From d90bb66daae7a57c7710cb6d09ec815181530134 Mon Sep 17 00:00:00 2001
From: bo0tzz <git@bo0tzz.me>
Date: Mon, 13 Jul 2026 16:54:01 +0200
Subject: [PATCH] feat: ML hardware acceleration testing on pokedex

---
 .github/workflows/test.yml        | 44 +++++++++++++++++++++++++++++++
 machine-learning/mise.toml        |  3 +++
 machine-learning/pyproject.toml   |  2 +-
 machine-learning/test_hardware.py | 35 ++++++++++++++++++++++++
 4 files changed, 83 insertions(+), 1 deletion(-)
 create mode 100644 machine-learning/test_hardware.py

diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 465f18ae8b..0f9d8c3c57 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -636,6 +636,50 @@ jobs:
       - name: Run ci-unit
         run: mise run ci-unit
 
+  ml-hwaccel-tests:
+    name: HW Accel Test ML (${{ matrix.name }})
+    needs: pre-job
+    if: ${{ fromJSON(needs.pre-job.outputs.should_run).machine-learning == true }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: nvidia-ampere
+            runner: pokedex-dgpu-nvidia-ampere
+            device: cuda
+            expected-ep: CUDAExecutionProvider
+    runs-on: ${{ matrix.runner }}
+    permissions:
+      contents: read
+    defaults:
+      run:
+        working-directory: ./machine-learning
+    env:
+      IMMICH_HWACCEL_EXPECTED_EP: ${{ matrix.expected-ep }}
+      HF_HUB_DISABLE_XET: '1'
+    steps:
+      - id: token
+        uses: immich-app/devtools/actions/create-workflow-token@9db058b2e6eec20e07760b0e17a0505c78ec3191 # create-workflow-token-action-v2.0.1
+        with:
+          client-id: ${{ secrets.PUSH_O_MATIC_APP_CLIENT_ID }}
+          private-key: ${{ secrets.PUSH_O_MATIC_APP_KEY }}
+
+      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
+        with:
+          persist-credentials: false
+          token: ${{ steps.token.outputs.token }}
+
+      - name: Setup Mise
+        uses: immich-app/devtools/actions/use-mise@3bca63ca3c15020293b36b51737a3ee2c773340b # use-mise-action-v3.1.0
+        with:
+          github_token: ${{ steps.token.outputs.token }}
+
+      - name: Install
+        run: mise run install --extra ${{ matrix.device }}
+
+      - name: Run hardware tests
+        run: mise run test-hardware
+
   github-files-formatting:
     name: .github Files Formatting
     needs: pre-job
diff --git a/machine-learning/mise.toml b/machine-learning/mise.toml
index e5e30c4fc2..aa37c90d88 100644
--- a/machine-learning/mise.toml
+++ b/machine-learning/mise.toml
@@ -14,6 +14,9 @@ run = "uv run pytest --cov=immich_ml --cov-report term-missing"
 [tasks.format]
 run = "uv run ruff format immich_ml"
 
+[tasks.test-hardware]
+run = "uv run pytest -m gpu test_hardware.py"
+
 [tasks.check]
 run = "uv run mypy --strict immich_ml/"
 
diff --git a/machine-learning/pyproject.toml b/machine-learning/pyproject.toml
index 6b2082930c..4fc4f48dd3 100644
--- a/machine-learning/pyproject.toml
+++ b/machine-learning/pyproject.toml
@@ -92,4 +92,4 @@ select = ["E", "F", "I"]
 per-file-ignores = { "test_main.py" = ["F403"] }
 
 [tool.pytest.ini_options]
-markers = ["providers", "ov_device_ids"]
+markers = ["providers", "ov_device_ids", "gpu"]
diff --git a/machine-learning/test_hardware.py b/machine-learning/test_hardware.py
new file mode 100644
index 0000000000..dde9abcc18
--- /dev/null
+++ b/machine-learning/test_hardware.py
@@ -0,0 +1,35 @@
+import os
+
+import pytest
+from PIL import Image
+
+from immich_ml.models import from_model_type
+from immich_ml.schemas import ModelTask, ModelType
+
+pytestmark = pytest.mark.gpu
+
+EXPECTED_EP = os.environ.get("IMMICH_HWACCEL_EXPECTED_EP")
+
+
+@pytest.fixture(autouse=True)
+def require_hwaccel_run() -> None:
+    if not EXPECTED_EP:
+        pytest.skip("IMMICH_HWACCEL_EXPECTED_EP unset; not a hardware run")
+
+
+MODELS = [
+    pytest.param("ViT-B-32__openai", ModelType.VISUAL, ModelTask.SEARCH, id="clip-visual"),
+]
+
+
+@pytest.mark.parametrize("model_name, model_type, model_task", MODELS)
+def test_hwaccel_engaged(model_name: str, model_type: ModelType, model_task: ModelTask) -> None:
+    model = from_model_type(model_name, model_type, model_task)
+    model.load()
+
+    # a CPU fallback returns the same output, so the provider list is the only fallback signal
+    assert EXPECTED_EP in model.session.providers, (
+        f"expected {EXPECTED_EP}, session resolved to {model.session.providers}"
+    )
+
+    model.predict(Image.new("RGB", (224, 224)))