[ET-VK] Introduce generic export pass for fusing Q/DQ nodes #10525
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 When quantizing models with the PT2E quantization flow, quantize/dequantize nodes will be inserted into the graph. However, these quantize/dequantize nodes must be fused with operators such as `aten.linear.default` to produce nodes corresponding to quantized operators (e.g. `weight_int8pack_mm`) in order for quantized operator implementations to be called at runtime. Currently, the op fusion is done by the `fuse_dequant_linear.py` pass, however, this only handles one specific fusion pattern to generate a `weight_int8pack_mm` operator. As more quantized operators are to be supported in ET-VK via the PT2E quantization flow, a more generic fusion pass is needed that can handle a variety of fusion patterns. ## Changes Introduce the `FuseQuantizedOpsTransform()` pass. I elected to introduce a new pass under the `backends/vulkan/_passes` directory, as opposed to modifying the existing pass because I anticipate the majority of the fusion patterns to be specific to ET-VK. Remove the existing `FuseDequantLinearPass()` Switch to using the `FuseQuantizedOpsTransform` pass instead of the old `FuseDequantLinear` pass. Add `test_vulkan_passes` Python test to test export passes. Some small refactors to `test_vulkan_delegate` Python test to improve code organizations. Differential Revision: [D73794042](https://our.internmc.facebook.com/intern/diff/D73794042/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Apr 28, 2025
## Context When quantizing models with the PT2E quantization flow, quantize/dequantize nodes will be inserted into the graph. However, these quantize/dequantize nodes must be fused with operators such as `aten.linear.default` to produce nodes corresponding to quantized operators (e.g. `weight_int8pack_mm`) in order for quantized operator implementations to be called at runtime. Currently, the op fusion is done by the `fuse_dequant_linear.py` pass, however, this only handles one specific fusion pattern to generate a `weight_int8pack_mm` operator. As more quantized operators are to be supported in ET-VK via the PT2E quantization flow, a more generic fusion pass is needed that can handle a variety of fusion patterns. ## Changes Introduce the `FuseQuantizedOpsTransform()` pass. I elected to introduce a new pass under the `backends/vulkan/_passes` directory, as opposed to modifying the existing pass because I anticipate the majority of the fusion patterns to be specific to ET-VK. Remove the existing `FuseDequantLinearPass()` Switch to using the `FuseQuantizedOpsTransform` pass instead of the old `FuseDequantLinear` pass. Add `test_vulkan_passes` Python test to test export passes. Some small refactors to `test_vulkan_delegate` Python test to improve code organizations. Differential Revision: [D73794042](https://our.internmc.facebook.com/intern/diff/D73794042/) ghstack-source-id: 280746102 Pull Request resolved: #10525
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/10525
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 6c0e1e8 with merge base 280db15 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D73794042 |
This PR needs a
|
|
Why are you relyin on |
## Context When quantizing models with the PT2E quantization flow, quantize/dequantize nodes will be inserted into the graph. However, these quantize/dequantize nodes must be fused with operators such as `aten.linear.default` to produce nodes corresponding to quantized operators (e.g. `weight_int8pack_mm`) in order for quantized operator implementations to be called at runtime. Currently, the op fusion is done by the `fuse_dequant_linear.py` pass, however, this only handles one specific fusion pattern to generate a `weight_int8pack_mm` operator. As more quantized operators are to be supported in ET-VK via the PT2E quantization flow, a more generic fusion pass is needed that can handle a variety of fusion patterns. ## Changes Introduce the `FuseQuantizedOpsTransform()` pass. I elected to introduce a new pass under the `backends/vulkan/_passes` directory, as opposed to modifying the existing pass because I anticipate the majority of the fusion patterns to be specific to ET-VK. Remove the existing `FuseDequantLinearPass()` Switch to using the `FuseQuantizedOpsTransform` pass instead of the old `FuseDequantLinear` pass. Add `test_vulkan_passes` Python test to test export passes. Some small refactors to `test_vulkan_delegate` Python test to improve code organizations. Differential Revision: [D73794042](https://our.internmc.facebook.com/intern/diff/D73794042/) [ghstack-poisoned]
This was referenced Apr 30, 2025
SS-JIA
requested review from
jackzhxng,
iseeyuan,
larryliu0820 and
swolchok
as code owners
|
This pull request was exported from Phabricator. Differential Revision: D73794042 |
## Context When quantizing models with the PT2E quantization flow, quantize/dequantize nodes will be inserted into the graph. However, these quantize/dequantize nodes must be fused with operators such as `aten.linear.default` to produce nodes corresponding to quantized operators (e.g. `weight_int8pack_mm`) in order for quantized operator implementations to be called at runtime. Currently, the op fusion is done by the `fuse_dequant_linear.py` pass, however, this only handles one specific fusion pattern to generate a `weight_int8pack_mm` operator. As more quantized operators are to be supported in ET-VK via the PT2E quantization flow, a more generic fusion pass is needed that can handle a variety of fusion patterns. ## Changes Introduce the `FuseQuantizedOpsTransform()` pass. I elected to introduce a new pass under the `backends/vulkan/_passes` directory, as opposed to modifying the existing pass because I anticipate the majority of the fusion patterns to be specific to ET-VK. Remove the existing `FuseDequantLinearPass()` Switch to using the `FuseQuantizedOpsTransform` pass instead of the old `FuseDequantLinear` pass. Add `test_vulkan_passes` Python test to test export passes. Some small refactors to `test_vulkan_delegate` Python test to improve code organizations. Differential Revision: [D73794042](https://our.internmc.facebook.com/intern/diff/D73794042/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
May 1, 2025
Pull Request resolved: #10525 ## Context When quantizing models with the PT2E quantization flow, quantize/dequantize nodes will be inserted into the graph. However, these quantize/dequantize nodes must be fused with operators such as `aten.linear.default` to produce nodes corresponding to quantized operators (e.g. `weight_int8pack_mm`) in order for quantized operator implementations to be called at runtime. Currently, the op fusion is done by the `fuse_dequant_linear.py` pass, however, this only handles one specific fusion pattern to generate a `weight_int8pack_mm` operator. As more quantized operators are to be supported in ET-VK via the PT2E quantization flow, a more generic fusion pass is needed that can handle a variety of fusion patterns. ## Changes Introduce the `FuseQuantizedOpsTransform()` pass. I elected to introduce a new pass under the `backends/vulkan/_passes` directory, as opposed to modifying the existing pass because I anticipate the majority of the fusion patterns to be specific to ET-VK. Remove the existing `FuseDequantLinearPass()` Switch to using the `FuseQuantizedOpsTransform` pass instead of the old `FuseDequantLinear` pass. Add `test_vulkan_passes` Python test to test export passes. Added some refactors to `test_vulkan_delegate` Python test to improve code organization. Introduce the `linear_qcsnw` nomenclature: * q - quantized * c - per-channel / channelswise * s - symmetric * n - number of bits (qcs4w for 4-bit quant, qcs8w for 8-bit quant) * w - weight quantized Added custom op for `linear_qcs4w` for 4-bit weight quantized linear and add the ability for the quantized op fusion pass to produce this op. Slight renaming/refactoring of quantization config retrieval functions in the `VulkanQuantizer` to improve clarity and API flexibility. ghstack-source-id: 281448174 @exported-using-ghexport Differential Revision: [D73794042](https://our.internmc.facebook.com/intern/diff/D73794042/)
|
This pull request was exported from Phabricator. Differential Revision: D73794042 |
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
When quantizing models with the PT2E quantization flow, quantize/dequantize nodes will be inserted into the graph. However, these quantize/dequantize nodes must be fused with operators such as
aten.linear.defaultto produce nodes corresponding to quantized operators (e.g.weight_int8pack_mm) in order for quantized operator implementations to be called at runtime.Currently, the op fusion is done by the
fuse_dequant_linear.pypass, however, this only handles one specific fusion pattern to generate aweight_int8pack_mmoperator. As more quantized operators are to be supported in ET-VK via the PT2E quantization flow, a more generic fusion pass is needed that can handle a variety of fusion patterns.Changes
Introduce the
FuseQuantizedOpsTransform()pass. I elected to introduce a new pass under thebackends/vulkan/_passesdirectory, as opposed to modifying the existing pass because I anticipate the majority of the fusion patterns to be specific to ET-VK.Remove the existing
FuseDequantLinearPass()Switch to using the
FuseQuantizedOpsTransformpass instead of the oldFuseDequantLinearpass.Add
test_vulkan_passesPython test to test export passes.Some small refactors to
test_vulkan_delegatePython test to improve code organizations.Differential Revision: D73794042