[ONNX] Add support for torch.cond/HOP in onnx exporter #137428
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/137428
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 3 New Failures, 1 Unrelated FailureAs of commit 9f081b8 with merge base 72943ba ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
titaiwangms
reviewed
Oct 7, 2024
|
Function vs subgraph: Could you share more details on how the fx subgraphs are represented differently than onnx subgraphs? This can inform us on what the best representation is. |
gramalingam
reviewed
Oct 8, 2024
justinchuby
reviewed
Nov 4, 2024
justinchuby
reviewed
Nov 4, 2024
justinchuby
requested changes
Nov 4, 2024
There was a problem hiding this comment.
Thanks!
- I think we should isolate logic handling a particular operator from the _core module. Any logic there would better be generic for handling HOPs, and specific logic for handling cond can be in a separate location which can be later integrated into torchlib (when it is migrated over)
- In general, we should avoid any proto object manipulation in the
exporterdirectory as they introduce unnecessary overhead and creates inconsistencies in how we manipulate the onnx graph.
justinchuby
reviewed
Nov 4, 2024
|
|
@xadupre and I talked and we have some ideas to simplify some of the function calls. I will play with it a little and update here if I make any progress |
titaiwangms
reviewed
Nov 5, 2024
titaiwangms
reviewed
Nov 19, 2024
| fx_graph = module.graph | ||
|
|
||
| graph_like: ir.Graph | ir.Function | ||
| if name == "": |
There was a problem hiding this comment.
Do you think torch.export would change this assumption? I am a bit concerned.
There was a problem hiding this comment.
Do you mean the name? I think the root module always have an empty name, but I should double check
titaiwangms
reviewed
Nov 19, 2024
gramalingam
reviewed
Nov 19, 2024
<
ir_version=9,
opset_imports={'pkg.onnxscript.torch_lib.common': 1, '': 18, 'pkg.torch.__subgraph__': 1, 'pkg.onnxscript.torch_lib': 1},
producer_name='pytorch',
producer_version='2.6.0a0+git3cd6dd5',
domain=None,
model_version=None,
>
graph(
name=main_graph,
inputs=(
%"x"<INT64,[2]>
),
outputs=(
%"getitem"<FLOAT,[2]>
),
initializers=(
%"weight"<FLOAT,[1]>,
%"submodule.weight"<FLOAT,[1]>
),
) {
0 | # node_ReduceSum_0
%"sum_1"<INT64,[]> ⬅️ ::ReduceSum(%"x") {keepdims=False, noop_with_empty_axes=0}
1 | # node_Constant_1
%"val_0"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(0), name=None)}
2 | # node_Greater_2
%"gt"<BOOL,[]> ⬅️ ::Greater(%"sum_1", %"val_0")
3 | # node_If_3
%"getitem"<FLOAT,[2]> ⬅️ ::If(%"gt") {then_branch=
graph(
name=true_graph_0,
inputs=(
),
outputs=(
%"getitem_true_graph_0"<?,?>
),
) {
0 | # node_true_graph_0_0
%"getitem_true_graph_0"<?,?> ⬅️ pkg.torch.__subgraph__::true_graph_0(%"weight", %"x", %"submodule.weight")
return %"getitem_true_graph_0"<?,?>
}, else_branch=
graph(
name=false_graph_0,
inputs=(
),
outputs=(
%"sub_false_graph_0"<?,?>
),
) {
0 | # node_false_graph_0_0
%"sub_false_graph_0"<?,?> ⬅️ pkg.torch.__subgraph__::false_graph_0(%"weight", %"x", %"submodule.weight")
return %"sub_false_graph_0"<?,?>
}}
return %"getitem"<FLOAT,[2]>
}
<
opset_imports={'': 18},
>
def pkg.torch.__subgraph__::false_graph_0(
inputs=(
%"p_weight"<FLOAT,[1]>,
%"x"<INT64,[2]>,
%"p_submodule_weight"<FLOAT,[1]>
),
outputs=(
%"sub"<FLOAT,[2]>
),
) {
0 | # node_Cast_0
%"convert_element_type_default"<FLOAT,[2]> ⬅️ ::Cast(%"x") {to=FLOAT}
1 | # node_Sub_1
%"sub"<FLOAT,[2]> ⬅️ ::Sub(%"convert_element_type_default", %"p_weight")
return %"sub"<FLOAT,[2]>
}
<
opset_imports={'pkg.onnxscript.torch_lib': 1},
>
def pkg.torch.__subgraph__::true_graph_0__false_graph_0(
inputs=(
%"p_submodule_weight"<FLOAT,[1]>,
%"sub"<FLOAT,[2]>
),
outputs=(
%"div"<FLOAT,[2]>
),
) {
0 | # node_aten_div_0
%"div"<FLOAT,[2]> ⬅️ pkg.onnxscript.torch_lib::aten_div(%"sub", %"p_submodule_weight")
return %"div"<FLOAT,[2]>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib::aten_div(
inputs=(
%"self"<?,?>,
%"other"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"return_val"<?,?> ⬅️ ::Div(%"self", %"other")
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.torch.__subgraph__::true_graph_0__true_graph_0(
inputs=(
%"p_submodule_weight"<FLOAT,[1]>,
%"sub"<FLOAT,[2]>
),
outputs=(
%"mul"<FLOAT,[2]>
),
) {
0 | # node_Mul_0
%"mul"<FLOAT,[2]> ⬅️ ::Mul(%"sub", %"p_submodule_weight")
return %"mul"<FLOAT,[2]>
}
<
opset_imports={'': 18, 'pkg.torch.__subgraph__': 1},
>
def pkg.torch.__subgraph__::true_graph_0(
inputs=(
%"p_weight"<FLOAT,[1]>,
%"x"<INT64,[2]>,
%"p_submodule_weight"<FLOAT,[1]>
),
outputs=(
%"getitem"<FLOAT,[2]>
),
) {
0 | # node_Cast_0
%"convert_element_type_default"<FLOAT,[2]> ⬅️ ::Cast(%"x") {to=FLOAT}
1 | # node_Sub_1
%"sub"<FLOAT,[2]> ⬅️ ::Sub(%"convert_element_type_default", %"p_weight")
2 | # node_ReduceSum_2
%"sum_1"<FLOAT,[]> ⬅️ ::ReduceSum(%"sub") {keepdims=False, noop_with_empty_axes=0}
3 | # node_Constant_3
%"val_0"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(0), name=None)}
4 | # node_Cast_4
%"scalar_tensor_default"<FLOAT,[]> ⬅️ ::Cast(%"val_0") {to=FLOAT}
5 | # node_LessOrEqual_5
%"le"<BOOL,[]> ⬅️ ::LessOrEqual(%"sum_1", %"scalar_tensor_default")
6 | # node_If_6
%"getitem"<FLOAT,[2]> ⬅️ ::If(%"le") {then_branch=
graph(
name=true_graph_0__true_graph_0,
inputs=(
),
outputs=(
%"mul_true_graph_0__true_graph_0"<?,?>
),
) {
0 | # node_true_graph_0__true_graph_0_0
%"mul_true_graph_0__true_graph_0"<?,?> ⬅️ pkg.torch.__subgraph__::true_graph_0__true_graph_0(%"p_submodule_weight", %"sub")
return %"mul_true_graph_0__true_graph_0"<?,?>
}, else_branch=
graph(
name=true_graph_0__false_graph_0,
inputs=(
),
outputs=(
%"div_true_graph_0__false_graph_0"<?,?>
),
) {
0 | # node_true_graph_0__false_graph_0_0
%"div_true_graph_0__false_graph_0"<?,?> ⬅️ pkg.torch.__subgraph__::true_graph_0__false_graph_0(%"p_submodule_weight", %"sub")
return %"div_true_graph_0__false_graph_0"<?,?>
}}
return %"getitem"<FLOAT,[2]>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib.common::Rank(
inputs=(
%"input"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"tmp"<?,?> ⬅️ ::Shape(%"input")
1 | # n1
%"return_val"<?,?> ⬅️ ::Size(%"tmp")
return %"return_val"<?,?>
}
<
opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib.common::IsScalar(
inputs=(
%"input"<?,?>
),
outputs=(
%"return_val"<?,?>
),
) {
0 | # n0
%"tmp"<?,?> ⬅️ ::Shape(%"input")
1 | # n1
%"tmp_0"<?,?> ⬅️ ::Size(%"tmp")
2 | # n2
%"tmp_1"<?,?> ⬅️ ::Constant() {value_int=0}
3 | # n3
%"return_val"<?,?> ⬅️ ::Equal(%"tmp_0", %"tmp_1")
return %"return_val"<?,?>
} |
|
@pytorchbot merge -i |
|
Will address new comments in a follow up PR |
Merge startedYour change will be merged while ignoring the following 4 checks: pull / linux-focal-py3_9-clang9-xla / test (xla, 1, 1, lf.linux.12xlarge), pull / linux-jammy-py3-clang12-executorch / test (executorch, 1, 1, lf.linux.2xlarge), pull / linux-jammy-py3.9-gcc11 / test (docs_test, 1, 1, lf.linux.2xlarge), pull / linux-jammy-py3.9-gcc11 / test (backwards_compat, 1, 1, lf.linux.2xlarge) Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
youssef62
pushed a commit
to youssef62/pytorch
that referenced
this pull request
Nov 23, 2024
This PR implements the framework for supporting HOP in the ONNX exporter. Refer to pytorch#140995 for the design. - Implement support for torch.cond - Refactor `_add_nodes` into `_translate_fx_graph` to handle nested subgraphs. To support building subgraphs as functions using the same logic, new handlers for `placeholder` and `output` nodes are added to register inputs and outputs on the onnx function. - Fuctions are created under the domain of `pkg.torch.__subgraph__` - Updated the type promotion pass to run on nested subgraphs. - Implement torch.cond in `_torchlib/ops/hop.py`. Updated the registry to discover these ops. - Improve opset_import handling robustness with `add_opset_imports` IR pass. To achieve this, we added opset version to all Nodes. Fixes pytorch#139503 Fixes pytorch#117655 Fixes pytorch#123972 Fixes pytorch#93743 Closes pytorch#140995 Pull Request resolved: pytorch#137428 Approved by: https://github.com/justinchuby Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
Ryo-not-rio
pushed a commit
to Ryo-not-rio/pytorch
that referenced
this pull request
Dec 2, 2024
This PR implements the framework for supporting HOP in the ONNX exporter. Refer to pytorch#140995 for the design. - Implement support for torch.cond - Refactor `_add_nodes` into `_translate_fx_graph` to handle nested subgraphs. To support building subgraphs as functions using the same logic, new handlers for `placeholder` and `output` nodes are added to register inputs and outputs on the onnx function. - Fuctions are created under the domain of `pkg.torch.__subgraph__` - Updated the type promotion pass to run on nested subgraphs. - Implement torch.cond in `_torchlib/ops/hop.py`. Updated the registry to discover these ops. - Improve opset_import handling robustness with `add_opset_imports` IR pass. To achieve this, we added opset version to all Nodes. Fixes pytorch#139503 Fixes pytorch#117655 Fixes pytorch#123972 Fixes pytorch#93743 Closes pytorch#140995 Pull Request resolved: pytorch#137428 Approved by: https://github.com/justinchuby Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
pobin6
pushed a commit
to pobin6/pytorch
that referenced
this pull request
Dec 5, 2024
This PR implements the framework for supporting HOP in the ONNX exporter. Refer to pytorch#140995 for the design. - Implement support for torch.cond - Refactor `_add_nodes` into `_translate_fx_graph` to handle nested subgraphs. To support building subgraphs as functions using the same logic, new handlers for `placeholder` and `output` nodes are added to register inputs and outputs on the onnx function. - Fuctions are created under the domain of `pkg.torch.__subgraph__` - Updated the type promotion pass to run on nested subgraphs. - Implement torch.cond in `_torchlib/ops/hop.py`. Updated the registry to discover these ops. - Improve opset_import handling robustness with `add_opset_imports` IR pass. To achieve this, we added opset version to all Nodes. Fixes pytorch#139503 Fixes pytorch#117655 Fixes pytorch#123972 Fixes pytorch#93743 Closes pytorch#140995 Pull Request resolved: pytorch#137428 Approved by: https://github.com/justinchuby Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
73 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR implements the framework for supporting HOP in the ONNX exporter. Refer to #140995 for the design.
_add_nodesinto_translate_fx_graphto handle nested subgraphs. To support building subgraphs as functions using the same logic, new handlers forplaceholderandoutputnodes are added to register inputs and outputs on the onnx function.pkg.torch.__subgraph___torchlib/ops/hop.py. Updated the registry to discover these ops.add_opset_importsIR pass. To achieve this, we added opset version to all Nodes. Fixes [ONNX] Add opset version to individual nodes when building the graph #139503Fixes #117655 Fixes #123972 Ref #93743 Closes #140995