[API] Create stable APIs for PyTorch 2.6

[justinchuby]

• optimize is turned on. It will be controlled by an option in PyTorch
• Remove the _TORCH_ONNX_SAVE_EXTERNAL_DATA_WITH_IR flag

[codecov]

❌ 13 Tests Failed:

Tests completed	Failed	Passed	Skipped
13304	13	13291	2298

View the top 3 failed tests by shortest run time

tests.ir.serde_roundtrip_test.SerdeTest test_serialization_deserialization_produces_same_model_0756_test_cast_FLOAT4E2M1_to_FLOAT_model_onnx

Stack Traces | 0.001s run time

onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:711: in deserialize_type_proto_for_type
    return _core.TensorType(_enums.DataType(elem_type), denotation=denotation)
.../hostedtoolcache/Python/3.11.10....../x64/lib/python3.11/enum.py:714: in __call__
    return cls.__new__(cls, value)
.../hostedtoolcache/Python/3.11.10....../x64/lib/python3.11/enum.py:1137: in __new__
    raise ve_exc
E   ValueError: 23 is not a valid DataType

The above exception was the direct cause of the following exception:
onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:657: in deserialize_value_info_proto
    value.type = deserialize_type_proto_for_type(proto.type)
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling deserialize_type_proto_for_type with: tensor_type {
E     elem_type: 23
E     shape {
E       dim {
E         dim_value: 3
E       }
E       dim {
E         dim_value: 5
E       }
E     }
E   }

The above exception was the direct cause of the following exception:
onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:568: in _deserialize_graph
    deserialize_value_info_proto(info, value)
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling deserialize_value_info_proto with: name: "input"
E   type {
E     tensor_type {
E       elem_type: 23
E       shape {
E         dim {
E           dim_value: 3
E         }
E         dim {
E           dim_value: 5
E         }
E       }
E     }
E   }

The above exception was the direct cause of the following exception:
..../test_onnx_weekly/lib/python3.11.../site-packages/parameterized/parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
tests/ir/serde_roundtrip_test.py:38: in test_serialization_deserialization_produces_same_model
    ir_model = ir.serde.deserialize_model(model)
onnxscript/ir/serde.py:459: in deserialize_model
    graph = _deserialize_graph(proto.graph, [])
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling _deserialize_graph with: main_graph

tests.ir.serde_roundtrip_test.SerdeTest test_serialization_deserialization_produces_same_model_1278_test_cast_FLOAT4E2M1_to_FLOAT16_model_onnx

Stack Traces | 0.001s run time

onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:711: in deserialize_type_proto_for_type
    return _core.TensorType(_enums.DataType(elem_type), denotation=denotation)
.../hostedtoolcache/Python/3.11.10....../x64/lib/python3.11/enum.py:714: in __call__
    return cls.__new__(cls, value)
.../hostedtoolcache/Python/3.11.10....../x64/lib/python3.11/enum.py:1137: in __new__
    raise ve_exc
E   ValueError: 23 is not a valid DataType

The above exception was the direct cause of the following exception:
onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:657: in deserialize_value_info_proto
    value.type = deserialize_type_proto_for_type(proto.type)
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling deserialize_type_proto_for_type with: tensor_type {
E     elem_type: 23
E     shape {
E       dim {
E         dim_value: 3
E       }
E       dim {
E         dim_value: 5
E       }
E     }
E   }

The above exception was the direct cause of the following exception:
onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:568: in _deserialize_graph
    deserialize_value_info_proto(info, value)
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling deserialize_value_info_proto with: name: "input"
E   type {
E     tensor_type {
E       elem_type: 23
E       shape {
E         dim {
E           dim_value: 3
E         }
E         dim {
E           dim_value: 5
E         }
E       }
E     }
E   }

The above exception was the direct cause of the following exception:
..../test_onnx_weekly/lib/python3.11.../site-packages/parameterized/parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
tests/ir/serde_roundtrip_test.py:38: in test_serialization_deserialization_produces_same_model
    ir_model = ir.serde.deserialize_model(model)
onnxscript/ir/serde.py:459: in deserialize_model
    graph = _deserialize_graph(proto.graph, [])
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling _deserialize_graph with: main_graph

tests.ir.serde_roundtrip_test.SerdeTest test_serialization_deserialization_produces_same_model_1438_test_dequantizelinear_float4e2m1_model_onnx

Stack Traces | 0.002s run time

onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:711: in deserialize_type_proto_for_type
    return _core.TensorType(_enums.DataType(elem_type), denotation=denotation)
.../hostedtoolcache/Python/3.11.10....../x64/lib/python3.11/enum.py:714: in __call__
    return cls.__new__(cls, value)
.../hostedtoolcache/Python/3.11.10....../x64/lib/python3.11/enum.py:1137: in __new__
    raise ve_exc
E   ValueError: 23 is not a valid DataType

The above exception was the direct cause of the following exception:
onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:657: in deserialize_value_info_proto
    value.type = deserialize_type_proto_for_type(proto.type)
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling deserialize_type_proto_for_type with: tensor_type {
E     elem_type: 23
E     shape {
E       dim {
E         dim_value: 5
E       }
E     }
E   }

The above exception was the direct cause of the following exception:
onnxscript/ir/serde.py:97: in wrapper
    return func(*args, **kwargs)
onnxscript/ir/serde.py:568: in _deserialize_graph
    deserialize_value_info_proto(info, value)
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling deserialize_value_info_proto with: name: "x"
E   type {
E     tensor_type {
E       elem_type: 23
E       shape {
E         dim {
E           dim_value: 5
E         }
E       }
E     }
E   }

The above exception was the direct cause of the following exception:
..../test_onnx_weekly/lib/python3.11.../site-packages/parameterized/parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
tests/ir/serde_roundtrip_test.py:38: in test_serialization_deserialization_produces_same_model
    ir_model = ir.serde.deserialize_model(model)
onnxscript/ir/serde.py:459: in deserialize_model
    graph = _deserialize_graph(proto.graph, [])
onnxscript/ir/serde.py:99: in wrapper
    raise SerdeError(
E   onnxscript.ir.serde.SerdeError: Error calling _deserialize_graph with: main_graph

To view individual test run time comparison to the main branch, go to the Test Analytics Dashboard

[titaiwangms]

Could you remind me again where to control the version of api when users try to export?

[justinchuby]

Each pytorch release will use a version that is design for that release. We are responsible to update the pytorch code to use the correct version. For example, we should go in now and change pytorch to use the 2.6 api.