Replace getattr with constant when buffer is a graph output

`remove_getattr_nodes` raised when a model directly returned a
registered buffer / parameter (e.g. `forward(self): return self.buf`):

  RuntimeError: <name> should not be in the graph outputs.

The pre-existing `flatten_graph_output_values` pass already lifts
buffer-returning getattr nodes into the graph outputs, so the bare
`raise` had no useful effect — there is no other op that lowers them.

Replace each output-producing getattr with a `constant` node carrying
the buffer value (read from `graph.params`). The downstream `constant`
op handler then materializes it as a const var. Intermediate getattr
nodes are still dropped as before, since their consumers read directly
from `graph.params`.

Fixes #2538.

Author: john-rocky

coremltools/converters/mil/frontend/torch/test/test_passes.py

@@ -17,6 +17,7 @@
 from ..torchir_passes import (
     flatten_graph_input_values,
     flatten_graph_output_values,
+    remove_getattr_nodes,
     transform_inplace_ops,
 )
 import coremltools as ct
@@ -113,6 +114,63 @@ def test_flatten_input_values():
         np.testing.assert_equal(graph.nodes[1].outputs[0], graph.nodes[2].inputs[0])
 
 
+    @staticmethod
+    def test_remove_getattr_nodes_with_output_buffer():
+        # Regression test for #2538: when a model directly returns a buffer,
+        # the corresponding getattr appears in graph.outputs and the pass
+        # used to raise. It now replaces the getattr with a constant node
+        # holding the buffer value so conversion can proceed.
+        params = {
+            "buf_a": np.array([1.0, 2.0, 3.0], dtype=np.float32),
+            "buf_b": np.array([4.0, 5.0], dtype=np.float32),
+        }
+        graph_nodes = [
+            InternalTorchIRNode(inputs=[], outputs=["buf_a"], kind="getattr", name="buf_a"),
+            InternalTorchIRNode(inputs=[], outputs=["buf_b"], kind="getattr", name="buf_b"),
+        ]
+        graph = InternalTorchIRGraph(
+            nodes=graph_nodes,
+            params=params,
+            inputs=OrderedDict(),
+            outputs=["buf_a", "buf_b"],
+        )
+
+        remove_getattr_nodes(graph)
+
+        np.testing.assert_equal(len(graph.nodes), 2)
+        for node in graph.nodes:
+            np.testing.assert_equal(node.kind, "constant")
+        np.testing.assert_array_equal(graph.nodes[0].attr["value"], params["buf_a"])
+        np.testing.assert_array_equal(graph.nodes[1].attr["value"], params["buf_b"])
+        # Original output names are preserved.
+        np.testing.assert_equal(graph.nodes[0].outputs, ["buf_a"])
+        np.testing.assert_equal(graph.nodes[1].outputs, ["buf_b"])
+
+
+    @staticmethod
+    def test_remove_getattr_nodes_drops_intermediate():
+        # Sanity check: a getattr node that is *not* in graph.outputs should
+        # still be dropped (the consuming op handler reads from graph.params).
+        params = {"weight": np.array([1.0], dtype=np.float32)}
+        graph_nodes = [
+            InternalTorchIRNode(inputs=[], outputs=["weight"], kind="getattr", name="weight"),
+            InternalTorchIRNode(
+                inputs=["x", "weight"], outputs=["y"], kind="mul", name="y"
+            ),
+        ]
+        graph = InternalTorchIRGraph(
+            nodes=graph_nodes,
+            params=params,
+            inputs=OrderedDict([("x", torch.rand(1))]),
+            outputs=["y"],
+        )
+
+        remove_getattr_nodes(graph)
+
+        np.testing.assert_equal(len(graph.nodes), 1)
+        np.testing.assert_equal(graph.nodes[0].kind, "mul")
+
+
     @staticmethod
     def test_flatten_output_values():
         graph = _build_flattening_test_graph()

coremltools/converters/mil/frontend/torch/torchir_passes.py

@@ -6,6 +6,8 @@
 from collections import OrderedDict, defaultdict
 from typing import Dict, Optional
 
+import numpy as np
+
 from coremltools import _logger as logger
 
 from .internal_graph import InternalTorchIRGraph, InternalTorchIRNode
@@ -229,30 +231,52 @@ def forward(self, x):
             node.model_hierarchy = cached_model_hierarchy[child_ops[node.name][0]]
 
 
-def remove_getattr_nodes(graph: InternalTorchIRGraph) -> None:
+def remove_getattr_nodes(
+    graph: InternalTorchIRGraph,
+    params: Optional[Dict[str, "np.ndarray"]] = None,
+) -> None:
     """
-    Remove the getattr nodes in the graph
+    Remove the getattr nodes from the graph.
+
+    A getattr node typically references a buffer / parameter that is consumed
+    by another op; the consuming op handler reads the value from the
+    surrounding graph's params dict, so dropping the getattr node is safe.
+    However, when a model directly returns a buffer (e.g. forward returns
+    `self.my_constant`), the getattr appears in the graph outputs. In that
+    case, replace the getattr with a constant node holding the buffer value
+    so the conversion does not crash.
     """
 
-    getattr_nodes = []
+    if params is None:
+        params = graph.params
+
     new_nodes = []
 
     for node in graph.nodes:
 
         for block in node.blocks:
-            remove_getattr_nodes(block)
+            remove_getattr_nodes(block, params=params)
 
         if node.kind == "getattr":
-            getattr_nodes.append(node)
+            if node.name in graph.outputs:
+                if node.name not in params:
+                    raise RuntimeError(
+                        "{} appears in the graph outputs but its value was not "
+                        "found in the graph params.".format(node.name)
+                    )
+                # Replace the getattr with a constant node carrying the value.
+                new_nodes.append(
+                    InternalTorchIRNode(
+                        kind="constant",
+                        inputs=[],
+                        outputs=node.outputs,
+                        name=node.name,
+                        attr={"value": params[node.name]},
+                    )
+                )
         else:
             new_nodes.append(node)
 
-    # check the getattr nodes not in the outputs
-    for node in getattr_nodes:
-        if node.name in graph.outputs:
-            raise RuntimeError("{} should not be in the graph outputs.".format(node.name))
-
-    # remove the getattr nodes
     graph.nodes = new_nodes