ROCm HW support

tests/test_real_model_backend.py

@@ -24,6 +24,7 @@ from meshnet_node.model_backend import (
     _should_partial_materialize_shard,
     _decoder_attention_mask,
     _int_tensor_header,
+    _torch_cuda_is_executable,
     build_quantization_config,
     validate_quantization,
 )
@@ -209,6 +210,26 @@ def test_bitsandbytes_configs_are_created_lazily(monkeypatch):
     ]
 
 
+def test_rocm_inventory_without_executable_kernels_is_not_used_as_cuda():
+    class FakeCuda:
+        @staticmethod
+        def is_available():
+            return True
+
+        @staticmethod
+        def synchronize():
+            raise AssertionError("synchronize should not run after empty() fails")
+
+    fake_torch = types.SimpleNamespace(
+        cuda=FakeCuda(),
+        empty=lambda *args, **kwargs: (_ for _ in ()).throw(
+            RuntimeError("HIP error: invalid device function")
+        ),
+    )
+
+    assert _torch_cuda_is_executable(fake_torch) is False
+
+
 def test_head_forward_accepts_text_prompt_and_returns_bfloat16_activations():
     node = TorchNodeServer(backend=_FakeBackend())
     port = node.start()