Arm backend: Make passes preserve and update node metadata (#9362)

When creating or updating nodes in passes, the metadata is not preserved
nor updated correctly. This patch adds an ArmPass base class which may
update the node metadata if super().call_operator(update=True) is used.
It also adds functionality to arm_pass_utils.create_node() to update the
node metadata. It will only update the 'stack_trace' field. All the
other fields will be preserved from the original node.


Signed-off-by: Oscar Andersson <[email protected]>

Author: oscarandersson8218

backends/arm/_passes/__init__.py

@@ -7,6 +7,7 @@
 from . import arm_pass_utils  # noqa
 from .annotate_channels_last_dim_order_pass import AnnotateChannelsLastDimOrder  # noqa
 from .annotate_decomposed_matmul import AnnotateDecomposedMatmulPass  # noqa
+from .arm_pass import ArmPass  # noqa
 from .cast_int64_pass import CastInt64BuffersToInt32Pass  # noqa
 from .cast_to_int32_pass import CastToInt32Pass  # noqa
 from .conv1d_unsqueeze_pass import Conv1dUnsqueezePass  # noqa

backends/arm/_passes/arm_pass.py

@@ -0,0 +1,33 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-unsafe
+
+import traceback
+from typing import Optional
+
+import torch
+from executorch.exir.pass_base import ExportPass, NodeMetadata
+
+
+class ArmPass(ExportPass):
+    """Base class for Arm passes"""
+
+    def __init__(self, exported_program: Optional[torch.export.ExportedProgram] = None):
+        super(ArmPass, self).__init__()
+        self.exported_program = exported_program
+
+    def call_operator(self, op, args, kwargs, meta, updated: Optional[bool] = False):
+        if not updated:
+            return super().call_operator(op, args, kwargs, meta)
+
+        # if updated we should update metadata
+        new_meta = {}
+        keys = meta.data.keys()
+        for key in keys:
+            new_meta[key] = meta[key]
+        old_stack_trace = new_meta.get("stack_trace", "")
+        new_meta["stack_trace"] = f"{old_stack_trace}\n{traceback.format_stack()[-2]}"
+        return super().call_operator(op, args, kwargs, NodeMetadata(new_meta))

backends/arm/_passes/arm_pass_utils.py

@@ -7,12 +7,12 @@
 
 # pyre-unsafe
 
+import traceback
 from inspect import isclass
 from typing import Optional, Sequence
 
 import torch
 import torch.fx
-
 from executorch.exir import ExportedProgram
 from executorch.exir.dialects._ops import ops as exir_ops
 
@@ -96,6 +96,7 @@ def create_node(
     kwargs: Optional[dict] = None,
     quantize: bool = False,
     q_params: Optional[tuple] = None,
+    from_node: Optional[torch.fx.Node] = None,
 ):
     """
     Adds a node to 'graph'. graph.inserting_before/after() should be used before the call to decide where to insert the node.
@@ -108,15 +109,26 @@ def create_node(
         args=args,
         kwargs=kwargs or {},
     )
+
+    new_meta = {}
+    if from_node:
+        keys = from_node.meta.keys()
+        for key in keys:
+            new_meta[key] = from_node.meta[key]
+    old_stack_trace = new_meta.get("stack_trace", "")
+    new_meta["stack_trace"] = f"{old_stack_trace}\n{traceback.format_stack()[-2]}"
+    node.meta = new_meta
+
     if quantize and q_params:
-        return insert_q_dq_pair(graph, node, q_params)
+        return insert_q_dq_pair(graph, node, q_params, from_node)
     return node
 
 
 def insert_q_dq_pair(
     graph: torch.fx.Graph,
     anchor: torch.fx.Node,
     q_params: tuple,
+    from_node: Optional[torch.fx.Node] = None,
 ):
     """
     Inserts a q dq node pair after the node 'anchor'.
@@ -127,13 +139,15 @@ def insert_q_dq_pair(
             graph=graph,
             op_target=exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
             args=(),  # We add the argument last
+            from_node=from_node if from_node else anchor,
         )
         q.meta = anchor.meta
     with graph.inserting_after(q):
         dq = create_node(
             graph=graph,
             op_target=exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
             args=(q,) + q_params,
+            from_node=from_node if from_node else anchor,
         )
         dq.meta = q.meta
     anchor.replace_all_uses_with(dq)

backends/arm/_passes/decompose_layernorm_pass.py

@@ -9,9 +9,10 @@
 import operator
 
 import torch
+from executorch.backends.arm._passes import ArmPass
 from executorch.backends.arm._passes.arm_pass_utils import create_node
 from executorch.exir.dialects._ops import ops as exir_ops
-from executorch.exir.pass_base import ExportPass, PassResult
+from executorch.exir.pass_base import PassResult
 
 
 def get_layer_norm_decomposition(op) -> tuple:
@@ -40,7 +41,7 @@ def get_layer_norm_decomposition(op) -> tuple:
     raise RuntimeError(f"Can't get layer_norm composition for op {op}")
 
 
-class DecomposeLayerNormPass(ExportPass):
+class DecomposeLayerNormPass(ArmPass):
     """
     layernorm is defined as: ((x - E[x]) / sqrt(Var[x] + eps)) * weights + bias
     Decompose layernorm(x, normalized_shape, weights, bias, eps) to a sequence of:
@@ -111,35 +112,56 @@ def call(self, graph_module: torch.fx.GraphModule):
                     var_op,
                     args=(x, dims),
                     kwargs={"correction": 0, "keepdim": keepdim},
+                    from_node=node,
                 )
                 full = create_node(
                     graph_module.graph,
                     full_op,
                     args=(epsilon_reshaped_shape, epsilon),
                     kwargs={"dtype": dtype},
+                    from_node=node,
+                )
+                add0 = create_node(
+                    graph_module.graph, add_op, args=(var, full), from_node=node
+                )
+                rsqrt = create_node(
+                    graph_module.graph, rsqrt_op, args=(add0,), from_node=node
+                )
+                mul0 = create_node(
+                    graph_module.graph, mul_op, args=(sub, rsqrt), from_node=node
                 )
-                add0 = create_node(graph_module.graph, add_op, args=(var, full))
-                rsqrt = create_node(graph_module.graph, rsqrt_op, args=(add0,))
-                mul0 = create_node(graph_module.graph, mul_op, args=(sub, rsqrt))
                 if weights is not None:
                     weights_reshaped = create_node(
                         graph_module.graph,
                         view_op,
                         args=(weights, weights_reshaped_shape),
+                        from_node=node,
                     )
                     mul1 = create_node(
-                        graph_module.graph, mul_op, args=(mul0, weights_reshaped)
+                        graph_module.graph,
+                        mul_op,
+                        args=(
+                            mul0,
+                            weights_reshaped,
+                        ),
+                        from_node=node,
                     )
                 else:
                     mul1 = mul0
                 output = mul1
                 if bias is not None:
                     bias_reshaped_shape = weights_reshaped_shape
                     bias_reshaped = create_node(
-                        graph_module.graph, view_op, args=(bias, bias_reshaped_shape)
+                        graph_module.graph,
+                        view_op,
+                        args=(bias, bias_reshaped_shape),
+                        from_node=node,
                     )
                     output = create_node(
-                        graph_module.graph, add_op, args=(mul1, bias_reshaped)
+                        graph_module.graph,
+                        add_op,
+                        args=(mul1, bias_reshaped),
+                        from_node=node,
                     )
 
                 users = [user for user in node.users if node != user]

backends/arm/_passes/decompose_meandim_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -7,9 +7,9 @@
 # pyre-unsafe
 
 import torch
+from executorch.backends.arm._passes import ArmPass
 from executorch.backends.arm._passes.arm_pass_utils import get_node_arg
 from executorch.exir.dialects._ops import ops as exir_ops
-from executorch.exir.pass_base import ExportPass
 
 
 def get_meandim_decomposition(op) -> tuple:
@@ -28,7 +28,7 @@ def get_meandim_decomposition(op) -> tuple:
     raise RuntimeError(f"Can't get meandim decomposition for op {op}")
 
 
-class DecomposeMeanDimPass(ExportPass):
+class DecomposeMeanDimPass(ArmPass):
     """
     This pass decomposes meandim into a sum and mul node.
 
@@ -62,8 +62,8 @@ def call_operator(self, op, args, kwargs, meta):
 
         sum_op, full_op, mul_op = get_meandim_decomposition(op)
 
-        sum = super().call_operator(sum_op, (x, dim, keepdim), {}, meta)
+        sum = super().call_operator(sum_op, (x, dim, keepdim), {}, meta, True)
         full = super().call_operator(
-            full_op, ([1] * len(shape), 1 / N), {"dtype": dtype}, meta
+            full_op, ([1] * len(shape), 1 / N), {"dtype": dtype}, meta, True
         )
-        return super().call_operator(mul_op, (sum, full), {}, meta)
+        return super().call_operator(mul_op, (sum, full), {}, meta, True)

backends/arm/_passes/decompose_softmax_unstable_pass.py

@@ -6,8 +6,8 @@
 # pyre-unsafe
 
 import torch
+from executorch.backends.arm._passes import ArmPass
 from executorch.exir.dialects._ops import ops as exir_ops
-from executorch.exir.pass_base import ExportPass
 
 # For BI case
 torch_softmax = (torch.ops.aten.softmax.int, torch.ops.aten.log_softmax.int)
@@ -45,7 +45,7 @@ def get_logsoftmax_ops(op) -> tuple:
     raise RuntimeError(f"Can't get softmax decomposition ops for op {op}")
 
 
-class DecomposeSoftmaxUnstablePass(ExportPass):
+class DecomposeSoftmaxUnstablePass(ArmPass):
     """
     This pass decomposes log softmax or softmax into more primitive ops.
 
@@ -66,10 +66,10 @@ def call_operator(self, op, args, kwargs, meta):
         _input = args[0]
         dim = [args[1]]
 
-        op1 = super().call_operator(exp_op, (_input,), {}, meta)
-        op2 = super().call_operator(sum_op, (op1, dim, True), {}, meta)
-        op3 = super().call_operator(reciprocal_op, (op2,), {}, meta)
-        op4 = super().call_operator(mul_op, (op1, op3), {}, meta)
+        op1 = super().call_operator(exp_op, (_input,), {}, meta, True)
+        op2 = super().call_operator(sum_op, (op1, dim, True), {}, meta, True)
+        op3 = super().call_operator(reciprocal_op, (op2,), {}, meta, True)
+        op4 = super().call_operator(mul_op, (op1, op3), {}, meta, True)
         if op in log_softmax:
-            op4 = super().call_operator(log_op, (op4,), {}, meta)
+            op4 = super().call_operator(log_op, (op4,), {}, meta, True)
         return op4

backends/arm/_passes/decompose_var_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -8,9 +8,9 @@
 
 
 import torch
+from executorch.backends.arm._passes import ArmPass
 from executorch.backends.arm._passes.arm_pass_utils import get_node_arg
 from executorch.exir.dialects._ops import ops as exir_ops
-from executorch.exir.pass_base import ExportPass
 
 
 def get_var_decomposition(op) -> tuple:
@@ -33,7 +33,7 @@ def get_var_decomposition(op) -> tuple:
     raise RuntimeError(f"Can't get var decomposition for op {op}")
 
 
-class DecomposeVarPass(ExportPass):
+class DecomposeVarPass(ArmPass):
     """
     This pass decomposes var.correction and var.dim into smaller ops (see https://pytorch.org/docs/stable/generated/torch.var.html)
 
@@ -77,14 +77,17 @@ def call_operator(self, op, args, kwargs, meta):
             N *= input_shape[d]
 
         mean_op, diff_op, mul_op, sum_op, full_op = get_var_decomposition(op)
-        mean = super().call_operator(mean_op, (x, dim, True), {}, meta)
-        diff = super().call_operator(diff_op, (x, mean), {}, meta)
-        squared_diff = super().call_operator(mul_op, (diff, diff), {}, meta)
-        sum = super().call_operator(sum_op, (squared_diff, dim, keepdim), {}, meta)
+        mean = super().call_operator(mean_op, (x, dim, True), {}, meta, True)
+        diff = super().call_operator(diff_op, (x, mean), {}, meta, True)
+        squared_diff = super().call_operator(mul_op, (diff, diff), {}, meta, True)
+        sum = super().call_operator(
+            sum_op, (squared_diff, dim, keepdim), {}, meta, True
+        )
         full = super().call_operator(
             full_op,
             ([], 1 / max(0, N - correction)),
             {"dtype": dtype},
             meta,
+            True,
         )
-        return super().call_operator(mul_op, (sum, full), {}, meta)
+        return super().call_operator(mul_op, (sum, full), {}, meta, True)

backends/arm/_passes/mm_to_bmm_pass.py

@@ -47,7 +47,7 @@ def call(self, graph_module: torch.fx.GraphModule):
 
                 with graph.inserting_before(node):
                     unsqueeze_before = create_node(
-                        graph, exir_ops.edge.aten.unsqueeze_copy.default
+                        graph, exir_ops.edge.aten.unsqueeze_copy.default, from_node=node
                     )
                     unsqueeze_before.args = (
                         input_node,  # Input is node's original input
@@ -58,13 +58,14 @@ def call(self, graph_module: torch.fx.GraphModule):
                 # If Quantized we must insert unsqueeze --> q --> dq --> node
                 if input_node.target == dq_op:
                     q_params = input_node.args[1:]
-                    insert_q_dq_pair(graph, unsqueeze_before, q_params)
+                    insert_q_dq_pair(graph, unsqueeze_before, q_params, from_node=node)
 
             # Replace mm node with bmm
             with graph.inserting_before(node):
                 bmm_node = create_node(
                     graph,
                     exir_ops.edge.aten.bmm.default,
+                    from_node=node,
                 )
                 bmm_node.args = node.args
                 node.replace_all_uses_with(bmm_node)
@@ -75,6 +76,7 @@ def call(self, graph_module: torch.fx.GraphModule):
                 squeeze_after = create_node(
                     graph,
                     exir_ops.edge.aten.squeeze_copy.dims,
+                    from_node=node,
                 )
                 squeeze_after.args = (
                     bmm_node,
@@ -89,7 +91,7 @@ def call(self, graph_module: torch.fx.GraphModule):
             # If quantized, insert mm --> q --> dq --> squeeze
             if all(original_user.target == q_op for original_user in original_users):
                 q_params = original_users[0].args[1:]
-                insert_q_dq_pair(graph, bmm_node, q_params)
+                insert_q_dq_pair(graph, bmm_node, q_params, from_node=node)
 
             modified_graph = True
 

backends/arm/tosa_utils.py

@@ -43,10 +43,17 @@ def get_node_debug_info(node: torch.fx.Node, graph_module: torch.fx.GraphModule)
         "  Node.meta = \n"
     )
     for k, v in node.meta.items():
-        output += f"    '{k}' = {v}\n"
-        if isinstance(v, list):
-            for i in v:
-                output += f"      {i}\n"
+        if k == "stack_trace":
+            matches = v.split("\n")
+            output += "      'stack_trace =\n"
+            for m in matches:
+                output += f"      {m}\n"
+        else:
+            output += f"    '{k}' = {v}\n"
+
+            if isinstance(v, list):
+                for i in v:
+                    output += f"      {i}\n"
     return output