Arm backend: Convert remaining asserts that should be exceptions

backends/arm/_passes/fuse_batchnorm2d_pass.py

@@ -12,6 +12,7 @@
     create_node,
     get_first_fake_tensor,
 )
+from executorch.backends.arm.common.debug import get_node_debug_info
 from executorch.backends.transforms.utils import (
     create_constant_placeholder,
     delete_constant_placeholder,
@@ -60,8 +61,16 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:  # noqa: C901
             input_node = node.all_input_nodes[0]
             is_single_user = len(input_node.users) == 1
             bn_weight_node, bn_bias_node, bn_mean_node, bn_var_node = node.args[1:5]
-            assert bn_mean_node is not None, "Batchnorm mean node cannot be None."
-            assert bn_var_node is not None, "Batchnorm var node cannot be None."
+            if bn_mean_node is None:
+                raise RuntimeError(
+                    "BatchNorm mean buffer missing for node: "
+                    f"{get_node_debug_info(node, graph_module)}"
+                )
+            if bn_var_node is None:
+                raise RuntimeError(
+                    "BatchNorm variance buffer missing for node: "
+                    f"{get_node_debug_info(node, graph_module)}"
+                )
 
             epsilon = node.args[-1]
 
@@ -133,14 +142,23 @@ def call(self, graph_module: torch.fx.GraphModule) -> PassResult:  # noqa: C901
                     input_node = new_input_node
             else:
                 input_weight_node, input_bias_node = input_node.args[1:3]
-                assert (
+                if not (
                     isinstance(input_weight_node, Node)
                     and input_weight_node.op == "placeholder"
-                ), "Parameter weight of convolution must be a placeholder"
-                assert (input_bias_node is None) or (
-                    isinstance(input_weight_node, Node)
-                    and input_weight_node.op == "placeholder"
-                ), "Parameter bias of convolution must be a placeholder or None"
+                ):
+                    raise RuntimeError(
+                        "Parameter weight of convolution must be a placeholder"
+                    )
+                if not (
+                    (input_bias_node is None)
+                    or (
+                        isinstance(input_weight_node, Node)
+                        and input_weight_node.op == "placeholder"
+                    )
+                ):
+                    raise RuntimeError(
+                        "Parameter bias of convolution must be a placeholder or None"
+                    )
 
                 input_weight_tensor = torch.Tensor(
                     get_param(self.exported_program, input_weight_node)

backends/arm/arm_vela.py

@@ -34,7 +34,10 @@ def vela_bin_pack_io(prefix, data):
         io_elem_size = data[prefix + "_elem_size"][i]
         io_offset = data[prefix + "_offset"][i]
         io_region = data[prefix + "_region"][i]
-        assert len(io_shape) == vela_io_shape_dims
+        if len(io_shape) != vela_io_shape_dims:
+            raise ValueError(
+                f"Expected {vela_io_shape_dims}D shape, got {len(io_shape)}D"
+            )
         inp_pad = io_shape.tolist()
         io_struct = struct.pack(
             "<iiiiiiiii", *inp_pad, io_elem_size, io_offset, io_region

backends/arm/common/arm_compile_spec.py

@@ -126,7 +126,8 @@ def validate(self):
 
     def to_list(self):
         """Get the ArmCompileSpec in list form."""
-        assert self.tosa_spec
+        if not self.tosa_spec:
+            raise ValueError("tosa_spec must be set before calling to_list()")
 
         # Always supply a TOSA version
         compile_spec = [

backends/arm/process_node.py

@@ -106,11 +106,16 @@ def process_inputs_to_parameters(
         ) from e
     parameter_data = get_param(edge_program, node)
 
-    assert isinstance(parameter_data, torch.Tensor), "Expect Attr to be tensor"
+    if not isinstance(parameter_data, torch.Tensor):
+        raise TypeError(
+            f"Expected parameter '{node.name}' to be a torch.Tensor, got "
+            f"{type(parameter_data).__name__}"
+        )
     parameter_values = parameter_data.detach().numpy()
 
     if tosa_arg.dtype == torch.float32:
-        assert tosa_spec.support_float(), f"{tosa_spec} doesn't support float"
+        if not tosa_spec.support_float():
+            raise ValueError(f"{tosa_spec} doesn't support float operations")
 
     # Handle special case for INT48 tensors
     special_type = node.meta.get(TosaSpecialDtype.meta_key(), None)
@@ -142,7 +147,11 @@ def process_inputs_to_buffers(
         ) from e
     buffer_data = get_buffer(edge_program, node)
 
-    assert isinstance(buffer_data, torch.Tensor), "Expect Attr to be tensor"
+    if not isinstance(buffer_data, torch.Tensor):
+        raise TypeError(
+            f"Expected buffer '{node.name}' to be a torch.Tensor, got "
+            f"{type(buffer_data).__name__}"
+        )
     buffer_values = buffer_data.detach().numpy()
 
     # TODO: fragile code for temporary fix
@@ -183,8 +192,12 @@ def process_placeholder(
     tosa_spec: TosaSpecification,
 ):
     """Wrapper for processing and serializing all types of placeholders"""
-    assert node.name == node.target, "Expect placeholder name and target to match"
-    assert 0 == len(node.args), "Can't handle default input values"
+    if node.name != node.target:
+        raise ValueError(
+            f"Placeholder name '{node.name}' does not match target '{node.target}'"
+        )
+    if len(node.args) != 0:
+        raise ValueError(f"Placeholder '{node.name}' must not have default values")
 
     if node.name in edge_program.graph_signature.user_inputs:
         process_inputs(node, tosa_graph, tosa_spec)