fix: normalize cache key dtype to prevent initializer name collisions (#2888)

When a Python int literal (e.g. `1`) is used in both untyped positions
(like a Gather index, where the ONNX schema type variable is unbound)
and typed positions (like Add with an INT64 tensor), the constant cache
created two entries: (1, None) and (1, INT64). Both generated the same
initializer name 'const_1_i64' but as different ir.Value objects,
causing register_initializer to raise ValueError.

Fix: before cache lookup, normalize dtype=None to the default ONNX dtype
for the Python type (_PYTHON_TYPE_TO_DTYPE: int->INT64, float->FLOAT).
This merges both entries into a single cache key and reuses the same
ir.Value. Applied to both scalar and sequence (list/tuple) branches.

---------

Signed-off-by: G Ramalingam <grama@microsoft.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: gramalingam

onnxscript/_internal/builder.py

@@ -346,13 +346,16 @@ def _input_to_ir_value(
         # For simple scalar/sequence constants, use a cache to avoid duplicate initializers.
         # These are shared across layers, so we don't qualify the name with context prefix.
         if isinstance(value, (int, float, bool, str)):
+            # Normalize dtype: when None, use the default ONNX type for the
+            # Python scalar so that (value, None) and (value, default_dtype)
+            # share one cache entry and one initializer name.
+            if dtype is None:
+                dtype = _PYTHON_TYPE_TO_DTYPE.get(type(value))
             cache_key = (value, dtype)
             if cache_key in self._constant_cache:
                 ir_value = self._constant_cache[cache_key]
             else:
-                type_suffix = (
-                    _dtype_suffix(dtype) if dtype is not None else _type_suffix(type(value))
-                )
+                type_suffix = _dtype_suffix(dtype) if dtype is not None else ""
                 name = _constant_name(value, type_suffix, len(self._constant_cache))
                 tensor = ir.tensor(value, dtype=dtype, name=name)
                 ir_value = self.initializer(tensor, name=name, qualify=False)
@@ -363,13 +366,14 @@ def _input_to_ir_value(
             and all(isinstance(v, type(value[0])) for v in value)
             and isinstance(value[0], (int, float, bool, str))
         ):
+            # Same normalization for sequences of scalars.
+            if dtype is None:
+                dtype = _PYTHON_TYPE_TO_DTYPE.get(type(value[0]))
             cache_key = (tuple(value), dtype)
             if cache_key in self._constant_cache:
                 ir_value = self._constant_cache[cache_key]
             else:
-                type_suffix = (
-                    _dtype_suffix(dtype) if dtype is not None else _type_suffix(type(value[0]))
-                )
+                type_suffix = _dtype_suffix(dtype) if dtype is not None else ""
                 name = _constant_name(value, type_suffix, len(self._constant_cache))
                 tensor = ir.tensor(list(value), dtype=dtype, name=name)
                 ir_value = self.initializer(tensor, name=name, qualify=False)

onnxscript/_internal/builder_test.py

@@ -514,6 +514,66 @@ def test_int_constant_with_unknown_type_uses_cast_like(self):
         # Add should use the CastLike output, not the raw constant
         self.assertIs(add_node.inputs[1], cast_like_node.outputs[0])
 
+    def test_int_literal_no_clash_across_typed_and_untyped_contexts(self):
+        """Test that the same int literal used in typed and untyped positions
+        does not cause an initializer name collision.
+
+        Regression test: previously, (1, None) and (1, INT64) were separate
+        cache keys but generated the same name 'const_1_i64', causing
+        register_initializer to raise ValueError.
+        """
+        graph = ir.Graph(
+            name="test_model",
+            inputs=[],
+            outputs=[],
+            nodes=[],
+            opset_imports={"": _default_opset_version},
+        )
+        x = ir.Value(name="x", type=ir.TensorType(ir.DataType.INT64), shape=ir.Shape([3]))
+        graph.inputs.append(x)
+
+        graph_builder = builder.GraphBuilder(graph)
+        op = graph_builder.op
+
+        # Gather index: int literal in untyped position (Tind has no binding)
+        _ = op.Gather(x, 1, axis=0)
+        # Add: int literal in typed position (T bound to INT64 from x)
+        _ = op.Add(x, 1)
+
+        # Both ops should share the same initializer (same ir.Value object)
+        gather_node = graph.node(0)
+        add_node = graph.node(1)
+        self.assertIs(gather_node.inputs[1], add_node.inputs[1])
+        self.assertEqual(gather_node.inputs[1].name, "const_1_i64")
+
+    def test_int_list_no_clash_across_typed_and_untyped_contexts(self):
+        """Test that the same int list used in typed and untyped positions
+        does not cause an initializer name collision (sequence variant).
+        """
+        graph = ir.Graph(
+            name="test_model",
+            inputs=[],
+            outputs=[],
+            nodes=[],
+            opset_imports={"": _default_opset_version},
+        )
+        x = ir.Value(name="x", type=ir.TensorType(ir.DataType.INT64), shape=ir.Shape([2, 3]))
+        graph.inputs.append(x)
+
+        graph_builder = builder.GraphBuilder(graph)
+        op = graph_builder.op
+
+        # Reshape target: int list in untyped position
+        _ = op.Reshape(x, [3, 2])
+        # Add with a constant tensor of same values in typed position
+        _ = op.Add(x, [3, 2])
+
+        # Should not raise; both should share the same initializer
+        nodes = list(graph)
+        reshape_node = nodes[0]
+        add_node = nodes[1]
+        self.assertIs(reshape_node.inputs[1], add_node.inputs[1])
+
     def test_pop_module_raises_on_empty_stack(self):
         """Test that pop_module raises RuntimeError when no module has been pushed."""
         op, _, _ = _create_builder_with_inputs()