[ET-VK] linear_qta8a_qga4w graph pass #12574
Merged
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
# Changes * Introduce `linear_qta8a_qga4w` custom operator in `custom_ops_lib.py` to handle dynamic activation + grouped weight quantized linear operations * Add pattern matching and fusion logic in `FuseQuantizedOpsTransform` to detect and replace dequant + dequant + linear sequences with the new fused operator * Implement comprehensive test coverage in `test_vulkan_passes.py` for the QTA8A_QGA4W fusion pattern validation * Add 4-bit weight packing utilities and grouped quantization support for efficient memory usage # Motivation The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through: * Multiple kernel dispatches instead of a single fused operation * Intermediate tensor allocations for dequantized weights and activations * Suboptimal memory bandwidth utilization The new `linear_qta8a_qga4w` operator fuses the entire sequence into a single operation that: * Directly processes 8-bit quantized activations with per-token scales/zero-points * Handles 4-bit grouped quantized weights with configurable group sizes * Eliminates intermediate dequantization steps by performing dequantization inline * Reduces memory footprint through packed 4-bit weight storage This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy. Differential Revision: [D78291269](https://our.internmc.facebook.com/intern/diff/D78291269/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12574
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 340530c with merge base b6b7a16 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Jul 17, 2025
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
|
This pull request was exported from Phabricator. Differential Revision: D78291269 |
SS-JIA
approved these changes
Jul 17, 2025
# Changes * Introduce `linear_qta8a_qga4w` custom operator in `custom_ops_lib.py` to handle dynamic activation + grouped weight quantized linear operations * Add pattern matching and fusion logic in `FuseQuantizedOpsTransform` to detect and replace dequant + dequant + linear sequences with the new fused operator * Implement comprehensive test coverage in `test_vulkan_passes.py` for the QTA8A_QGA4W fusion pattern validation * Add 4-bit weight packing utilities and grouped quantization support for efficient memory usage # Motivation The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through: * Multiple kernel dispatches instead of a single fused operation * Intermediate tensor allocations for dequantized weights and activations * Suboptimal memory bandwidth utilization The new `linear_qta8a_qga4w` operator fuses the entire sequence into a single operation that: * Directly processes 8-bit quantized activations with per-token scales/zero-points * Handles 4-bit grouped quantized weights with configurable group sizes * Eliminates intermediate dequantization steps by performing dequantization inline * Reduces memory footprint through packed 4-bit weight storage This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy. Differential Revision: [D78291269](https://our.internmc.facebook.com/intern/diff/D78291269/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D78291269 |
# Changes * Introduce `linear_qta8a_qga4w` custom operator in `custom_ops_lib.py` to handle dynamic activation + grouped weight quantized linear operations * Add pattern matching and fusion logic in `FuseQuantizedOpsTransform` to detect and replace dequant + dequant + linear sequences with the new fused operator * Implement comprehensive test coverage in `test_vulkan_passes.py` for the QTA8A_QGA4W fusion pattern validation * Add 4-bit weight packing utilities and grouped quantization support for efficient memory usage # Motivation The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through: * Multiple kernel dispatches instead of a single fused operation * Intermediate tensor allocations for dequantized weights and activations * Suboptimal memory bandwidth utilization The new `linear_qta8a_qga4w` operator fuses the entire sequence into a single operation that: * Directly processes 8-bit quantized activations with per-token scales/zero-points * Handles 4-bit grouped quantized weights with configurable group sizes * Eliminates intermediate dequantization steps by performing dequantization inline * Reduces memory footprint through packed 4-bit weight storage This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy. Differential Revision: [D78291269](https://our.internmc.facebook.com/intern/diff/D78291269/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D78291269 |
# Changes * Introduce `linear_qta8a_qga4w` custom operator in `custom_ops_lib.py` to handle dynamic activation + grouped weight quantized linear operations * Add pattern matching and fusion logic in `FuseQuantizedOpsTransform` to detect and replace dequant + dequant + linear sequences with the new fused operator * Implement comprehensive test coverage in `test_vulkan_passes.py` for the QTA8A_QGA4W fusion pattern validation * Add 4-bit weight packing utilities and grouped quantization support for efficient memory usage # Motivation The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through: * Multiple kernel dispatches instead of a single fused operation * Intermediate tensor allocations for dequantized weights and activations * Suboptimal memory bandwidth utilization The new `linear_qta8a_qga4w` operator fuses the entire sequence into a single operation that: * Directly processes 8-bit quantized activations with per-token scales/zero-points * Handles 4-bit grouped quantized weights with configurable group sizes * Eliminates intermediate dequantization steps by performing dequantization inline * Reduces memory footprint through packed 4-bit weight storage This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy. Differential Revision: [D78291269](https://our.internmc.facebook.com/intern/diff/D78291269/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D78291269 |
# Changes * Introduce `linear_qta8a_qga4w` custom operator in `custom_ops_lib.py` to handle dynamic activation + grouped weight quantized linear operations * Add pattern matching and fusion logic in `FuseQuantizedOpsTransform` to detect and replace dequant + dequant + linear sequences with the new fused operator * Implement comprehensive test coverage in `test_vulkan_passes.py` for the QTA8A_QGA4W fusion pattern validation * Add 4-bit weight packing utilities and grouped quantization support for efficient memory usage # Motivation The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through: * Multiple kernel dispatches instead of a single fused operation * Intermediate tensor allocations for dequantized weights and activations * Suboptimal memory bandwidth utilization The new `linear_qta8a_qga4w` operator fuses the entire sequence into a single operation that: * Directly processes 8-bit quantized activations with per-token scales/zero-points * Handles 4-bit grouped quantized weights with configurable group sizes * Eliminates intermediate dequantization steps by performing dequantization inline * Reduces memory footprint through packed 4-bit weight storage This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy. Differential Revision: [D78291269](https://our.internmc.facebook.com/intern/diff/D78291269/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D78291269 |
facebook-github-bot
merged commit 3178517
into
gh/ahmtox/42/base
101 of 103 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
Gasoonjia
pushed a commit
that referenced
this pull request
Jul 30, 2025
This PR was created by the merge bot to help merge the original PR into the main branch. ghstack PR number: #12574 by @ahmtox ^ Please use this as the source of truth for the PR details, comments, and reviews ghstack PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/42/base ghstack PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/42/head Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/41/orig Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/42/orig @diff-train-skip-merge --------- Co-authored-by: morelos <morelos@devvm4573.ash0.facebook.com> Co-authored-by: ahmtox <69552192+ahmtox@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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):
Changes
linear_qta8a_qga4wcustom operator incustom_ops_lib.pyto handle dynamic activation + grouped weight quantized linear operationsFuseQuantizedOpsTransformto detect and replace dequant + dequant + linear sequences with the new fused operatortest_vulkan_passes.pyfor the QTA8A_QGA4W fusion pattern validationMotivation
The existing quantization workflow in Vulkan backend processes dynamic activation + grouped weight quantized linear operations as separate quantize/dequantize/linear steps, which creates performance overhead through:
The new
linear_qta8a_qga4woperator fuses the entire sequence into a single operation that:This aligns with the broader goal of optimizing quantized model inference in the Vulkan backend by leveraging graph-level transformations to improve computational efficiency while maintaining numerical accuracy.
Differential Revision: D78291269