[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp by alexdean08 · Pull Request #14415 · pytorch/executorch

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 · 2025-09-18T19:14:26Z

With the motivation of improving performance, this change adds the functionality for fusing the following ops:

conv2d PW s1p0 and binary ops (add, sub, mul, div)
clamp and binary ops (add, sub, mul, div)
clamp and clamp

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

pytorch-bot · 2025-09-18T19:14:30Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 8 Pending

As of commit ecc521f with merge base 2eb8994 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6508b71c03793e86e0884165215be6514d4319839ae5ad6884ef564b975b8413 /exec failed with exit code 1
pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
RuntimeError: Command docker exec -t 1968a27aca9b7cc1e61cd5413d552a246ece245b60a56ba93dd946f81127bb43 /exec failed with exit code 1
Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
RuntimeError: Command docker exec -t 7afe43e3298b5d7815e13952437c7fbed77e4f1ca3bbccb9add267709e941e46 /exec failed with exit code 2

This comment was automatically generated by Dr. CI and updates every 15 minutes.

SS-JIA · 2025-09-30T21:20:22Z

+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys


@alexdean08 would you mind taking a stab at implementing the fusions via the new patterns/ folder I added? This folder is meant to abstract matching fusable operator patterns and fusing them into a single op.
Here is an example for quantized convolution: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/quantized_convolution.py
Basically, you will need to implement 3 things in a file called, say patterns/fused_convolution.py
# patterns/fused_convolution.pyFUSEABLE_OPS= [ exir_ops.edge.aten.relu.default, exir_ops.edge.aten.hardtanh.default, exir_ops.edge.aten.clamp.default, ] FUSEABLE_BINARY_OPS= [ exir_ops.edge.aten.add.Tensor, exir_ops.edge.aten.sub.Tensor, exir_ops.edge.aten.mul.Tensor, exir_ops.edge.aten.div.Tensor, ] # Implement a Match class to represent the patterns involved in a pattern matchclassFusedConvolutionMatch(PatternMatch): def__init__(self, conv_node: torch.fx.Node) ->None: self.anchor_node=conv_node# Store all relevant args here ... first_user=list(conv_node.users)[0] iffirst_usernotinFUSABLE_OPS: returnself.output_node=first_user# Store other relevant args here ... self.match_found=Trueconvolution_anchor_nodes= { exir_ops.edge.aten.conv2d.default, exir_ops.edge.aten.convolution.default, } @register_pattern_detector("fused_convolution")deffind_fused_convolution_patterns( node: torch.fx.Node, ) ->Optional[FusedConvolutionMatch]: ifnode.targetnotinconvolution_anchor_nodes: returnNonematched_pattern=FusedConvolutionMatch(node) ifmatched_pattern.match_found: returnmatched_patternreturnNone#### Pattern Replacement##FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET= { exir_ops.edge.aten.add.Tensor: exir_ops.edge.et_vk.conv_with_binary_add.default ... } @register_pattern_replacement("fused_convolution")defmake_fused_conv2d_custom_op( ep: ExportedProgram, graph_module: torch.fx.GraphModule, match: QuantizedConvolutionMatch, ): withgraph_module.graph.inserting_before(match.output_node): op_target=FUSED_NODE_TARGET_TO_CUSTOM_OP_TARGET[match.output_node.target] custom_op_node=graph_module.graph.create_node( "call_function", op_target, args=( ... ), ) qconv_node.meta["val"] =match.output_node.meta["val"] match.output_node.replace_all_uses_with(custom_op_node)
This should simplify the logic a bit and allow us to reduce the amount of transforms we need to maintain. Although I think fuse_clamps.py still needs to be its own pass.
Once you've added the pattern file, remember to import it here: https://github.com/pytorch/executorch/blob/main/backends/vulkan/patterns/__init__.py so that registrations are triggered when importing the module.

Also for fusing clamps with binaries, how impactful is it? If the impact is not super significant, I think I would prefer to avoid fusing for the sake of simplifying our codebase.

SS-JIA

Runtime side LGTM. Just comments about how op fusions are implemented. Thanks for working on this!

SS-JIA

LGTM. All tests are passing.

SS-JIA · 2025-10-09T20:20:16Z

@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature

@SS-JIA

…lamp (#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin (cherry picked from commit a5d7e5c)

pytorchbot · 2025-10-09T20:22:36Z

Cherry picking #14415

The cherry pick PR is at #14960 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated:

[v1.0.0] Release Tracker #14288 (comment)

Details for Dev Infra team

Raised by workflow job

… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

@SS-JIA

#15066) … Clamp/Clamp (#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in #15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

@SS-JIA

…lamp (pytorch#14415) With the motivation of improving performance, this change adds the functionality for fusing the following ops: - conv2d PW s1p0 and binary ops (add, sub, mul, div) - clamp and binary ops (add, sub, mul, div) - clamp and clamp cc @SS-JIA @manuelcandales @digantdesai @cbilgin

@SS-JIA

pytorch#15066) … Clamp/Clamp (pytorch#14415)" This reverts commit a5d7e5c. Broke internal builds @SS-JIA is trying to fix this in pytorch#15058 will leave relanding to him ### Summary [PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines. [PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line. [PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests). ### Test plan [PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

alexdean08 added 5 commits September 10, 2025 14:11

Optimize conv2d s1p0
8e3d21e

Stylistic changes to pw conv2d s1p0
41ec8b6

Merge branch 'main' into PWConvOptimized
7f4345d

Add fp16 to conv2d pw s1p0
c1910fe

Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and Clamp/Clamp
a72e2e4

alexdean08 requested review from SS-JIA and kimishpatel as code owners September 18, 2025 19:14

alexdean08 added module: vulkan Issues related to the Vulkan delegate and code under backends/vulkan/ release notes: vulkan Changes to the Vulkan backend delegate labels Sep 18, 2025

meta-claBot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 18, 2025

Lint fixes
7e8d9b1

SS-JIA reviewed Sep 30, 2025
View reviewed changes

SS-JIA requested changes Sep 30, 2025
View reviewed changes

Remove conv + binary ops fusing and change binary op + clamp fusing
ffb0cdf

alexdean08 force-pushed the FuseOps branch from ccf4aeb to ffb0cdfCompare October 9, 2025 06:53

alexdean08 added 2 commits October 8, 2025 23:58

Merge branch 'main' into FuseOps
58c1838

Lint fix
4e64691

SS-JIA added the ciflow/nightly label Oct 9, 2025

Change order of fusing passes
ecc521f

SS-JIA approved these changes Oct 9, 2025
View reviewed changes

SS-JIA merged commit a5d7e5c into pytorch:mainOct 9, 2025
176 of 180 checks passed

pytorchbot mentioned this pull request Oct 9, 2025
[v1.0.0] Release Tracker #14288
Closed

JacobSzwejbka added a commit that referenced this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and…
b196b1e
… Clamp/Clamp (#14415)" This reverts commit a5d7e5c.

JacobSzwejbka mentioned this pull request Oct 13, 2025
Revert "[ET-VK] Add Fusing for Conv/Binary Ops, Clamp/Binary Ops, and… #15066
Merged

xuyanwen2012 deleted the FuseOps branch June 30, 2026 21:58

Uh oh!

Conversation

alexdean08 commented Sep 18, 2025• edited by pytorch-bot BotLoading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-botBot commented Sep 18, 2025• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

❌ 3 New Failures, 8 Pending

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

SS-JIA commented Oct 9, 2025

Uh oh!

pytorchbot commented Oct 9, 2025

Cherry picking #14415

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

alexdean08 commented Sep 18, 2025• edited by pytorch-bot BotLoading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-botBot commented Sep 18, 2025• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

❌ 3 New Failures, 8 Pending

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

SS-JIA commented Oct 9, 2025

Uh oh!

pytorchbot commented Oct 9, 2025

Cherry picking #14415

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

alexdean08 commented Sep 18, 2025• edited by pytorch-bot BotLoading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-botBot commented Sep 18, 2025• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14415

❌ 3 New Failures, 8 Pending

alexdean08 commented Sep 18, 2025•
edited by pytorch-bot Bot
Loading

pytorch-botBot commented Sep 18, 2025•
edited
Loading

alexdean08 commented Sep 18, 2025•
edited by pytorch-bot Bot
Loading

pytorch-botBot commented Sep 18, 2025•
edited
Loading

alexdean08 commented Sep 18, 2025•
edited by pytorch-bot Bot
Loading

pytorch-botBot commented Sep 18, 2025•
edited
Loading

alexdean08 commented Sep 18, 2025•
edited by pytorch-bot Bot
Loading

pytorch-botBot commented Sep 18, 2025•
edited
Loading