Skip to content

fix export doc #26

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed

fix export doc #26

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ea3fa51
push
msaroufim Jul 26, 2023
f63a500
upsh
msaroufim Jul 26, 2023
05abcdf
push
msaroufim Jul 27, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
45 changes: 26 additions & 19 deletions docs/website/docs/tutorials/exporting_to_executorch.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,16 +72,20 @@ class MyModule(torch.nn.Module):

aten_dialect = exir.capture(MyModule(), (torch.randn(3, 4),))

print(aten_dialect.exported_program)
print(aten_dialect)
"""
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, arg0_1: f32[3, 4], arg1_1: f32[5, 4], arg2_1: f32[5], arg3_1: f32[3, 4]):
add: f32[3, 4] = torch.ops.aten.add.Tensor(arg3_1, arg0_1);
permute: f32[4, 5] = torch.ops.aten.permute_copy.default(arg1_1, [1, 0]);
addmm: f32[3, 5] = torch.ops.aten.addmm.default(arg2_1, add, permute);
clamp: f32[3, 5] = torch.ops.aten.clamp.default(addmm, 0.0, 1.0);
return (clamp,)
class GraphModule(torch.nn.Module):
def forward(self, arg0_1: f32[4, 4]):
# File: /Users/marksaroufim/Dev/zzz/test3.py:10, code: return self.linear(x)
_param_constant0 = self._param_constant0
t: f32[4, 4] = torch.ops.aten.t.default(_param_constant0); _param_constant0 = None
_param_constant1 = self._param_constant1
addmm: f32[4, 4] = torch.ops.aten.addmm.default(_param_constant1, arg0_1, t); _param_constant1 = arg0_1 = t = None
return [addmm]

Graph Signature: ExportGraphSignature(parameters=[], buffers=[], user_inputs=[], user_outputs=[], inputs_to_parameters={}, inputs_to_buffers={}, buffers_to_mutate={}, backward_signature=None, assertion_dep_token=None)
Symbol to range: {}
"""
```

Expand All @@ -106,18 +110,22 @@ This lowering will be done through the `to_edge()` API.

```python
aten_dialect = exir.capture(MyModule(), (torch.randn(3, 4),))
edge_dialect = aten_dialect.to_edge()
edge_dialect = aten_dialect.to_edge(exir.EdgeCompileConfig(_check_ir_validity=False))
Copy link
Member Author

@msaroufim msaroufim Jul 26, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

without this you get this error torch._export.verifier.SpecViolationError: Operator torch._ops.aten.t.default is not Aten Canonical. - should probably figure out how to make this error go away

I could for example get rid of this error by reworking the example to just do vector multiplication but like matmuls are probably more interesting lol https://gist.github.com/msaroufim/629b5c623fade2d5a30bec379f9e08da

Copy link
Contributor

@angelayi angelayi Jul 27, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This error should be fixed by D47346723 (which will be landed before PP) where aten.t will get decomposed to aten.permute which is ATen Canonical. We want to avoid users using the _check_ir_validity flag, but we should probably provide a better error message like "Please file an issue to executorch team, or turn on _check_ir_validity flag to unblock yourself for now"


print(edge_dialect.exported_program)
print(edge_dialect)
"""
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, arg0_1: f32[3, 4], arg1_1: f32[5, 4], arg2_1: f32[5], arg3_1: f32[3, 4]):
add: f32[3, 4] = executorch_exir_dialects_edge__ops_aten_add_Tensor(arg3_1, arg0_1);
permute: f32[4, 5] = executorch_exir_dialects_edge__ops_permute_copy_default(arg1_1, [1, 0]);
addmm: f32[3, 5] = executorch_exir_dialects_edge__ops_addmm_default(arg2_1, add, permute);
clamp: f32[3, 5] = executorch_exir_dialects_edge__ops_clamp_default(addmm, 0.0, 1.0);
return (clamp,)
class GraphModule(torch.nn.Module):
def forward(self, arg0_1: f32[3, 3]):
# File: /Users/marksaroufim/Dev/zzz/test3.py:10, code: return self.linear(x)
_param_constant0: f32[3, 3] = self._param_constant0
t_copy_default: f32[3, 3] = torch.ops.aten.t_copy.default(_param_constant0); _param_constant0 = None
_param_constant1: f32[3] = self._param_constant1
addmm_default: f32[3, 3] = torch.ops.aten.addmm.default(_param_constant1, arg0_1, t_copy_default); _param_constant1 = arg0_1 = t_copy_default = None
return [addmm_default]

Graph Signature: ExportGraphSignature(parameters=[], buffers=[], user_inputs=[], user_outputs=[], inputs_to_parameters={}, inputs_to_buffers={}, buffers_to_mutate={}, backward_signature=None, assertion_dep_token=None)
Symbol to range: {}
"""
```

Expand Down Expand Up @@ -185,8 +193,7 @@ be loaded in the Executorch runtime.

```python
edge_dialect = exir.capture(MyModule(), (torch.randn(3, 4),)).to_edge()
# edge_dialect = to_backend(edge_dialect.exported_program, CustomBackendPartitioner)
executorch_program = edge_dialect.to_executorch(executorch_backend_config)
executorch_program = edge_dialect.to_executorch()
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

partitioner and custom config could probably be their own section, regardless backend config is not defined here

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I also tried instead dumping in get_executorch_backend_config() but that wasn't right

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

partitioner and custom config could probably be their own section

This should be covered in the details of section 1.3, but I think you're right that we should move it.

backend config is not defined here

Backend config should be something passed in by the user, and covered in section 1.4.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah whoops I just found ExecutorchBackendConfig lemme just use an instance of that

buffer = executorch_program.buffer
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also one more issue I'm not sure how to resolve yet is you can't actually import in OSS

from executorch.extension.pybindings.portable import (
    _load_for_executorch_from_buffer,
)

I tried different buck incantations on the target but can't get the right one

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, from the pip built package we can't import files that are in the c++ land. :(
I'm not sure if the buck build will be able to help with this @dbort @cccclai

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maybe @malfet has some thoughts too since we'll need to figure this out soon enough


# Save it to a file and load it in the Executorch runtime
Expand Down