[ET-VK][Ops] affine quantization operators registration #12369
Merged
+239
−7
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
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
This was referenced Jul 10, 2025
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12369
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 78377d8 with merge base 1540659 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Jul 7, 2025
facebook-github-bot
added
the
CLA Signed
This was referenced Jul 10, 2025
ahmtox
pushed a commit
that referenced
this pull request
Jul 10, 2025
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) ghstack-source-id: 295489089 Pull Request resolved: #12369
This PR needs a
|
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
ahmtox
pushed a commit
that referenced
this pull request
Jul 11, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295685042 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
SS-JIA
approved these changes
Jul 11, 2025
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
ahmtox
pushed a commit
that referenced
this pull request
Jul 11, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. A full implementation for the affine operators will be done at a later time, since they are required for usage. However, if you wan't to just use the default as per_tensor then you must remove the checks made in `op_registry` and in the vulkan implementation so that the per_tensor version can be used. Without it they will not be registered. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295698280 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
ahmtox
pushed a commit
that referenced
this pull request
Jul 11, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. A full implementation for the affine operators will be done at a later time, since they are required for usage. However, if you wan't to just use the default as per_tensor then you must remove the checks made in `op_registry` and in the vulkan implementation so that the per_tensor version can be used. Without it they will not be registered. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295719089 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
ahmtox
pushed a commit
that referenced
this pull request
Jul 11, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. A full implementation for the affine operators will be done at a later time, since they are required for usage. However, if you wan't to just use the default as per_tensor then you must remove the checks made in `op_registry` and in the vulkan implementation so that the per_tensor version can be used. Without it they will not be registered. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295735692 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
ahmtox
pushed a commit
that referenced
this pull request
Jul 11, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. A full implementation for the affine operators will be done at a later time, since they are required for usage. However, if you wan't to just use the default as per_tensor then you must remove the checks made in `op_registry` and in the vulkan implementation so that the per_tensor version can be used. Without it they will not be registered. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295746674 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
# Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore. Furthermore, we have `_annotate_for_static_quantization_config` for parity/legacy reasons, and we simply create an equivalent dynamic quantization config method. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/) [ghstack-poisoned]
ahmtox
pushed a commit
that referenced
this pull request
Jul 14, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. A full implementation for the affine operators will be done at a later time, since they are required for usage. However, if you wan't to just use the default as per_tensor then you must remove the checks made in `op_registry` and in the vulkan implementation so that the per_tensor version can be used. Without it they will not be registered. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295972790 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
|
This pull request was exported from Phabricator. Differential Revision: D78035354 |
facebook-github-bot
merged commit 2ac8693
into
gh/ahmtox/39/base
98 of 100 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
SS-JIA
pushed a commit
that referenced
this pull request
Jul 14, 2025
Pull Request resolved: #12369 # Context In order to enable dynamic quantization, especially for the source transform method using `Int8DynActInt4WeightQuantizer` we need to have vulkan versions for `quantize_affine`, `dequantize_affine`, and `choose_qparams_affine`. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss. A full implementation for the affine operators will be done at a later time, since they are required for usage. However, if you wan't to just use the default as per_tensor then you must remove the checks made in `op_registry` and in the vulkan implementation so that the per_tensor version can be used. Without it they will not be registered. # Changes This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. We also changed `Linear.cpp`, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref. ghstack-source-id: 295972790 @exported-using-ghexport Differential Revision: [D78035354](https://our.internmc.facebook.com/intern/diff/D78035354/)
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.
Stack from ghstack (oldest at bottom):
Context
In order to enable dynamic quantization, especially for the source transform method using
Int8DynActInt4WeightQuantizerwe need to have vulkan versions forquantize_affine,dequantize_affine, andchoose_qparams_affine. Currently we do not have a shader that performs block-based quantization as expected from these shaders, so we delegate to the per_tensor variant just to get unblocked. At a later stage, this will likely be developed more on in order to ensure we don't get too much accuracy loss.Changes
This creates a schema reference in the TorchAO library for out variants of these respective operators. Then there is a VK_REGISTER_OP done on them to ensure that we can properly register them when lowering the ET model with vulkan. Also the vulkan_quantizer is changed a bit in this to enable a dynamic quantization config so that we aren't purely working with just static quantization anymore.
Furthermore, we have
_annotate_for_static_quantization_configfor parity/legacy reasons, and we simply create an equivalent dynamic quantization config method.We also changed
Linear.cpp, particularly to allow a passthrough for weight_data since during dynamic quantization it's possible that it'll be a tensor_data than tensor_ref.Differential Revision: D78035354