diff --git a/.github/workflows/linux_cuda_aarch64_wheel.yaml b/.github/workflows/linux_cuda_aarch64_wheel.yaml
index 3be5539b0..bfb55fdca 100644
--- a/.github/workflows/linux_cuda_aarch64_wheel.yaml
+++ b/.github/workflows/linux_cuda_aarch64_wheel.yaml
@@ -56,3 +56,80 @@ jobs:
       architecture: aarch64
       build-platform: "python-build-package"
       build-command: "BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1 ENABLE_CUDA=1 python -m build --wheel -vvv --no-isolation"
+
+  install-and-test:
+    runs-on: linux.arm64.2xlarge
+    container:
+      image: pytorch/manylinuxaarch64-builder:cuda12.6
+    env:
+      cuda_version_without_periods: "126"
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ['3.10']
+        ffmpeg-version-for-tests: ['4.4.2', '5.1.2', '6.1.1', '7.0.1', '8.0']
+    needs: build
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: meta-pytorch_torchcodec__${{ matrix.python-version }}_cu${{ env.cuda_version_without_periods }}_aarch64
+          path: pytorch/torchcodec/dist/
+      - name: Setup conda env
+        uses: conda-incubator/setup-miniconda@v3
+        with:
+          auto-update-conda: true
+          # Using miniforge instead of miniconda ensures that the default
+          # conda channel is conda-forge instead of main/default. This ensures
+          # ABI consistency between dependencies:
+          # https://conda-forge.org/docs/user/transitioning_from_defaults/
+          miniforge-version: latest
+          activate-environment: test
+          python-version: ${{ matrix.python-version }}
+      - name: Update pip
+        run: python -m pip install --upgrade pip
+      - name: Install PyTorch
+        run: |
+          ${CONDA_RUN} python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu${{ env.cuda_version_without_periods }}
+      - name: Install torchcodec from the wheel
+        run: |
+          wheel_path=`find pytorch/torchcodec/dist -type f -name "*.whl"`
+          echo Installing $wheel_path
+          python -m pip install $wheel_path -vvv
+
+      - name: Check out repo
+        uses: actions/checkout@v3
+      - name: Install ffmpeg, post build
+        run: |
+          # Ideally we would have checked for that before installing the wheel,
+          # but we need to checkout the repo to access this file, and we don't
+          # want to checkout the repo before installing the wheel to avoid any
+          # side-effect. It's OK.
+          source packaging/helpers.sh
+          assert_ffmpeg_not_installed
+
+          conda install "ffmpeg=${{ matrix.ffmpeg-version-for-tests }}" -c conda-forge
+          ffmpeg -version
+          echo LD_LIBRARY_PATH=$CONDA_PREFIX/lib:/usr/local/cuda/lib64/:${LD_LIBRARY_PATH} >> $GITHUB_ENV
+
+      - name: Install test dependencies
+        run: |
+          # Ideally we would find a way to get those dependencies from pyproject.toml
+          python -m pip install numpy pytest pillow
+
+      - name: Delete the src/ folder just for fun
+        run: |
+          # The only reason we checked-out the repo is to get access to the
+          # tests. We don't care about the rest. Out of precaution, we delete
+          # the src/ folder to be extra sure that we're running the code from
+          # the installed wheel rather than from the source.
+          # This is just to be extra cautious and very overkill because a)
+          # there's no way the `torchcodec` package from src/ can be found from
+          # the PythonPath: the main point of `src/` is precisely to protect
+          # against that and b) if we ever were to execute code from
+          # `src/torchcodec`, it would fail loudly because the built .so files
+          # aren't present there.
+          rm -r src/
+          ls
+      - name: Run Python tests
+        run: |
+          pytest --override-ini="addopts=-v" test
diff --git a/test/test_encoders.py b/test/test_encoders.py
index bc781bb83..e78fbd630 100644
--- a/test/test_encoders.py
+++ b/test/test_encoders.py
@@ -1,6 +1,7 @@
 import io
 import json
 import os
+import platform
 import re
 import subprocess
 import sys
@@ -328,7 +329,11 @@ def test_against_cli(
 
         assert_close = torch.testing.assert_close
         if sample_rate != asset.sample_rate:
-            rtol, atol = 0, 1e-3
+            if platform.machine().lower() == "aarch64":
+                rtol, atol = 0, 1e-2
+            else:
+                rtol, atol = 0, 1e-3
+
             if sys.platform == "darwin":
                 assert_close = partial(assert_tensor_close_on_at_least, percentage=99)
         elif format == "wav":
diff --git a/test/utils.py b/test/utils.py
index 6f4c77b76..ef6e7ec4d 100644
--- a/test/utils.py
+++ b/test/utils.py
@@ -2,6 +2,7 @@
 import json
 import os
 import pathlib
+import platform
 import subprocess
 import sys
 
@@ -138,7 +139,7 @@ def psnr(a, b, max_val=255) -> float:
 # not guarantee bit-for-bit equality across systems and architectures, so we
 # also cannot. We currently use Linux on x86_64 as our reference system.
 def assert_frames_equal(*args, **kwargs):
-    if sys.platform == "linux":
+    if sys.platform == "linux" and "x86" in platform.machine().lower():
         if args[0].device.type == "cuda":
             atol = 3 if cuda_version_used_for_building_torch() >= (13, 0) else 2
             if get_ffmpeg_major_version() == 4:
@@ -150,6 +151,7 @@ def assert_frames_equal(*args, **kwargs):
         else:
             torch.testing.assert_close(*args, **kwargs, atol=0, rtol=0)
     else:
+        # Here: Windows, MacOS, and Linux for non-x86 architectures like aarch64
         torch.testing.assert_close(*args, **kwargs, atol=3, rtol=0)