ROCm HW support

packages/node/meshnet_node/model_backend.py

@@ -39,6 +39,26 @@ class KVCacheMiss(ModelBackendError):
     """
 
 
+def _torch_cuda_is_executable(torch_module: Any) -> bool:
+    """Return True only when this process can actually execute a CUDA/HIP op.
+
+    On ROCm, ``torch.cuda.is_available()`` can be true for an AMD GPU even when
+    the installed PyTorch wheel has no runnable kernels for that GPU target.
+    Loading weights onto such a device can segfault in native code, so the model
+    backend must use the same executable-device check as startup hardware
+    detection rather than trusting inventory alone.
+    """
+    try:
+        if not torch_module.cuda.is_available():
+            return False
+        probe = torch_module.empty((1,), device="cuda")
+        probe += 1
+        torch_module.cuda.synchronize()
+        return True
+    except Exception:
+        return False
+
+
 @dataclass(frozen=True)
 class TensorPayload:
     body: bytes
@@ -209,7 +229,7 @@ class TorchModelShard:
             ) from exc
 
         self.torch = torch
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.device = torch.device("cuda" if _torch_cuda_is_executable(torch) else "cpu")
         load_source = str(cache_dir) if cache_dir is not None and (cache_dir / "config.json").exists() else model_id
         quant_config, dtype, uses_quantized_weights = _model_load_plan(
             AutoConfig,