diff --git a/.github/workflows/linux_cuda.yaml b/.github/workflows/linux_cuda.yaml
index 1a5e042..f8fa8ce 100644
--- a/.github/workflows/linux_cuda.yaml
+++ b/.github/workflows/linux_cuda.yaml
@@ -71,7 +71,7 @@ jobs:
 
       - uses: docker/build-push-action@v5
         with:
-          context: ./docker
+          context: ./docker/cuda${{ inputs.cuda}}
           push: true
           target: ${{ steps.build.outputs.target }}
           tags: scverse/scvi-tools:py${{ inputs.python }}-cu${{ inputs.cuda }}-${{ steps.build.outputs.dependencies }}${{ steps.build.outputs.versiontag }}
diff --git a/.github/workflows/linux_cuda_base.yaml b/.github/workflows/linux_cuda_base.yaml
index d6cab7f..f1df4ba 100644
--- a/.github/workflows/linux_cuda_base.yaml
+++ b/.github/workflows/linux_cuda_base.yaml
@@ -11,7 +11,7 @@ jobs:
       fail-fast: false
       matrix:
         python: ["3.9", "3.10", "3.11"]
-        cuda: ["11.8", "12.1"]
+        cuda: ["11", "12"]
         version: ["cuda"]
         dependencies: ["base"]
 
diff --git a/.github/workflows/linux_cuda_latest_scheduled.yaml b/.github/workflows/linux_cuda_latest_scheduled.yaml
index 3366a33..c9a6f3d 100644
--- a/.github/workflows/linux_cuda_latest_scheduled.yaml
+++ b/.github/workflows/linux_cuda_latest_scheduled.yaml
@@ -11,7 +11,7 @@ jobs:
       fail-fast: false
       matrix:
         python: ["3.9", "3.10", "3.11"]
-        cuda: ["11.8", "12.1"]
+        cuda: ["11", "12"]
         version: ["latest"]
         dependencies: ["", "dev", "tutorials"]
 
diff --git a/.github/workflows/linux_cuda_manual.yaml b/.github/workflows/linux_cuda_manual.yaml
index f87e0bc..1bc90b5 100644
--- a/.github/workflows/linux_cuda_manual.yaml
+++ b/.github/workflows/linux_cuda_manual.yaml
@@ -20,7 +20,7 @@ jobs:
       fail-fast: false
       matrix:
         python: ["3.9", "3.10", "3.11"]
-        cuda: ["11.8", "12.1"]
+        cuda: ["11", "12"]
 
     uses: ./.github/workflows/linux_cuda.yaml
     with:
diff --git a/.github/workflows/linux_cuda_release.yaml b/.github/workflows/linux_cuda_release.yaml
new file mode 100644
index 0000000..776a562
--- /dev/null
+++ b/.github/workflows/linux_cuda_release.yaml
@@ -0,0 +1,50 @@
+name: Linux (CUDA, manual)
+
+on:
+  workflow_dispatch:
+    inputs:
+      version:
+        description: "semver"
+        required: true
+        default: "latest"
+        type: string
+
+      dependencies:
+        description: "optional dependencies"
+        required: false
+        type: string
+
+jobs:
+  build-semver:
+    strategy:
+      fail-fast: false
+      matrix:
+        python: ["3.9", "3.10", "3.11"]
+        cuda: ["11", "12"]
+        dependencies: ["", "dev", "tutorials"]
+
+    uses: ./.github/workflows/linux_cuda.yaml
+    with:
+      python: ${{ matrix.python }}
+      cuda: ${{ matrix.cuda }}
+      version: ${{ inputs.version }}
+      dependencies: ${{ matrix.dependencies }}
+
+    secrets: inherit
+
+  build-stable:
+    strategy:
+      fail-fast: false
+      matrix:
+        python: ["3.9", "3.10", "3.11"]
+        cuda: ["11", "12"]
+        dependencies: ["", "dev", "tutorials"]
+
+    uses: ./.github/workflows/linux_cuda.yaml
+    with:
+      python: ${{ matrix.python }}
+      cuda: ${{ matrix.cuda }}
+      version: stable
+      dependencies: ${{ matrix.dependencies }}
+
+    secrets: inherit
diff --git a/docker/cuda11/Dockerfile b/docker/cuda11/Dockerfile
new file mode 100644
index 0000000..2692d78
--- /dev/null
+++ b/docker/cuda11/Dockerfile
@@ -0,0 +1,56 @@
+FROM condaforge/miniforge3:latest AS base
+
+RUN head -n -1 "${HOME}/.bashrc" > "${HOME}/tmp.bashrc" ; \
+    mv "${HOME}/tmp.bashrc" "${HOME}/.bashrc"
+
+ARG PYTHON_VERSION
+RUN mamba create --copy -y -p /env \
+    python=${PYTHON_VERSION} -c conda-forge && \
+    mamba clean -ayf
+
+RUN mamba install -y -p /env \
+    gcc cxx-compiler -c conda-forge && \
+    mamba clean -ayf
+
+ARG CUDA_VERSION=11
+RUN CONDA_OVERRIDE_CUDA=${CUDA_VERSION} mamba install -y -p /env \
+    pytorch torchvision torchaudio pytorch-cuda=${CUDA_VERSION} \
+    -c pytorch -c nvidia && \
+    mamba clean -ayf
+
+ARG CUDA_VERSION=11
+RUN mamba run -p /env \
+    python -m pip install --no-cache-dir -U "jax[cuda${CUDA_VERSION}_pip]" \
+    -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html && \
+    mamba run -p /env \
+    python -m pip cache purge
+
+ENV PATH "/env/bin:$PATH"
+
+RUN conda init --reverse && \
+    rm -rf $(conda info --base) && \
+    rm -f "${HOME}/.condarc" && \
+    rm -rf ${HOME}/.conda && \
+    rm -rf /opt/conda
+
+FROM base AS latest
+ARG DEPENDENCIES=""
+RUN pip install --no-cache-dir "scvi-tools[${DEPENDENCIES}] @ git+https://github.com/scverse/scvi-tools" && \
+    pip cache purge
+
+FROM base AS stable
+ARG DEPENDENCIES=""
+RUN pip install -U --no-cache-dir "scvi-tools[${DEPENDENCIES}]" && \
+    pip cache purge
+
+FROM base AS branch
+ARG SCVI_TOOLS_VERSION=main
+ARG DEPENDENCIES=""
+RUN pip install --no-cache-dir "scvi-tools[${DEPENDENCIES}] @ git+https://github.com/scverse/scvi-tools@${SCVI_TOOLS_VERSION}" && \
+    pip cache purge
+
+FROM base AS semver
+ARG SCVI_TOOLS_VERSION=latest
+ARG DEPENDENCIES=""
+RUN pip install --no-cache-dir "scvi-tools[${DEPENDENCIES}]==${SCVI_TOOLS_VERSION}" && \
+    pip cache purge
diff --git a/docker/Dockerfile b/docker/cuda12/Dockerfile
similarity index 98%
rename from docker/Dockerfile
rename to docker/cuda12/Dockerfile
index 8d92aad..79b00ad 100644
--- a/docker/Dockerfile
+++ b/docker/cuda12/Dockerfile
@@ -8,16 +8,16 @@ RUN mamba create --copy -y -p /env \
     python=${PYTHON_VERSION} -c conda-forge && \
     mamba clean -ayf
 
-ARG CUDA_VERSION
+RUN mamba install -y -p /env \
+    gcc cxx-compiler -c conda-forge && \
+    mamba clean -ayf
+
+ARG CUDA_VERSION=12
 RUN CONDA_OVERRIDE_CUDA=${CUDA_VERSION} mamba install -y -p /env \
     pytorch torchvision torchaudio pytorch-cuda=${CUDA_VERSION} \
     jaxlib=*=*cuda* jax cuda-nvcc -c pytorch -c conda-forge -c nvidia && \
     mamba clean -ayf
 
-RUN mamba install -y -p /env \
-    gcc cxx-compiler -c conda-forge && \
-    mamba clean -ayf
-
 ENV PATH "/env/bin:$PATH"
 
 RUN conda init --reverse && \