanceStatus, InstanceType
 from dstack._internal.core.models.profiles import StartupOrder, UtilizationPolicy
+from dstack._internal.core.models.resources import ResourcesSpec
 from dstack._internal.core.models.runs import (
+    JobProvisioningData,
     JobRuntimeData,
+    JobSpec,
     JobStatus,
     JobTerminationReason,
+    Requirements,
     RunStatus,
 )
 from dstack._internal.core.models.volumes import (
@@ -24,8 +29,11 @@
     VolumeMountPoint,
     VolumeStatus,
 )
-from dstack._internal.server import settings
-from dstack._internal.server.background.tasks.process_running_jobs import process_running_jobs
+from dstack._internal.server import settings as server_settings
+from dstack._internal.server.background.tasks.process_running_jobs import (
+    _patch_base_image_for_aws_efa,
+    process_running_jobs,
+)
 from dstack._internal.server.schemas.runner import (
     HealthcheckResponse,
     JobStateEvent,
@@ -221,7 +229,7 @@ async def test_updates_running_job(self, test_db, session: AsyncSession, tmp_pat
             patch(
                 "dstack._internal.server.services.runner.client.RunnerClient"
             ) as RunnerClientMock,
-            patch.object(settings, "SERVER_DIR_PATH", tmp_path),
+            patch.object(server_settings, "SERVER_DIR_PATH", tmp_path),
         ):
             runner_client_mock = RunnerClientMock.return_value
             runner_client_mock.pull.return_value = PullResponse(
@@ -330,7 +338,7 @@ async def test_provisioning_shim_with_volumes(
             name="test-run-0-0",
             registry_username="",
             registry_password="",
-            image_name="dstackai/base:py3.13-0.9-cuda-12.1",
+            image_name="dstackai/base:0.10-base-ubuntu22.04",
             container_user="root",
             privileged=privileged,
             gpu=None,
@@ -878,3 +886,129 @@ async def test_master_job_waits_for_workers(self, test_db, session: AsyncSession
             await process_running_jobs()
         await session.refresh(master_job)
         assert master_job.status == JobStatus.RUNNING
+
+
+class TestPatchBaseImageForAwsEfa:
+    @staticmethod
+    def _create_job_spec(image_name: str) -> "JobSpec":
+        return JobSpec(
+            job_num=0,
+            job_name="test-job",
+            commands=["echo hello"],
+            env={},
+            image_name=image_name,
+            requirements=Requirements(resources=ResourcesSpec()),
+        )
+
+    @staticmethod
+    def _create_job_provisioning_data_with_instance_type(
+        backend: BackendType, instance_type: str
+    ) -> JobProvisioningData:
+        job_provisioning_data = get_job_provisioning_data(backend=backend)
+        job_provisioning_data.instance_type = InstanceType(
+            name=instance_type,
+            resources=job_provisioning_data.instance_type.resources,
+        )
+        return job_provisioning_data
+
+    @staticmethod
+    def _call_patch_base_image_for_aws_efa(
+        image_name: str, backend: BackendType, instance_type: str
+    ) -> str:
+        job_spec = TestPatchBaseImageForAwsEfa._create_job_spec(image_name)
+        job_provisioning_data = (
+            TestPatchBaseImageForAwsEfa._create_job_provisioning_data_with_instance_type(
+                backend, instance_type
+            )
+        )
+        return _patch_base_image_for_aws_efa(job_spec, job_provisioning_data)
+
+    @pytest.mark.parametrize(
+        "suffix,instance_type",
+        [
+            ("-base", "p6-b200.48xlarge"),
+            ("-devel", "p5.48xlarge"),
+        ],
+    )
+    def test_patch_aws_efa_instance_with_suffix(self, suffix: str, instance_type: str):
+        image_name = f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}{suffix}-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+        result = self._call_patch_base_image_for_aws_efa(
+            image_name, BackendType.AWS, instance_type
+        )
+        expected = f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}-devel-efa-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+        assert result == expected
+
+    @pytest.mark.parametrize("suffix", ["-base", "-devel"])
+    @pytest.mark.parametrize(
+        "instance_type",
+        [
+            "p5.48xlarge",
+            "p5e.48xlarge",
+            "p4d.24xlarge",
+            "p4de.24xlarge",
+            "g6.8xlarge",
+            "g6e.8xlarge",
+        ],
+    )
+    def test_patch_all_efa_instance_types(self, instance_type: str, suffix: str):
+        image_name = f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}{suffix}-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+        result = self._call_patch_base_image_for_aws_efa(
+            image_name, BackendType.AWS, instance_type
+        )
+        expected = f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}-devel-efa-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+        assert result == expected
+
+    @pytest.mark.parametrize("suffix", ["-base", "-devel"])
+    @pytest.mark.parametrize(
+        "backend",
+        [BackendType.GCP, BackendType.AZURE, BackendType.LAMBDA, BackendType.LOCAL],
+    )
+    @pytest.mark.parametrize(
+        "instance_type",
+        [
+            "standard-4",
+            "p5.xlarge",
+            "p6.2xlarge",
+            "g6.xlarge",
+        ],  # Mix of generic and EFA-named types
+    )
+    def test_no_patch_non_aws_backends(
+        self, backend: BackendType, suffix: str, instance_type: str
+    ):
+        image_name = f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}{suffix}-ubuntu{settings.DSTACK_BASE_IMAGE_UBUNTU_VERSION}"
+        result = self._call_patch_base_image_for_aws_efa(image_name, backend, instance_type)
+        assert result == image_name
+
+    @pytest.mark.parametrize("suffix", ["-base", "-devel"])
+    @pytest.mark.parametrize(
+        "instance_type",
+        ["t3.micro", "m5.large", "c5.xlarge", "r5.2xlarge", "m6i.large", "g6.xlarge"],
+    )
+    def test_no_patch_non_efa_aws_instances(self, instance_type: str, suffix: str):
+        image_name = f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}{suffix}"
+        result = self._call_patch_base_image_for_aws_efa(
+            image_name, BackendType.AWS, instance_type
+        )
+        assert result == image_name
+
+    @pytest.mark.parametrize(
+        "instance_type",
+        ["p5.xlarge", "p6.2xlarge", "t3.micro", "m5.large"],  # Mix of EFA and non-EFA instances
+    )
+    @pytest.mark.parametrize(
+        "image_name",
+        [
+            "ubuntu:20.04",
+            "nvidia/cuda:11.8-runtime-ubuntu20.04",
+            "python:3.9-slim",
+            "custom/image:latest",
+            f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}-custom",
+            f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}-devel-efa",
+            f"{settings.DSTACK_BASE_IMAGE}:{settings.DSTACK_BASE_IMAGE_VERSION}",
+        ],
+    )
+    def test_no_patch_other_images(self, instance_type: str, image_name: str):
+        result = self._call_patch_base_image_for_aws_efa(
+            image_name, BackendType.AWS, instance_type
+        )
+        assert result == image_name
diff --git a/src/tests/_internal/server/routers/test_runs.py b/src/tests/_internal/server/routers/test_runs.py
index 4a40c3273f..3c2181a209 100644
--- a/src/tests/_internal/server/routers/test_runs.py
+++ b/src/tests/_internal/server/routers/test_runs.py
@@ -170,7 +170,7 @@ def get_dev_env_run_plan_dict(
                         "/bin/bash",
                         "-i",
                         "-c",
-                        "uv venv --prompt workflow --seed /workflow/.venv > /dev/null 2>&1"
+                        "uv venv --python 3.13 --prompt workflow --seed /workflow/.venv > /dev/null 2>&1"
                         " && echo 'source /workflow/.venv/bin/activate' >> ~/.bashrc"
                         " && source /workflow/.venv/bin/activate"
                         " && (echo pip install ipykernel... && "
@@ -188,7 +188,7 @@ def get_dev_env_run_plan_dict(
                     ],
                     "env": {},
                     "home_dir": "/root",
-                    "image_name": "dstackai/base:py3.13-0.9-cuda-12.1",
+                    "image_name": "dstackai/base:0.10-base-ubuntu22.04",
                     "user": None,
                     "privileged": privileged,
                     "job_name": f"{run_name}-0-0",
@@ -334,7 +334,7 @@ def get_dev_env_run_dict(
                         "/bin/bash",
                         "-i",
                         "-c",
-                        "uv venv --prompt workflow --seed /workflow/.venv > /dev/null 2>&1"
+                        "uv venv --python 3.13 --prompt workflow --seed /workflow/.venv > /dev/null 2>&1"
                         " && echo 'source /workflow/.venv/bin/activate' >> ~/.bashrc"
                         " && source /workflow/.venv/bin/activate"
                         " && (echo pip install ipykernel... && "
@@ -352,7 +352,7 @@ def get_dev_env_run_dict(
                     ],
                     "env": {},
                     "home_dir": "/root",
-                    "image_name": "dstackai/base:py3.13-0.9-cuda-12.1",
+                    "image_name": "dstackai/base:0.10-base-ubuntu22.04",
                     "user": None,
                     "privileged": privileged,
                     "job_name": f"{run_name}-0-0",
diff --git a/src/tests/_internal/server/services/jobs/configurators/test_task.py b/src/tests/_internal/server/services/jobs/configurators/test_task.py
index 07c8da7328..e954e6a01f 100644
--- a/src/tests/_internal/server/services/jobs/configurators/test_task.py
+++ b/src/tests/_internal/server/services/jobs/configurators/test_task.py
@@ -88,7 +88,7 @@ async def test_with_commands_and_image(self, shell: Optional[str], expected_shel
         ],
     )
     async def test_with_commands_no_image(self, shell: Optional[str], expected_shell: str):
-        configuration = TaskConfiguration(commands=["sleep inf"], shell=shell)
+        configuration = TaskConfiguration(python="3.12", commands=["sleep inf"], shell=shell)
         run_spec = get_run_spec(run_name="run", repo_id="id", configuration=configuration)
         configurator = TaskJobConfigurator(run_spec)
 
@@ -98,7 +98,7 @@ async def test_with_commands_no_image(self, shell: Optional[str], expected_shell
             expected_shell,
             "-i",
             "-c",
-            "uv venv --prompt workflow --seed /workflow/.venv > /dev/null 2>&1"
+            "uv venv --python 3.12 --prompt workflow --seed /workflow/.venv > /dev/null 2>&1"
             " && echo 'source /workflow/.venv/bin/activate' >> ~/.bashrc"
             " && source /workflow/.venv/bin/activate"
             " && sleep inf",