[ONNX] handle aten::_set_item on Dict in convertInplaceOpsAndTrackAlias #58317

garymm · 2021-05-14T19:32:52Z

It seems the JIT produces an output for aten::_set_item on lists but
not on dicts. Previously the code would crash because it assumed it
was operating on a list.

The different behavior can be seen with the following test:

class DictModule(torch.nn.Module):
    def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]:
        x_out = {}
        x_out["test_key_out"] = x_in
        return x_out
    
x_in = torch.tensor(1)
dms = torch.jit.script(DictModule())
torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),))

Before this change:
RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.

After this change:
RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.

This is a more useful error message.

facebook-github-bot · 2021-05-14T19:32:58Z

💊 CI failures summary and remediations

As of commit c80837f (more details on the Dr. CI page):

1/1 failures introduced in this PR

🕵️ 1 new failure recognized by patterns

The following CI failures do not appear to be due to upstream breakages:

pytorch_xla_linux_bionic_py3_6_clang9_build (1/1)

Step: "Build" (full log | diagnosis details | 🔁 rerun)

May 18 21:47:06 AssertionError: Found an invalid operator name: _copy_from_and_resize

May 18 21:47:06   File "/opt/conda/lib/python3.6/runpy.py", line 193, in _run_module_as_main
May 18 21:47:06     "__main__", mod_spec)
May 18 21:47:06   File "/opt/conda/lib/python3.6/runpy.py", line 85, in _run_code
May 18 21:47:06     exec(code, run_globals)
May 18 21:47:06   File "/var/lib/jenkins/workspace/tools/codegen/gen_backend_stubs.py", line 126, in <module>
May 18 21:47:06     main()
May 18 21:47:06   File "/var/lib/jenkins/workspace/tools/codegen/gen_backend_stubs.py", line 87, in main
May 18 21:47:06     cpp_namespace, external_backend_functions = parse_backend_yaml(options.source_yaml, grouped_native_functions)
May 18 21:47:06   File "/var/lib/jenkins/workspace/tools/codegen/gen_backend_stubs.py", line 61, in parse_backend_yaml
May 18 21:47:06     raise AssertionError(f"Found an invalid operator name: {op_name}")
May 18 21:47:06 AssertionError: Found an invalid operator name: _copy_from_and_resize
May 18 21:47:06 ~/workspace/xla
May 18 21:47:06 + OPTS=()
May 18 21:47:06 + getopts O: OPTION
May 18 21:47:06 + case $OPTION in
May 18 21:47:06 + for i in ${OPTARG}
May 18 21:47:06 + OPTS+=("--cxxopt=${i}")
May 18 21:47:06 + getopts O: OPTION
May 18 21:47:06 + shift 2
May 18 21:47:06 + CMD=install
May 18 21:47:06 ++ dirname /var/lib/jenkins/workspace/xla/build_torch_xla_libs.sh

XLA failure

Job pytorch_xla_linux_bionic_py3_6_clang9_build is failing. Please create an issue with title prefixed by [PT_BREAK] in pytorch/xla and link to to this PR. If you have questions, please reach out to @ailzhang / @dlibenzi / @JackCaoG.

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

BowenBao · 2021-05-17T23:42:22Z

Thanks for the fix! PR looks good, could you include the test case in the description in unit test as well, with assertRaises?

torch/csrc/jit/passes/onnx/remove_inplace_ops_for_onnx.cpp

garymm · 2021-05-18T00:00:54Z

Thanks for the fix! PR looks good, could you include the test case in the description in unit test as well, with assertRaises?

Good idea. Done!
BTW, Is there a way that we can more easily differentiate between assertion failures in the JIT and unsupported operations? Perhaps with a different exception type?

BowenBao · 2021-05-18T00:59:55Z

@garymm That would be a great idea, i think currently we don't have a specialized exception type for onnx export. It'd be interesting to see how to use a same exception type across C++ and Python code, maybe need to touch torch/csrc/Exceptions.h and make sure it works with pybind.
Another point I see in this example is that there are cases of assertion errors that were actually caused by unsupported operations, yet it failed early at assertions before reaching the place to print a more meaningful error. I guess we need to be more careful about what we assume when writing passes.

garymm · 2021-05-18T01:52:16Z

Filed https://msdata.visualstudio.com/Vienna/_workitems/edit/1185416/.

Please take another look at this PR and approve and merge if ready.

Thanks!

It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) Before this change: RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch. After this change: RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub. This is a more useful error message.

…as (#58317) It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Co-authored-by: Gary Miguel <garymiguel@microsoft.com> [ghstack-poisoned]

…as (#58317) It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Co-authored-by: Gary Miguel <garymiguel@microsoft.com>

…as (#58317) It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Co-authored-by: Gary Miguel <garymiguel@microsoft.com> [ghstack-poisoned]

…AndTrackAlias (#58317)" It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Co-authored-by: Gary Miguel <garymiguel@microsoft.com> [ghstack-poisoned]

…AndTrackAlias (#58317)" It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Co-authored-by: Gary Miguel <garymiguel@microsoft.com> Differential Revision: [D28714804](https://our.internmc.facebook.com/intern/diff/D28714804) [ghstack-poisoned]

…as (#58317) (#58696) Summary: Pull Request resolved: #58696 It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Test Plan: Imported from OSS Reviewed By: driazati Differential Revision: D28714804 Pulled By: SplitInfinity fbshipit-source-id: 1e5dc5fb44d1e3f971a22a79b5cf009d7590bf84 Co-authored-by: Gary Miguel <garymiguel@microsoft.com>

…as (pytorch#58317) (pytorch#58696) Summary: Pull Request resolved: pytorch#58696 It seems the JIT produces an output for aten::_set_item on lists but not on dicts. Previously the code would crash because it assumed it was operating on a list. The different behavior can be seen with the following test: ```python class DictModule(torch.nn.Module): def forward(self, x_in: torch.Tensor) -> typing.Dict[str, torch.Tensor]: x_out = {} x_out["test_key_out"] = x_in return x_out x_in = torch.tensor(1) dms = torch.jit.script(DictModule()) torch.onnx.export(dms, (x_in,), "/dev/null", example_outputs=(dms(x_in),)) ``` Before this change: `RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":452, please report a bug to PyTorch.` After this change: `RuntimeError: Exporting the operator prim_DictConstruct to ONNX opset version 9 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub.` This is a more useful error message. Test Plan: Imported from OSS Reviewed By: driazati Differential Revision: D28714804 Pulled By: SplitInfinity fbshipit-source-id: 1e5dc5fb44d1e3f971a22a79b5cf009d7590bf84 Co-authored-by: Gary Miguel <garymiguel@microsoft.com>

garymm requested review from BowenBao, neginraoof and spandantiwari as code owners May 14, 2021 19:32

facebook-github-bot added the cla signed label May 14, 2021

facebook-github-bot added the oncall: jit Add this issue/PR to JIT oncall triage queue label May 14, 2021

pytorchbot added the open source label May 14, 2021

garymm force-pushed the set-item branch from 8b0622b to 61cda7b Compare May 14, 2021 19:36

BowenBao reviewed May 17, 2021

View reviewed changes

torch/csrc/jit/passes/onnx/remove_inplace_ops_for_onnx.cpp Show resolved Hide resolved

garymm force-pushed the set-item branch from 61cda7b to e6c5142 Compare May 17, 2021 23:58

garymm force-pushed the set-item branch from e6c5142 to 782afb7 Compare May 18, 2021 21:17

garymm force-pushed the set-item branch from 782afb7 to c80837f Compare May 18, 2021 21:23

BowenBao merged commit d2f2a03 into pytorch:onnx_ms_1 May 19, 2021

garymm deleted the set-item branch May 19, 2021 16:40

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[ONNX] handle aten::_set_item on Dict in convertInplaceOpsAndTrackAlias #58317

[ONNX] handle aten::_set_item on Dict in convertInplaceOpsAndTrackAlias #58317

garymm commented May 14, 2021 •

edited

Loading

facebook-github-bot commented May 14, 2021 •

edited

Loading

BowenBao commented May 17, 2021

garymm commented May 18, 2021

BowenBao commented May 18, 2021

garymm commented May 18, 2021

[ONNX] handle aten::_set_item on Dict in convertInplaceOpsAndTrackAlias #58317

[ONNX] handle aten::_set_item on Dict in convertInplaceOpsAndTrackAlias #58317

Conversation

garymm commented May 14, 2021 • edited Loading

facebook-github-bot commented May 14, 2021 • edited Loading

💊 CI failures summary and remediations

🕵️ 1 new failure recognized by patterns

pytorch_xla_linux_bionic_py3_6_clang9_build (1/1)

XLA failure

BowenBao commented May 17, 2021

garymm commented May 18, 2021

BowenBao commented May 18, 2021

garymm commented May 18, 2021

garymm commented May 14, 2021 •

edited

Loading

facebook-github-bot commented May 14, 2021 •

edited

Loading