diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 8a2ba9f841434..f7f2c110ef186 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -108,7 +108,7 @@ jobs:
           - platforms: linux/amd64,linux/arm64
             device: cpu
 
-          - platforms: linux/amd64
+          - platforms: linux/amd64,linux/arm64
             device: cuda
             suffix: -cuda
 
diff --git a/machine-learning/Dockerfile b/machine-learning/Dockerfile
index 12fb183c953d4..348b418e4d86b 100644
--- a/machine-learning/Dockerfile
+++ b/machine-learning/Dockerfile
@@ -17,7 +17,7 @@ RUN mkdir /opt/armnn && \
 
 FROM builder-${DEVICE} AS builder
 
-ARG DEVICE
+ARG DEVICE TARGETARCH
 ENV PYTHONDONTWRITEBYTECODE=1 \
     PYTHONUNBUFFERED=1 \
     PIP_NO_CACHE_DIR=true \
@@ -32,7 +32,11 @@ RUN poetry config installer.max-workers 10 && \
 RUN python3 -m venv /opt/venv
 
 COPY poetry.lock pyproject.toml ./
-RUN poetry install --sync --no-interaction --no-ansi --no-root --with ${DEVICE} --without dev
+RUN if [ "$DEVICE" = "cuda" ] && [ "$TARGETARCH" = "arm64" ]; then \
+    # hack to work around poetry not setting the right filename for the wheel https://github.com/python-poetry/poetry/issues/4472
+    wget -q -O onnxruntime_gpu-1.18.0-cp311-cp311-manylinux_aarch64.whl https://nvidia.box.com/shared/static/fy55jvniujjbigr4gwkv8z1ma6ipgspg.whl; fi && \
+    poetry install --sync --no-interaction --no-ansi --no-root --with ${DEVICE} --without dev && \
+    if [ "$DEVICE" = "cuda" ] && [ "$TARGETARCH" = "arm64" ]; then rm onnxruntime_gpu-1.18.0-cp311-cp311-manylinux_aarch64.whl; fi
 
 FROM python:3.11-slim-bookworm@sha256:ed4e985674f478c90ce879e9aa224fbb772c84e39b4aed5155b9e2280f131039 AS prod-cpu
 
@@ -49,7 +53,20 @@ RUN apt-get update && \
     apt-get remove wget -yqq && \
     rm -rf /var/lib/apt/lists/*
 
-FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04@sha256:fa44193567d1908f7ca1f3abf8623ce9c63bc8cba7bcfdb32702eb04d326f7a8 AS prod-cuda
+FROM nvidia/cuda:12.2.2-runtime-ubuntu22.04 AS prod-cuda-amd64
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends libcudnn9-cuda-12 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+FROM nvidia/cuda:12.2.2-runtime-ubuntu22.04 AS prod-cuda-arm64
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends libcudnn8 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+ENV LD_LIBRARY_PATH=/usr/local/cuda-12/compat:$LD_LIBRARY_PATH
+
+FROM prod-cuda-${TARGETARCH} AS prod-cuda
 
 COPY --from=builder-cuda /usr/local/bin/python3 /usr/local/bin/python3
 COPY --from=builder-cuda /usr/local/lib/python3.11 /usr/local/lib/python3.11
@@ -76,10 +93,10 @@ COPY --from=builder-armnn \
     /opt/armnn/
 
 FROM prod-${DEVICE} AS prod
-ARG DEVICE
+ARG DEVICE TARGETARCH
 
 RUN apt-get update && \
-    apt-get install -y --no-install-recommends tini $(if ! [ "$DEVICE" = "openvino" ]; then echo "libmimalloc2.0"; fi) && \
+    apt-get install -y --no-install-recommends tini $(if ! { [ "$DEVICE" = "openvino" ] || { [ "$DEVICE" = "cuda" ] && [ "$TARGETARCH" = "arm64" ]; }; }; then echo "libmimalloc2.0"; fi) && \
     apt-get autoremove -yqq && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
diff --git a/machine-learning/poetry.lock b/machine-learning/poetry.lock
index bd09bd8469e67..13e9f35c5420c 100644
--- a/machine-learning/poetry.lock
+++ b/machine-learning/poetry.lock
@@ -2003,6 +2003,28 @@ packaging = "*"
 protobuf = "*"
 sympy = "*"
 
+[[package]]
+name = "onnxruntime-gpu"
+version = "1.18.0"
+description = "ONNX Runtime is a runtime accelerator for Machine Learning models"
+optional = false
+python-versions = "*"
+files = [
+    {file = "onnxruntime_gpu-1.18.0-cp311-cp311-manylinux_aarch64.whl", hash = "sha256:7bdd6c373611235e43c8707fa528539327ff17a969448adf956ddf177d5fc8e7"},
+]
+
+[package.dependencies]
+coloredlogs = "*"
+flatbuffers = "*"
+numpy = ">=1.26.4"
+packaging = "*"
+protobuf = "*"
+sympy = "*"
+
+[package.source]
+type = "file"
+url = "onnxruntime_gpu-1.18.0-cp311-cp311-manylinux_aarch64.whl"
+
 [[package]]
 name = "onnxruntime-gpu"
 version = "1.18.1"
@@ -2636,7 +2658,6 @@ files = [
     {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"},
-    {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"},
     {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"},
     {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"},
     {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"},
@@ -3607,4 +3628,4 @@ testing = ["coverage (>=5.0.3)", "zope.event", "zope.testing"]
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.10,<4.0"
-content-hash = "b2b053886ca1dd3a3305c63caf155b1976dfc4066f72f5d1ecfc42099db34aab"
+content-hash = "87e047b423bdf74fa1825e509d6a31d2bac57e8e7d4419cfefcc86aa98b139a2"
diff --git a/machine-learning/pyproject.toml b/machine-learning/pyproject.toml
index a69fb33a8d50e..c33868f13809e 100644
--- a/machine-learning/pyproject.toml
+++ b/machine-learning/pyproject.toml
@@ -4,7 +4,7 @@ version = "1.114.0"
 description = ""
 authors = ["Hau Tran <alex.tran1502@gmail.com>"]
 readme = "README.md"
-packages = [{include = "app"}]
+packages = [{ include = "app" }]
 
 [tool.poetry.dependencies]
 python = ">=3.10,<4.0"
@@ -12,7 +12,7 @@ insightface = ">=0.7.3,<1.0"
 opencv-python-headless = ">=4.7.0.72,<5.0"
 pillow = ">=9.5.0,<11.0"
 fastapi-slim = ">=0.95.2,<1.0"
-uvicorn = {extras = ["standard"], version = ">=0.22.0,<1.0"}
+uvicorn = { extras = ["standard"], version = ">=0.22.0,<1.0" }
 pydantic = "^1.10.8"
 aiocache = ">=0.12.1,<1.0"
 rich = ">=13.4.2"
@@ -45,7 +45,10 @@ onnxruntime = "^1.15.0"
 optional = true
 
 [tool.poetry.group.cuda.dependencies]
-onnxruntime-gpu = {version = "^1.17.0", source = "cuda12"}
+onnxruntime-gpu = [
+    { version = "^1.17.0", source = "cuda12", markers = "platform_machine == 'x86_64'" },
+    { python = "3.11", path = "onnxruntime_gpu-1.18.0-cp311-cp311-manylinux_aarch64.whl", markers = "platform_machine == 'aarch64'" }
+]
 
 [tool.poetry.group.openvino]
 optional = true
diff --git a/machine-learning/start.sh b/machine-learning/start.sh
index c3fda523df832..9cddd6d6eb2bc 100755
--- a/machine-learning/start.sh
+++ b/machine-learning/start.sh
@@ -1,19 +1,26 @@
 #!/usr/bin/env sh
 
-lib_path="/usr/lib/$(arch)-linux-gnu/libmimalloc.so.2"
 # mimalloc seems to increase memory usage dramatically with openvino, need to investigate
-if ! [ "$DEVICE" = "openvino" ]; then
+mimalloc="/usr/lib/$(arch)-linux-gnu/libmimalloc.so.2"
+if [ -f "$mimalloc" ]; then
+	export LD_PRELOAD="$mimalloc"
+fi
+
+if { [ "$DEVICE" = "cuda" ] && [ "$(arch)" = "aarch64" ]; }; then
+	lib_path="/usr/lib/$(arch)-linux-gnu/libmimalloc.so.2"
 	export LD_PRELOAD="$lib_path"
-	export LD_BIND_NOW=1
-	: "${MACHINE_LEARNING_WORKER_TIMEOUT:=120}"
-else
-	: "${MACHINE_LEARNING_WORKER_TIMEOUT:=300}"
 fi
+export LD_BIND_NOW=1
 
 : "${IMMICH_HOST:=[::]}"
 : "${IMMICH_PORT:=3003}"
 : "${MACHINE_LEARNING_WORKERS:=1}"
 : "${MACHINE_LEARNING_HTTP_KEEPALIVE_TIMEOUT_S:=2}"
+if [ "$DEVICE" = "openvino" ]; then
+	: "${MACHINE_LEARNING_WORKER_TIMEOUT:=300}"
+else
+	: "${MACHINE_LEARNING_WORKER_TIMEOUT:=120}"
+fi
 
 gunicorn app.main:app \
 	-k app.config.CustomUvicornWorker \