diff --git a/docs/models/supported_models.md b/docs/models/supported_models.md
index 66dfc740eee..0fd03536e3d 100644
--- a/docs/models/supported_models.md
+++ b/docs/models/supported_models.md
@@ -66,7 +66,7 @@ th {
 | `MiniMaxH3Pipeline` | FastH3 V2 eight-step T2VA (tile-64 VSA; full-model qualification pending) | `FastVideo/FastVideo-FastH3-8-Step-V2` | Experimental | | | | [Recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/FastVideo/FastH3-V2.md) |
 | `MiniMaxH3Pipeline` | FastH3 Omni 8-Step V1 Ref2VA (PDD, logical tile-128 VSA; experimental) | `FastVideo/FastVideo-FastH3-Omni-8-Step-V1` | Experimental | | | | [Recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/FastVideo/FastH3-Omni.md) · [Single-card guide](https://github.com/vllm-project/vllm-omni/blob/main/recipes/FastVideo/FastH3%20RefOmni%20Single%20Card.md) |
 | `TaoMateH3Pipeline` | TaoMate-H3 realtime streaming presenter (MiniMax-H3 + TaoMate LoRA, USP2/USP4) | `MiniMaxAI/MiniMax-H3` FL2VA + `TaoLiveAIGC/TaoMate-H3` | Experimental | | | | [Recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/TaoLiveAIGC/TaoMate-H3.md) |
-| `Magi2Pipeline` | MAGI-2 native Preview-stage T2VA and I2VA (272p/540p) | `sand-ai/MAGI-2-preview` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/SandAI/MAGI-2-preview-L20X.md) |
+| `Magi2Pipeline` | MAGI-2 native Preview-stage T2VA and I2VA (272p/540p) | `sand-ai/MAGI-2-preview` | ✅︎ | | | ✅︎ | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/SandAI/MAGI-2-preview-L20X.md) |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Pi0Pipeline` | π0 (Pi-Zero) VLA | `lerobot/pi0_base` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/lerobot/Pi0.md) |
 | `Pi05Pipeline` | π0.5 VLA | `lerobot/pi05_base` | ✅︎ | | | ✅︎ | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/lerobot/Pi05.md) |
diff --git a/tests/diffusion/models/magi2/test_musa_pipeline.py b/tests/diffusion/models/magi2/test_musa_pipeline.py
index f58f49a56dd..93d107b63e8 100644
--- a/tests/diffusion/models/magi2/test_musa_pipeline.py
+++ b/tests/diffusion/models/magi2/test_musa_pipeline.py
@@ -60,11 +60,11 @@ def test_constructor_uses_selected_device(monkeypatch, device_type):
 
 
 @pytest.mark.cpu
-@pytest.mark.parametrize("device_type,available", [("cuda", False), ("musa", False), ("cpu", True), ("xpu", True)])
+@pytest.mark.parametrize("device_type,available", [("cuda", False), ("musa", False), ("cpu", True)])
 def test_constructor_rejects_unsupported_or_unavailable(monkeypatch, device_type, available):
     monkeypatch.setattr(pipeline, "current_omni_platform", fake_platform(device_type, available))
     resolver = stop_before_weights(monkeypatch)
-    with pytest.raises(RuntimeError, match="available CUDA or MUSA"):
+    with pytest.raises(RuntimeError, match="requires an accelerator"):
         pipeline.Magi2Pipeline(od_config=config())
     resolver.assert_not_called()
 
diff --git a/tests/diffusion/models/magi2/test_pipeline_magi2.py b/tests/diffusion/models/magi2/test_pipeline_magi2.py
index 280caff9783..9870b6ef217 100644
--- a/tests/diffusion/models/magi2/test_pipeline_magi2.py
+++ b/tests/diffusion/models/magi2/test_pipeline_magi2.py
@@ -300,8 +300,8 @@ def test_forward_stops_peak_monitor_when_initial_synchronize_fails(monkeypatch):
     pipe.enable_diffusion_pipeline_profiler = True
     monitor = Mock(peak_bytes=0)
     monkeypatch.setattr(
-        "vllm_omni.diffusion.models.magi2.pipeline_magi2.current_omni_platform.is_cuda",
-        lambda: True,
+        "vllm_omni.diffusion.models.magi2.pipeline_magi2.current_omni_platform.device_type",
+        "cuda",
     )
     monkeypatch.setattr(
         "vllm_omni.diffusion.models.magi2.pipeline_magi2.current_omni_platform.is_available",
diff --git a/vllm_omni/diffusion/models/magi2/pipeline_magi2.py b/vllm_omni/diffusion/models/magi2/pipeline_magi2.py
index 09f590b97b7..f3cb4b3bc68 100644
--- a/vllm_omni/diffusion/models/magi2/pipeline_magi2.py
+++ b/vllm_omni/diffusion/models/magi2/pipeline_magi2.py
@@ -550,9 +550,8 @@ def __init__(self, od_config: OmniDiffusionConfig, **kwargs: object) -> None:
         if not od_config.model:
             raise ValueError("MAGI-2 requires od_config.model")
         _validate_native_topology(od_config)
-        supported_gpu = current_omni_platform.is_cuda() or current_omni_platform.is_musa()
-        if not supported_gpu or not current_omni_platform.is_available():
-            raise RuntimeError("MAGI-2 Preview requires available CUDA or MUSA GPUs")
+        if current_omni_platform.device_type == "cpu" or not current_omni_platform.is_available():
+            raise RuntimeError("MAGI-2 Preview requires an accelerator")
 
         self.od_config = od_config
         self.dtype = od_config.dtype or torch.bfloat16
@@ -1085,9 +1084,7 @@ def forward(self, req: DiffusionRequestBatch) -> DiffusionOutput:
         seed = _resolve_request_seed(sampling)
         _seed_request(seed)
 
-        has_accelerator = (
-            current_omni_platform.is_cuda() or current_omni_platform.is_musa()
-        ) and current_omni_platform.is_available()
+        has_accelerator = current_omni_platform.device_type != "cpu" and current_omni_platform.is_available()
         device_index = torch.accelerator.current_device_index() if has_accelerator else None
         # Sampling reserved memory is qualification instrumentation, not part
         # of ordinary serving. It starts only when the pipeline profiler is