test: keep Hunyuan LoRA CPU path on native linear (#8166)

Signed-off-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: wangyu <53896905+yenuo26@users.noreply.github.com>

tests/diffusion/lora/test_lora_manager.py

@@ -1158,6 +1158,18 @@ def test_hunyuan_image3_peft_qkv_forward(tmp_path, monkeypatch, tp_size):
     from vllm.model_executor import parameter
     from vllm.model_executor.layers import linear
 
+    def cpu_unquantized_gemm(_layer, inputs, weight, bias=None):
+        return torch.nn.functional.linear(inputs, weight, bias)
+
+    # This numerical contract intentionally runs on CPU. Accelerator builds
+    # select the unquantized GEMM from the process-wide platform, which would
+    # otherwise dispatch CPU tensors to the ROCm-only custom operator.
+    monkeypatch.setattr(
+        linear,
+        "dispatch_unquantized_gemm",
+        lambda *_args, **_kwargs: cpu_unquantized_gemm,
+    )
+
     adapter_dir = tmp_path / "adapter"
     adapter_dir.mkdir()
     rank = 2