ROCm HW support

tests/test_node_startup.py

@@ -128,6 +128,59 @@ def test_nvidia_smi_without_torch_cuda_keeps_cpu_execution(monkeypatch):
     assert hw["ram_mb"] == 80 * 1024
 
 
+def test_torch_rocm_inventory_is_reported_when_kernels_are_not_executable(monkeypatch):
+    """ROCm can expose GPU metadata even when this torch wheel cannot run kernels."""
+    import meshnet_node.hardware as hardware_mod
+
+    class FakeProps:
+        total_memory = 64 * 1024 * 1024 * 1024
+        gcnArchName = "gfx1151"
+
+    class FakeCuda:
+        @staticmethod
+        def is_available():
+            return True
+
+        @staticmethod
+        def device_count():
+            return 1
+
+        @staticmethod
+        def current_device():
+            return 0
+
+        @staticmethod
+        def get_device_name(_idx):
+            return "AMD Radeon 8060S"
+
+        @staticmethod
+        def get_device_properties(_idx):
+            return FakeProps()
+
+        @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")
+        ),
+    )
+
+    monkeypatch.setattr(hardware_mod, "_detect_ram_mb", lambda: 125 * 1024)
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+
+    hw = hardware_mod.detect_hardware()
+
+    assert hw["device"] == "cpu"
+    assert hw["cuda_available"] is False
+    assert hw["gpu_name"] == "AMD Radeon 8060S"
+    assert hw["vram_mb"] == 64 * 1024
+    assert hw["shared_vram_mb"] == 64_000
+    assert hw["gcn_arch"] == "gfx1151"
+
+
 def test_memory_budget_uses_ram_for_cpu_and_shared_memory_for_cuda():
     assert _memory_budget("cpu", vram_mb=8192, ram_mb=80 * 1024, shared_vram_mb=40 * 1024) == (
         80 * 1024,

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()