diff --git a/.github/workflows/_qdc-test.yml b/.github/workflows/_qdc-test.yml
index 82f88064b..0de9998af 100644
--- a/.github/workflows/_qdc-test.yml
+++ b/.github/workflows/_qdc-test.yml
@@ -45,7 +45,9 @@ jobs:
         if: inputs.gate_on_changed_files
         uses: ./.github/actions/changed-files
         with:
-          include: '^(sdk|third-party|tests/qdc)/|^\.github/workflows/_qdc-test\.yml$'
+          # QDC harness bundles the full tests/ tree onto the device, so any
+          # tests/** change (not just tests/qdc/**) affects the on-device run.
+          include: '^(sdk|third-party|tests)/|^\.github/workflows/_qdc-test\.yml$'
 
       - name: Decide skip
         id: skip
diff --git a/tests/README.md b/tests/README.md
index 512501135..8c8ef8a92 100644
--- a/tests/README.md
+++ b/tests/README.md
@@ -30,7 +30,7 @@ model-manager-pull check that runs first:
 | `test_vlm_quality_keywords`   | Golden-retriever caption must match one of the canonical keywords. |
 | `test_mtp_multi_turn`         | (llama_cpp only) same "Alice" convo with `spec_type='draft-mtp'`.  |
 
-Backends per plugin: `llama_cpp` on `cpu` + `npu`; `qairt` on `npu` only.
+Backends per plugin: `llama_cpp` on `cpu` + `gpu` + `npu`; `qairt` on `npu` only.
 Model-manager pull failures are treated as FAIL, not SKIP, so a broken
 download surfaces as a red CI leg instead of a silent green skip.
 
@@ -44,8 +44,9 @@ The conftest auto-tags items by location and `device_map` value:
 | `llama_cpp`     | items in `tests/test_llama_cpp.py`                       |
 | `qairt`         | items in `tests/test_qairt.py`                           |
 | `device_cpu`    | parametrised with `device_map='cpu'`                     |
+| `device_gpu`    | parametrised with `device_map='gpu'`                     |
 | `device_npu`    | parametrised with `device_map='npu'`                     |
-| `snapdragon`    | any `device_map='npu'` cell (auto-applied)               |
+| `snapdragon`    | any `device_map` in {`gpu`, `npu`} cell (auto-applied)   |
 | `llm` / `vlm`   | applied per-test via `@pytest.mark.llm` / `.vlm`         |
 
 `snapdragon`-marked and `qairt` items skip automatically unless
diff --git a/tests/conftest.py b/tests/conftest.py
index 142034253..823c30c8d 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -33,9 +33,11 @@
 
 _DEVICE_MARKER = {
     'cpu': 'device_cpu',
+    'gpu': 'device_gpu',
     'npu': 'device_npu',
 }
-_SNAPDRAGON_DEVICES = {'npu'}
+# GPU uses the Snapdragon OpenCL backend, so it needs real hardware just like NPU.
+_SNAPDRAGON_DEVICES = {'gpu', 'npu'}
 
 
 def _is_snapdragon_host() -> bool:
diff --git a/tests/pytest.ini b/tests/pytest.ini
index 4a07c4352..9aa640f34 100644
--- a/tests/pytest.ini
+++ b/tests/pytest.ini
@@ -9,6 +9,7 @@ markers =
     llm: text-only generation
     vlm: vision-language generation
     device_cpu: cpu backend (works anywhere)
+    device_gpu: OpenCL GPU backend (Snapdragon only)
     device_npu: HTP NPU backend (Snapdragon only)
     snapdragon: requires real Snapdragon host (gated by GENIEX_DEVICE_TEST=1)
 norecursedirs = qdc
diff --git a/tests/test_llama_cpp.py b/tests/test_llama_cpp.py
index 5d10f03f0..59495c498 100644
--- a/tests/test_llama_cpp.py
+++ b/tests/test_llama_cpp.py
@@ -29,8 +29,8 @@
     VLM_QUALITY_TEMPERATURE,
 )
 
-_LLM_BACKENDS = ['cpu', 'npu']
-_VLM_BACKENDS = ['cpu', 'npu']
+_LLM_BACKENDS = ['cpu', 'gpu', 'npu']
+_VLM_BACKENDS = ['cpu', 'gpu', 'npu']
 
 
 def test_model_manager_pull(llama_cpp_llm_paths, llama_cpp_vlm_paths, llama_cpp_mtp_paths):
@@ -80,11 +80,12 @@ def test_llm_multi_turn(llama_cpp_llm_paths, device_map):
 
 
 @pytest.mark.vlm
-def test_vlm_multi_turn(llama_cpp_vlm_paths, test_image):
+@pytest.mark.parametrize('device_map', _VLM_BACKENDS)
+def test_vlm_multi_turn(llama_cpp_vlm_paths, test_image, device_map):
     with geniex.AutoModelForVision2Seq.from_pretrained(
         LLAMA_CPP_VLM_MODEL,
         precision=LLAMA_CPP_VLM_PRECISION,
-        device_map='cpu',
+        device_map=device_map,
     ) as vlm:
         history = [
             {